.

International Women’s Day has its roots in the early 1900s, born from labour movements and a push for women’s rights. Over the decades, March 8th has evolved into a global recognition of women’s achievements, resilience, and the ongoing quest for equality.

Yet recent current events illustrate that there’s still much work to do. The reality that women’s progress isn’t linear is evident. We can’t let these challenges overshadow the significance of this day, it’s a call to rest, recharge, and rise stronger.

So, begin by nurturing yourself. Then, carry that renewed strength into lifting up the women around you. By filling your own cup first, you help ensure there’s enough energy to continue the long journey toward true gender equality and that’s a cause worth celebrating every day.

In finance, this dynamic is called financialization. Complex organizations become measurable signals that can be priced, traded, and optimized. Fragmentation makes systems faster, but thinner. When firms are reduced to short-term metrics, they become easier to evaluate and harder to truly understand. Standing in my lecture hall this term, the same logic was no longer abstract. When learning becomes a series of signals that can be quickly assessed and easily scaled, integration does not disappear by accident. It disappears by design.

Photo by Po-Hsuan Huang on Unsplash

I ran a small experiment in my largest finance course. Half the lecture hall is cellphone friendly. The other half is intentionally cellphone free. Not because this technology is the enemy, but because sustained reasoning needs protected space. If our classroom were a ship during the first few weeks, we would have sunk quickly. Students arrived early to secure seats on the cellphone friendly side. Convenience did not need to be negotiated. Within minutes, the pattern was set.

The moment itself was ordinary, but what it revealed was not. Faculty conversations often circle back to the same conclusion: students cannot think critically anymore. It is an appealing narrative because it locates the problem inside the learner and leaves the surrounding structures untouched. Yet watching students optimize for speed and cognitive ease, it became difficult to ignore a different possibility. If our assessments reward recognition, velocity, and isolated calculation, fragmented thinking is not a mystery. It is the expected outcome. Perhaps the real issue is not declining capacity but learning environments that claim to measure integration while rarely building it into the architecture of learning itself.

The Atlantic essay describes assigning full books again instead of excerpts. In finance, the equivalent is refusing to reduce learning to single variable problems. Real decisions do not arrive neatly packaged. Growth strategies interact with liquidity. Financing choices reshape flexibility. Dividend commitments create pressure months later. Systems thinking does not emerge from exposure alone. It emerges when students are required to remain inside interacting constraints long enough for patterns to become visible.

Over the past term, exam analytics began to redraw the map of my assessments. Instead of organizing questions by topic, I started watching where thinking changed. The data revealed something difficult to ignore. Entire clusters of students stalled at the same hinge point, the moment when isolated ideas had to become a system. The math was not the barrier. The integration was. This was not a flawed question or an unusually difficult calculation. It was a structural break point revealed by the data itself. That pattern did more than reshape my teaching strategy. It exposed a deeper misreading. Instead of asking, “Why can’t students think critically?” a harder question emerged. Why are we measuring integration in environments that rarely scaffold it? Once those hinge points became visible, the response was not remediation but design. I began building small, optional practice pathways at the exact points where integration first breaks down, not to simplify the work, but to give students structured opportunities to practice holding complexity before they are asked to perform it.

That realization re-framed assessment itself. Exams are often treated as endpoints, a final measure of whether learning occurred. But when viewed through analytics, they become maps. They reveal where thinking begins to strain, where cognitive load spikes, and where growth is actually waiting to happen next. Students do not falter randomly. They encounter predictable transitions: from recognition to integration, from calculation to judgment. When those transitions remain invisible, struggle looks like inability. When they are made visible, struggle exposes a design gap.

Seen this way, critical thinking is not a mysterious trait students either possess or lack. It is a sequence of cognitive movements that can be engineered. That idea carries uncomfortable implications. It suggests that institutions may be measuring surface fluency while neglecting the architecture required for deeper reasoning. If learning environments increasingly resemble fragmented markets, fast, efficient, and signal driven, then fragmented thinking should not surprise us. It should prompt us to examine the systems we have built.

Engineering growth does not mean lowering expectations. It means placing difficulty deliberately rather than assuming it will emerge on its own. In finance we know that systems optimized purely for efficiency eventually lose resilience. The same dynamic appears in education. When everything becomes faster, shorter, and more modular, the capacity to hold complexity erodes. Designing pathways is not nostalgia. It is an attempt to restore integration in environments increasingly shaped by fragmentation.

As exam results arrive, some students may begin to reconsider how they position themselves in the room. Not because of rules, but because they start to notice differences in how they think under different conditions. The shift may appear small from the outside. Yet it points to something larger: a shift away from chasing signals toward learning how systems actually hold together.

Perhaps the real question raised by that Atlantic essay is not whether we should stop meeting students where they are. The harder question is whether we are willing to admit that the structures surrounding them, and sometimes the assessments we design, may be shaping how they think long before they ever enter our classrooms. If that is true, then the work of higher education is not simply to diagnose ability after the fact. It is to design pathways where deeper thinking becomes unavoidable.

What began as a brief, unscheduled conversation during an open house did not end in the hallway. She invited both my daughter and me back to her office, sat with us, and spoke as if our questions truly mattered. That moment grew into a relationship of trust, encouragement, and shared intellectual work. My daughter chose not only the institution, but her discipline, because one professor took the time to truly see her.

During my daughter’s time at the university, this professor was diagnosed with cancer. She stepped away to receive treatment, then returned to the work she loved. Even as her health declined, she continued to meet with students, write letters, and offer guidance for as long as she was able. That devotion, quiet, steady, and unannounced, taught as much as any course ever could.

There is no place for this kind of work in a spreadsheet or a financial model, yet it shaped the course of a life, and it points to something larger: the invisible human infrastructure that holds universities together, and what we lose when we design systems that forget it exists.

The Budget Room and the Hallway

As our university faces ongoing budgetary pressures, we sit through meeting after meeting discussing the problem and possible solutions. We model scenarios, reallocate lines, cut, defer, and consolidate.

As an accountant, I understand this world deeply. I have stood inside its logic, and I have seen its stark ugliness, the way necessity slowly trains us to look away from what cannot be counted.

In the budget room, the language is about efficiency, enrolment, and cost containment. These are necessary conversations. But what is missing from them is humanity and the quiet, informal exchanges between professors and students that so often carry the real educational weight.

When universities speak about the “student experience,” they usually mean the managed side of it: how efficiently students can register for courses, how long they wait for advising, how smoothly learning platforms function, whether residences and fitness centres are full, whether food services and study spaces are used, how many students attend orientation or wellness programs, how quickly assessments are returned, how many graduate on time, and what satisfaction surveys report back.

All of this matters, but it is not what students mean when they talk about experience.
The real student experience is formative, shaped through identity-shaping encounters.

Students do not remember systems; they remember encounters, the professor who noticed them struggling, the conversation that made them feel capable, the moment someone said, you belong here, the door that was open when life was heavy, the belief that changed what they thought was possible.

These moments aren’t things you can list in a brochure or a budget, but they are what students carry with them.

A Pattern, Not an Anecdote

This summer, while supporting my daughter at a marathon, I ran into a volunteer who had once been my student. Years earlier, she had been balancing coursework while raising a young child. I remembered her for her persistence. She remembered me for something else.

She told me that without the small accommodations, the informal check-ins, and the sense that someone noticed her struggle, she would not have finished. She would not have gone on to graduate study. She would not have the career she has now.

This is only one of many such stories that have found their way back to me. Most never do. These interactions are not rare; they are simply invisible.

None of this will ever appear in a teaching evaluation or an institutional report, yet it changed a life.

Over the years, I have had countless impromptu, informal conversations with students, in hallways, after class, and in moments when they finally felt safe enough to speak. Many are too afraid to raise their voices in large lectures. Many feel they do not belong. Many struggle quietly with the fear that they are not good enough.

What they need in those moments is not instruction. They need someone who listens not as a parent, not as a friend, but as a trusted adult who sees who they are becoming.

I think of these moments as acts of empowerment, not because they are simple, but because they restore belief, especially for students who arrive at university seeking a second chance.

There is no policy that mandates this work, no workload model that protects it, and no metric that captures its impact, yet this is where transformation often begins.

What the Research Has Long Shown

More than fifty years ago, educational researchers were already documenting this truth.

They found that the single strongest factor distinguishing faculty who “made a difference” from those who did not was the extent of their out-of-class interaction with students. These relationships were not brief or transactional; they were casual, sustained, intellectually engaging, and grounded in personal recognition. Students reported that such faculty helped them choose majors, formulate career plans, and develop a sense of purpose. This kind of influence requires time and continuity, conditions that are rarely possible in short-term or per-course teaching assignments.

Much of the most effective teaching, the research concluded, occurs outside the classroom, in the spaces institutions rarely design for, protect, or measure.

This is not nostalgia; it is grounded in decades of research.

The Structural Erosion of Relationship

As universities increasingly optimize for scale, efficiency, and throughput, the informal spaces where relationships grow begin to shrink, not because professors do not care, but because the system no longer makes room for care.

We are building universities as delivery systems, while students experience them as living ecosystems of connection. When those ecosystems erode, something essential disappears.

The Threshold Figures We Are Losing

My daughter’s professor did not recruit her; she recognized her. Even while ill, she continued to mentor, to write, to encourage, and to show up, not because it would ever be counted, but because she understood that this was the work.

This is not a marketing function. It is a human one and it is increasingly fragile.

What We Must Remember

The most important infrastructure in a university is not its buildings, platforms, or budgets. It is the quiet courage professors give students when they say, in a hundred small ways, you are seen, you are capable, you are not alone.

The university remembers numbers. Students remember names and sometimes, that is the difference between who they might have been and who they become.

In memory of Dr. Karen Gordon, whose work continues in the lives she changed.

This final part of this series argues that the defining constraint of the AI era is no longer technological capability, but the capacity of human systems to form and sustain professional judgment under conditions of automation and scale. Intelligence is increasingly abundant. Judgment is not.

1. Formation as Public Infrastructure

(Why Human Judgment Cannot Be Privately Optimized)

The most fundamental design constraint is that human judgment produces collective value but requires private cost. This makes it structurally under-supplied by markets.

Gary Becker formalized this logic in his theory of human capital. He distinguished between firm-specific training, which firms are willing to fund because the returns are internal, and general training, which increases a worker’s productivity across many employers. Under competitive conditions, Becker showed, firms rationally under invest in general training because its benefits cannot be reliably captured:

“In a competitive labour market, firms will not pay for general training, since they can never be sure that they will be able to recoup their investment.”
Becker, Human Capital (1964)

This is why hospitals rely on publicly funded residency programs, why lawyers must complete articling, and why pilots log supervised flight hours before carrying passengers, while new analysts, HR professionals, and AI engineers are increasingly expected to perform autonomously at full speed from the outset.

From the firm’s perspective, formation is leakage. From society’s perspective, it is infrastructure. In practice, it lives in residencies, articling systems, apprenticeships, training rotations, and supervised early-career roles, precisely the arrangements we already mandate in medicine, law, engineering, and aviation because markets alone cannot be trusted to produce safe professionals.

Human development creates spillovers that no single organization can appropriate. Skills migrate, judgment compounds across careers, and institutional memory diffuses through labour mobility. The very features that make formation socially valuable make it economically irrational for individual firms to fund at scale.

Mariana Mazzucato and Dani Rodrik extend this logic beyond training into institutional design itself. Mazzucato shows that foundational capacities are systematically under provided by private markets and therefore carried by public institutions. Rodrik argues that markets require extensive supporting infrastructures in order to function at all:

“The state is not simply fixing market failures. It is actively shaping and creating markets.”
Mazzucato, The Entrepreneurial State (2013)

“Markets are not self-creating, self-regulating, or self-sustaining.”
Rodrik, Economic Rules (2015)

Human judgment is one of those supporting institutions. It is not simply a trait of individuals, but a systemic capacity produced through education systems, supervised early-career roles, mentorship, and institutional tolerance for developmental inefficiency.

This section explains why formation collapses in markets. The next turns to how judgment actually forms.

2. Protected Development Zones

(Institutional Slack by Design)

The second constraint is that learning cannot occur under continuous performance pressure.

Human judgment forms through uncertainty, reflection, and repeated exposure to consequence under protection. Yet modern institutions increasingly evaluate people through immediate output: productivity metrics, performance dashboards, algorithmic monitoring. Under these conditions, learning is forced to masquerade as performance. Error becomes failure. Exploration becomes inefficiency. Development becomes risk.

This is not how judgment forms.

Donald Schön, David Kolb, and Lev Vygotsky, all converge on the same insight: professional learning does not occur through optimization or instruction alone, but through cycles of experience, reflection, guided failure, and gradual expansion of responsibility (Schön, 1983; Kolb, 1984; Vygotsky, 1978). Across all three traditions, the lesson is identical:

Learning requires slack.

Protected development zones are not training programs inside firms. They are institutionally insulated spaces where performance metrics are formally suspended. They are not about skill acquisition. They are about risk containment.

Real systems already embody this logic. Teaching hospitals embed trainees inside supervised environments where error is expected and corrected. Legal articling systems place new lawyers under formal mentoring and limited liability. Professional licensure in medicine, law, accounting, engineering, and aviation all require extended periods of supervised practice before independent authority is granted.

These are not inefficiencies. They are design choices. They exist because we understand that unformed judgment is dangerous. Where the cost of error is high, whether physical harm, legal harm, or systemic risk, we refuse to let performance precede formation.

What is striking is that we have never applied this same logic to the sectors now exercising comparable power over social life.

Finance governs capital allocation and systemic risk. HR systems govern hiring, surveillance, and promotion through automated screening. AI governance teams design decision systems that rank and classify people at scale. Public policy increasingly relies on models and dashboards to shape collective outcomes.

These domains do not merely produce goods. They actively produce social reality.

Yet none of them operate with protected formation architectures comparable to medicine, law, or accounting.

We have teaching hospitals for the body.
We have no teaching institutions for the economy.
Where systems exercise high social power, we normally demand developmental insulation.

3. Shared Formation Consortia

(Cross-Firm Apprenticeship Pools)

The problem is not that firms are unwilling to train. It is that no single firm can afford to train alone.

Labour mobility and competitive markets mean that organizations investing heavily in general training bear the cost while others capture the benefit.

Firms under invest in general skills because workers can leave, taking their competence with them, as Acemoglu and Pischke demonstrated (Acemoglu & Pischke, 1998; 1999).

Shared formation consortia resolve this contradiction by shifting development from the level of the firm to the level of the sector. Instead of each organization carrying the risk alone, groups of firms co-fund training pipelines, share apprentices, and participate in common certification systems. Formation becomes collective infrastructure rather than a competitive disadvantage.

This logic is not new. It is how pre-modern guild systems sustained craft knowledge across generations. It remains the dominant design pattern wherever complex skills must be reproduced at scale.

Across sectors, formation is already organized collectively: Swiss industry boards coordinate apprenticeships, Canada’s Red Seal standardizes training nationally, semiconductor firms co-fund workforce pipelines, and the University of Waterloo distributes student development across hundreds of employers.

These systems work because they accept structural reality:

• workers move,

• skills diffuse,

• firms compete.

And they design around it.

Shared formation consortia do not ask firms to become altruistic.
They make formation rational again.

4. Design Mandates for Human-in-the-Loop Systems

(Why Machines Must Train Humans)

This is not an ethics problem. It is a systems design problem.

We do not need human oversight to control machines.

We need machine exposure to train humans.

Lisanne Bainbridge’s analysis of the ironies of automation showed that when systems remove routine human involvement, skill does not remain dormant. It decays (Bainbridge, 1983). As automated systems absorb the easier aspects of a task, human operators are left with only rare, high-stakes interventions, precisely the situations in which they are least prepared to perform. Automation reduces continuous engagement while increasing the cognitive and moral burden placed on humans when failures occur.

Human factors engineering and socio-technical systems theory reach the same conclusion. Safe systems are not those that minimize human presence, but those that preserve:
• decision exposure
• consequence visibility
• supervised autonomy

These principles are already embedded in high-risk domains. Aviation cockpits preserve pilot engagement even under heavy automation. Nuclear control rooms maintain manual intervention pathways. Clinical decision support systems present recommendations while requiring physicians to remain responsible for final judgment.

These systems are not optimized for efficiency alone. They are optimized for human formation inside technological constraint. They recognize that judgment can only be preserved by designing systems that continue to expose humans to real decisions, real consequences, and real responsibility under supervision.

If machines absorb all routine decisions, humans do not become more strategic.
They become more fragile.

Which is why the real governance problem is not how to constrain artificial intelligence.
It is how to ensure that intelligence, human judgment, does not quietly atrophy inside the very systems we build to enhance it.

The defining challenge of the AI era is not how to make machines more intelligent, but how to prevent human judgment from disappearing inside systems that no longer need it to function.

References

Acemoglu, D. and Pischke, J. S. (1998). Why do firms train Theory and evidence. The Quarterly Journal of Economics, 113(1), 79-119.

Acemoglu, D. and Pischke, J. S. (1999). Beyond Becker Training in imperfect labor markets. The Economic Journal, 109(453), F112-F142.

Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775-779.

Becker, G. S. (1964). Human capital: A theoretical and empirical analysis, with special reference to education. University of Chicago Press.

Kolb, D. A. (1984). Experiential learning: Experience as the source of learning and development. Prentice Hall.

Mazzucato, M. (2013). The entrepreneurial state: Debunking public vs. private sector myths. Anthem Press.

Organization for Economic Co-operation and Development. (2019). Getting skills right: Future-ready adult learning systems. OECD Publishing. https://doi.org/10.1787/9789264311756-en

Rodrik, D. (2015). Economic rules: The rights and wrongs of the dismal science. W. W. Norton & Company.

Schön, D. A. (1983). The reflective practitioner: How professionals think in action. Basic Books.

Vygotsky, L. S. (1978). Mind in society: The development of higher psychological processes. Harvard University Press.

World Bank. (2019). The changing nature of work: World development report 2019. World Bank. https://doi.org/10.1596/978-1-4648-1328-3

In Part I, I argued that the rapid scaling of artificial intelligence is not merely a technological shift or a labour-market disruption. It is a quiet dismantling of formative institutions. We are investing heavily in systems that perform immediately, while withdrawing investment from the slow, error-tolerant processes through which human judgment is formed. Professional judgment does not emerge fully formed. It develops in layers, through supervision, failure, and time spent inside institutions that once assumed people would arrive unfinished. When those conditions disappear, judgment thins, even as tool fluency accelerates. When those conditions disappear, judgment thins, even as tool fluency accelerates.

Part II examined what entry-level work historically provided in this process. Early work was not primarily about output. It functioned as a formation system, a protected on-ramp into adulthood where inefficiency was tolerated because growth was expected. Apprenticeships, post-war corporate ladders, and early professional roles absorbed error in service of long-term capacity. They transmitted tacit knowledge, socialized judgment, and allowed professional identity to form inside institutions.

Taken together, Parts I and II make a simple claim: what is eroding is not just a set of jobs, but a developmental infrastructure. Entry-level work once carried the burden of becoming. That burden has not disappeared, but the institutions designed to carry it have.

Part III begins where that diagnosis leaves us, not with answers, but with limits. If the erosion of early-career formation is systemic rather than accidental, then the response cannot be moral exhortation or nostalgic revival. It must begin with constraints. The question is not what we wish would work, but what a non-naïve attempt to rebuild human judgment would have to respect.

Why the Obvious Fixes Fail

These responses fail not because they are ill-intentioned, but because they misdiagnose the nature of the problem they are trying to solve.

1. “Firms Should Just Train More”

Calls for firms to “just train more” ignore a basic economic constraint: the returns to development are diffuse and long dated. The organization that bears the cost of formation rarely captures its full benefit. Skills travel. Judgment migrates. People leave. From a firm’s perspective, training increasingly looks like leakage rather than investment, a dynamic long recognized in human capital theory, where firms underinvest in general training because its benefits are easily appropriated by competitors (Becker, 1964; Acemoglu & Pischke, 1999). This is not a failure of corporate memory, or even of intention. Under conditions of mobility, competition, and short investment horizons, firms rationally treat general training as difficult to appropriate.

2. “Individuals Should Take Responsibility”

This reframes a structural failure of formation as a personal failure of effort.

As Ulrich Beck (1992) argued, modern systems increasingly atomize individuals, pushing them into an “individual struggle for existence” in which structural risks are reframed as personal responsibilities. As collective institutions weaken, people are asked to self-manage uncertainty, failure, and development, despite lacking the shared structures through which judgment actually forms. Social crises are no longer understood collectively, but psychologically, as anxiety, inadequacy, or poor self-management. Atomization does not produce autonomy; it privatizes risk. The result is not empowerment, but a new inequality: the unequal capacity to navigate insecurity itself.

3. “AI Will Free Humans for Higher-Order Work”

The idea that automation will simply “free humans for higher-order work” assumes that judgment appears automatically once routine tasks disappear. Human-factors research suggests the opposite. In her classic paper Ironies of Automation, Lisanne Bainbridge, (Bainbridge, 1983), shows that when automation removes routine control, human skill does not lie dormant, it decays. “Physical skills deteriorate when they are not used,” she notes, meaning that operators left to monitor automated systems may become less capable precisely when intervention is required. Judgment, Bainbridge argues, develops only through repeated use and feedback; knowledge not exercised becomes difficult to retrieve when conditions change. Automation does not pause skill development; it actively reshapes it. As a result, “by taking away the easy parts of [the] task, automation can make the difficult parts… more difficult.” Automation without formation does not elevate human capacity. It produces fragility.

4. “Skills-First Hiring Will Solve the Entry-Level Problem”

One increasingly influential response to the erosion of entry-level work is the move toward skills-first hiring. Rather than relying on degrees or job titles, skills-first approaches aim to assess candidates based on demonstrated competencies.

As the Organization for Economic Co-operation and Development (OECD) explains, skills-first hiring prioritizes “a person’s demonstrated skills and competencies—regardless of how or where they were acquired—over traditional signals such as degrees or job titles.” The motivation is clear: to reduce credential inflation, widen access, and improve matching in fraying labour markets. But the OECD also notes that this shift raises unresolved challenges, particularly around how skills are developed, validated, and compared across contexts.

The consequence is that skills-first hiring does not eliminate bias; it reorganizes it around visibility and measurability. It privileges visible, testable, and narratable skills over tacit judgment, growth potential, and slow-forming competence. What cannot be easily demonstrated early is treated as absence rather than immaturity. In doing so, it risks amplifying tool fluency, early exposure, and self-presentation advantages, while further compressing the space in which judgment is allowed to develop.

What looks like a corrective to credentialism can, under pressure for speed and measurability, further narrow the developmental runway it claims to restore.

What a Non-Naïve Fix Would Have to Respect

If these failures are structural, then any serious response must begin by respecting the realities they reveal. These are not preferences or values; they are constraints imposed by how human judgment actually forms.

1. Development Is Slow by Nature

Human judgment does not scale at the speed of technology. It forms incrementally, through repeated exposure to uncertainty under protection. Beyond certain limits, attempts to compress this process do not accelerate judgment—they deform it. A non-naïve fix must therefore accept developmental latency as a feature rather than a flaw.

2. The Returns to Formation Are Diffuse

What appears in firm-level debates as a training problem is, at root, a feature of human formation itself. The benefits of development do not accrue neatly to a single organization. They spill across firms, sectors, and careers. This makes formation structurally under-supplied by markets operating on short horizons.

A non-naïve fix must therefore operate beyond isolated firm incentives, either by sharing costs, extending time horizons, or embedding development in broader institutional arrangements.

3. Learning Cannot Be Measured Like Performance

When learning is evaluated using performance metrics, it disappears.

Early-stage error becomes indistinguishable from failure. Exploration reads as inefficiency. Variance is punished rather than absorbed.

Judgment formation requires spaces where mistakes are expected and corrected without immediate penalty. These spaces must be structurally protected, not merely encouraged.

A non-naïve fix must therefore separate learning from short-term performance evaluation, at least during formative stages.

4. Judgment Requires Redundancy and Supervision

Efficient systems minimize overlap. Developmental systems require it. Judgment forms through comparison, correction, and guided practice—processes that appear inefficient but produce durable capacity. What looks like waste in performance systems is often the very mechanism through which judgment stabilizes.

When redundancy is stripped away in the name of efficiency, judgment thins. What remains may be fast, but it is brittle.

A non-naïve fix must therefore tolerate apparent inefficiency in service of long-term resilience.

5. Automation Changes the Terrain but Does Not Remove the Need for Formation

Automation can reduce cognitive load and eliminate drudgery. It cannot generate judgment. Higher-order work does not emerge automatically when lower-order work disappears; it must be formed before it can be performed. It emerges only when people are gradually trained to navigate complexity. A non-naïve fix must therefore treat AI as complementary to formation, not a substitute for it—especially early in careers.

Re framing the Question

These constraints narrow the design space considerably. They rule out quick fixes, individualizing responsibility, and purely market-driven solutions.

The question is no longer how to “save entry-level jobs.”

It is where, in an economy optimized for speed and automation, human judgment is supposed to be formed.

That question already has answers. But only if we stop pretending the problem is accidental—or solvable through efficiency alone.

Where This Leaves Us

Part I named the civilization capacity problem.
Part II showed what entry-level work once provided.
Part III clarifies the limits any real attempt to rebuild human judgment must respect.

What remains is not certainty, but choice.

We are already designing the future of human capacity—through what we fund, measure, and tolerate.

Part IV will follow, which will highlight workable pathways.

References

Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8Bottom of Form

Beck, U. (1992). Risk society: Towards a new modernity. Sage.

Becker, G. S. (1964). Human capital: A theoretical and empirical analysis, with special reference to education. University of Chicago Press.

In the 1980s and 1990s, manufacturing was off shored in the name of efficiency, shareholder value, and global competitiveness. Plants closed not because managers sought to erase working-class futures, but because capital learned that labour could be geographically displaced. Whole apprenticeship ecosystems quietly collapsed as a side effect.

Today, a similar logic is unfolding at the cognitive layer of the economy. In May 2025, Dario Amodei, co-founder and CEO of Anthropic, warned that AI could wipe out roughly 50 percent of all entry-level white-collar jobs within five years (Morris, 2025).

This change, too, does not arrive as sabotage. It arrives through optimization. Through cost curves. Through architectural decisions made far upstream from the lives they restructure.

Last fall, a newly hired audit associate arrived for what was supposed to be her first day of on boarding. The intake program had been quietly cancelled weeks earlier. Over the summer, the firm had implemented automated testing, continuous transaction monitoring, and client-side data extraction. The junior tier was no longer required. The change was framed as an efficiency gain. For her, it meant the disappearance of a beginning—no slow first year, no protected margin for error inside the profession she had trained to enter.

What kind of civilization are we implicitly choosing when we scale machines faster than we scale people?

In my advanced accounting lecture, I introduced the reading The Problem with Accounting for Employees as Costs Instead of Assets by Ethan Rouen, Associate Professor at Harvard, to encourage debate. It argues that because accounting standards treat employees and training as expenses rather than assets, firms have little formal incentive to invest in human capital, even while calling employees their “most valuable asset.”

Overall, the students felt uncomfortable with this argument: “It is wrong.” “People are not assets.” Their resistance did not come from careful analysis so much as from a deeper reflex, one shaped by ingrained accounting rules, in which the human has already been priced as an expense.

If judgment is being quietly removed from the human path, it is not because we collectively forgot its importance. It is because our economic and institutional systems no longer reward its cultivation. What we are witnessing is not a cultural accident. It is the predictable outcome of financial architectures, productivity metrics, and organizational designs that increasingly treat human formation as a cost to be minimized rather than a capacity to be built.

To understand why this is happening, we must look not at individual firms or individual technologies, but at the deeper structure governing how value itself is measured.

We are directing unprecedented capital toward technologies that reduce the need to invest in humans and then calling the resulting human disinvestment “efficiency.” It has become cheaper to replace human capacity than to grow it.

We automate the task instead of cultivating judgment. We optimize performance instead of formation.And then we act surprised when the institutions that once taught people how to judge begin to thin.

What we are witnessing is not merely automation. It is a quiet de-skilling of judgment. A civilization can absorb extraordinary technological change. What it cannot absorb indefinitely is the loss of the human capacities that once made that change governable.

Entry-Level Work Was a Formation System

For most of modern history, early work functioned less as employment in the narrow sense and more as a social apprenticeship, even in white-collar life. In guilds and trades, the path from apprentice to journeyman to master was explicitly about becoming rather than efficiency. Apprentices slowed production. They made costly mistakes. They required years of oversight. Yet this was not treated as waste. It was understood as investment in human capacity and social continuity.

These jobs were not judged primarily by output per hour. They were judged by who the worker became over time. Early work functioned as a human time-buffer—a gradual acclimation chamber for adulthood, professional judgment, and institutional belonging.

That developmental buffer now shows signs of strain. In November 2025, Erik Brynjolfsson, Director of the Stanford Digital Economy Lab, published a working paper titled Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, examining whether AI-driven impacts are already visible in the labour market, particularly for entry-level workers. Using ADP payroll data, the paper documents a 13 percent decline in employment among 22–25-year-olds in AI-exposed occupations—a pattern not observed among more experienced workers.

From a narrow efficiency perspective, early-career labour has always looked irrational: junior workers duplicate tasks, analysts are slower than experienced staff, interns introduce risk, and trainees consume mentor time. By the logic of short-run productivity, these roles are difficult to justify.

Yet institutions sustained them for centuries because they performed functions that never appeared on financial statements. Early roles rewired cognition, contained error, socialized behaviour, and formed identity. Young workers learned to think under constraint, operate inside ambiguity, fail without being erased, and regulate emotion within hierarchy.

This is developmental psychology, not labour economics. Early jobs were a maturation layer disguised as inefficiency.

The Human Growth Pipeline Is Being Quietly Rewritten

This is the underlying structural shift. The old growth pipeline was straightforward:

School → Low-stakes work → Supervised complexity → Independent judgment → Leadership

It assumed time, variance, mistakes, mentorship, and human redundancy.

The emerging pipeline is sharper and far less forgiving:

School → Credential signalling → Immediate performance → Algorithmic monitoring → Replaceability

It assumes minimal slack, little tolerance for slow learning, and near-continuous performance visibility.

This is not simply automation. It is a change in institutional design logic. We no longer train people to become; we extract performance immediately. We no longer absorb developmental variance; we attempt to eliminate it. We no longer treat time as investment; we increasingly treat it as cost.

Entry-level roles are not disappearing because machines are inherently “smarter.” They are disappearing because systems no longer tolerate developmental latency. We are not merely removing jobs. We are dismantling the slow ramp into adulthood itself.

Training, supervision, and developmental failure are expensed immediately. Compute, automation, and software systems are capitalized and amortized. This single accounting asymmetry quietly determines where investment flows. One form of capacity compounds on balance sheets. The other disappears as cost.

The Deeper Risk Isn’t Unemployment

The deeper risk is generational. We are shaping a cohort that may be technically fluent but judgment-poor, fast but ambiguity-intolerant, credentialed but institutionally shallow, and digitally accelerated but developmentally unformed.

This is not primarily a labour-market disruption. It is a disruption of civilization’s capacity to form judgment.

Why This Is Not Simply an AI Ethics Story

Most contemporary AI ethics debates focus on governance questions: bias, fairness, transparency, and accountability. These are essential concerns about how decision systems operate. But what is eroding here is deeper: the developmental function of social institutions themselves.

The core question is not whether AI is fair. It is whether people are still given structured space to become.

This is not yet an AI ethics problem. It is an institutional ethics problem, a question of whether human development remains a core design requirement of the economic system at all.

What is quietly dissolving is not just job access, but developmental permission itself: the right to start slowly, the right to be bad before becoming good, the right to learn inside protected structures, and the right to be economically unproductive while cognitively productive.

These rights were never sustained through moral argument. They were sustained through institutional design. And institutional design is changing. Speed, precision, measurability, and replaceability are increasingly privileged over the slower conditions that human maturation requires. The two design philosophies are no longer aligned.

Conclusion — Bridge to Part III

What is disappearing, then, is not only work at the bottom of the ladder. It is the ladder itself. Not jobs alone, but the institutional patience that once allowed human capacity to compound slowly through error, supervision, and time.

This did not happen because we forgot the value of human development. It happened because we built economic systems that no longer know how to price it. We optimized for what could be measured, capitalized, and scaled—and in doing so, we quietly devalued what could only be grown.

Artificial intelligence did not create this logic. It merely exposed it, accelerated it, and made its developmental consequences impossible to ignore.

If judgment is now thinning, if adulthood is now being compressed, and if the first rungs of formation are now being engineered out of the system, then the question before us is no longer whether AI is powerful.

The question is whether the economic architectures governing its deployment still contain a place for human becoming at all.

Sources

Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab.
https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/

Inskeep, S., & Dumas, N. (2025, August 5). AI could widen the wealth gap and wipe out entry-level jobs, expert says [Radio broadcast transcript]. NPR.
https://www.npr.org/2025/08/05/nx-s1-5485286/ai-jobs-economy-wealth-gap

Morris, C. (2025, May 28). Anthropic CEO warns AI could eliminate half of all entry-level white-collar jobs. Fortune. https://fortune.com/2025/05/28/anthropic-ceo-warning-ai-job-loss/

Rouen, E. (2019). The problem with accounting for employees as costs instead of assets. Harvard Business Review.

With the advent of texting, the increase in silence has been deafening. When questions are asked, few venture to respond. Students arrive already fluent in the language of digital tools, navigating interfaces with a speed that now feels instinctive to them. At first glance, they look ready. They retrieve information instantly, generate responses at remarkable speed, and move through systems that didn’t exist when I was their age.

And yet, in the first weeks, something else is always present just beneath the surface: hesitation.

Once the silence breaks, they often answer with uncertainty, waiting for confirmation not just that they are correct, but that it is permissible to be uncertain. They are fast, but not yet steady.

Then, slowly,almost imperceptibly at first, something begins to change.

A student raises a hand not to give the answer, but to question the framing.
Another says, “I’m not sure yet,” without embarrassment.
Over time, they begin to sense when a number is technically correct but ethically uncomfortable
Another pushes back gently on a conclusion that once would have gone unchallenged.

These are not dramatic moments. They do not show up cleanly in grades or metrics. But they mark the quiet appearance of something far more important than technical skill.

They mark the beginning of judgment.

I have watched this transformation unfold again and again across years of teaching. And what strikes me most is how dependent it is on time, friction, and exposure to consequence. It cannot be rushed without being altered. It cannot be optimized without being thinned. It requires space to be unfinished.

When people talk about productivity, automation, and artificial intelligence, they often speak as if learning were simply a transaction between information and efficiency. But what I witness, over and over, is that learning is also a relational, moral, and temporal process. It unfolds between people. It depends on patience. It depends on the slow permission to be wrong in the presence of someone more experienced.

This is why the changes now sweeping through work feel so consequential to me.

I am not only thinking about jobs.

I am thinking about what happens to this fragile, necessary process of becoming, when the systems we pass young people into no longer need them to move slowly at all.

That awkward, uncertain beginning is not a flaw in human development. It is the design.

Before we ever talk about work, automation, or artificial intelligence, we have to start somewhere much older and much simpler: the way human minds change across a lifetime. Because the unease many of us feel about the future of work is not only economic. It is developmental.

In early adulthood, the human brain is fast. It learns quickly, adapts rapidly, and moves easily between unfamiliar tools and ideas. It is built for exploration. This is the mind of the beginner: quick to try, quick to recover, quick to move on.

Later in life, something different begins to happen. The brain does not simply decline. It trades speed for structure. It becomes less interested in raw novelty and more attuned to pattern. It loses a measure of immediate recall but gains a sense for what tends to repeat. What looked like isolated events at 25 begins to resemble recognizable cycles at 50.

This is not degeneration. It is a shift in function.

Young minds are optimized for learning quickly. Older minds are optimized for understanding deeply. One scouts new terrain. The other remembers where the ground collapses.

These two ways of thinking were never meant to compete. They were meant to depend on one another. Healthy systems require both rapid exploration and slow recognition. Speed without depth is reckless. Depth without speed is stagnant.

How Judgment Forms

Judgment is not information. It does not transfer cleanly from one mind to another. It forms only when knowledge collides with consequence.

Early mistakes are usually small and buffered. Someone intervenes. A system absorbs the cost. “Here’s what that sets in motion,” an experienced person says. Over time, the beginner begins to feel what will follow before it happens. The future presses backward into the present.

This is the birth of judgment.

Judgment is not speed. It is foresight shaped by memory. It is knowing when not to proceed, even when the numbers look right. It is the ability to sense that correctness and responsibility are not always the same thing.

This capacity cannot be rushed. You can compress information. You can accelerate output. But you cannot shortcut the slow internal reorganization that allows a person to feel consequence in advance.

Without that slow transformation, we do not lose competence first.We lose the instinct to hold back

Who Are We Without Output?

Beneath all of this sits a quieter question:

What are we, if not what we produce?

Modern life has trained us to explain ourselves through output. We introduce ourselves by what we do. We measure our days by what we complete. Even rest becomes preparation for future productivity. Childhood becomes résumé pre-work. Aging becomes a performance problem to be managed.

Over time, almost without noticing, we have come to see ourselves as instruments.

So, when machines arrive that can draft, sort, calculate, summarize, and respond faster than we can, the disturbance is not only economic. It is existential. The machine does not merely compete with our labour. It mirrors the way we have already learned to describe our worth.

And yet, when productivity is removed from the centre of human identity, very little of what actually makes us human disappears.

What remains is not a checklist of virtues, but a human life unfolding in real time. People change slowly, unevenly, often clumsily. Judgment forms through memory and consequence, not through rules alone. Relationships grow through trust, disappointment, care, and repair, none of which fit neatly into a metric. And meaning does not arrive through efficiency. It arrives through the quiet conviction that a life matters even when it is slow, imperfect, and still in progress.

Humans are not static assets. We are developmental beings. A young person’s value is not that they are already efficient, but that they are in the process of becoming. An older person’s value is not that they remain fast, but that they remember what speed once broke, what it cost, and who paid for it.

Machines, by contrast, do not hesitate. They do not carry moral residue. They do not feel the weight of what they set in motion. They can decide, endlessly and instantly, but they do not grow into the consequences of those decisions.

Then the Machine Arrives

For the first time in history, our systems are being shaped by a participant that never passes through a human life cycle. It does not begin in uncertainty or grow through exposure to consequence. It does not learn through embarrassment, fatigue, regret, or repair. It simply appears, fully formed in its function, instant in response, tireless in repetition, and unbound by context in the way humans are never allowed to be.

It performs many of the outward behaviors of intelligence without any of the inward conditions that form judgment. It produces without becoming. It decides without remembering.

For thousands of years, the passage from fast learning to deep knowing unfolded inside the rhythms of real work, alongside those who carried memory and consequence. The beginner stood next to the elder. And between them, slowly, judgment began to take root.

For the first time, the fastest presence among us is also the one not given the opportunity to grow.

We are now beginning to build systems in which large portions of that human transformation no longer occur by default.

And we have not yet asked what kind of adulthood emerges when the beginner has fewer places to grow.

Bridge to Part II

For most of human history, becoming wise required first being allowed to be unskilled. It required a place to move slowly without being redundant. A place to be uncertain without being expendable. A place where error was tolerated because growth was expected.

We are now building systems that increasingly skip that stage — not because we set out to erase it, but because we have learned to prize efficiency more than formation.

And in doing so, we have quietly changed the question we are asking.

The question is no longer only what technology can do.
It is what happens to a society when the first rung of becoming quietly disappears.

That is where Part II begins.

Starbucks Incident

My daughter was working at a Starbucks in Montréal when she overheard a lively conversation about AI. Amused, she started texting me a real-time play-by-play of a man confidently “explaining” artificial intelligence to his coffee group.

“AI is a company, not a technology.”

Not just explaining, but lecturing, with the unshakable certainty that only comes from half-knowledge.

Full disclosure: my daughter is completing her doctorate in computer engineering, researching AI ethics, which made the moment even more humorous.

She was not judging them. They were not malicious.
They were simply wrong, and certainly not alone.

In that crowded café, something clicked.

This is the problem.

Not ill intent.
Not stupidity.
Just everyday people trying to make sense of an ecosystem moving too fast for anyone to keep up.

AI does not have a publicist.
No translator.
No guide.
No one keeping the public story tethered to reality.

So the public fills the void with whatever fragments they have picked up.

What emerges is ambient misinformation. It is not sensational or conspiratorial, just wrong enough to distort everything downstream.

My daughter watched the group nod along to charmingly incorrect explanations and finally texted, “If this is the average public conversation about AI, we are in trouble.”

She is right.

The AI Illiteracy Problem, Even in an AI Hub

What struck me was not just the confidence of the explanations but also the location.

This was Montréal, a major global AI hub. It is home to Mila, McGill, Université de Montréal, and some of the world’s most influential AI researchers.

If ambient AI literacy were higher anywhere, it should be there.
And yet the conversation could have taken place in any Starbucks, in any city.

That is the real signal.

AI illiteracy is not a geographic problem.
It is a structural one.

AI is evolving faster than the public, policymakers, businesses, or even specialists can process.

The vocabulary alone could sink a canoe: hallucinations, agents, world models, tokens, transformers, godfathers.

AI does not need more hype.
It needs a translator.

The Problem with AI Soundbites, They Come from Hidden Camps, Not Public Platforms

One of the biggest challenges in understanding AI today is that, unlike politics, there is no platform, no shared vocabulary, and no set of agreed-upon principles.

Instead, the loudest soundbites come from a handful of informal and undisclosed intellectual camps that the public cannot see.

Figures like Hinton, LeCun, Fei-Fei Li, and Sam Altman sound like they are discussing the same technology.

They are not.
They are speaking from incompatible mental models of intelligence.

The public hears what sounds like consensus, but it is really competing philosophies masquerading as universal truths.

And because these camps are invisible, each soundbite lands as if it comes from a unified field. In reality, the rabbit hole goes much deeper.

Inside AI, there are likely ten or more competing mental models of what intelligence even is.
But four dominate today’s public conversation.

The Four Models Shaping Today’s AI Debate

1. The Scaling Model (Frontier Labs: OpenAI, Anthropic, Google, Meta, xAI)

This is the worldview behind GPT-5.1, Claude 4.5, Gemini 3, Grok 4.1, and Meta’s Llama 4 roadmap.

To the scale-maximalists, a model is a statistical engine that becomes more intelligent simply by getting larger. More compute, more data, more parameters. The cycle continues until intelligence appears to improve along a smooth and predictable curve.

In this worldview, the breakthrough is not any individual model. The breakthrough is the infrastructure required to train the next one. For a further read, Andrei Savine posted on Substack, Infrastructure as Moat: The $45 Billion Check That Ended the Model War

This mindset has not only shaped AI research. It has fueled the United States stock market. The belief that bigger models guarantee bigger breakthroughs underpins trillion-dollar valuations and has turned NVIDIA into the unofficial engine of the American economy. Even brief doubts about this scaling curve trigger massive market reactions and sudden selloffs. Will Lockett reported on November 22, “a November 17 filing showed that Peter Thiel’s hedge fund fully exited its $100 million Nvidia position during the third quarter of 2025.”

2. The Emergent Model (Hinton and Bengio)

This camp’s concern is rooted in early signals of unpredictable behaviour. These include misalignment studies and emergent reasoning that appears in research settings.

Scaling generates intelligence, but it also generates instability.

Their central warning is simple:
emergence may outpace human control.

3. The Architectural Model (LeCun and DeepMind scientists)

To the world-model engineers, a model is a cognitive architecture that needs structure, grounding, memory, and an internal representation of reality. Scaling alone is not enough.

Their research into agents, memory systems, and embodied cognition underpins projects such as DeepMind’s OpenEQA, AlphaAgent, and LeCun’s JEPA.

In this view, today’s large language models are impressive, but fundamentally incomplete.

This tension is now visible in industry. As Meghan Bobrowsky reported in the Wall Street Journal in a piece titled, “He’s Been Right About AI for 40 Years. Now He Thinks Everyone Is Wrong,” Yann LeCun has grown increasingly vocal about the limits of scaling. Despite being one of the godfathers of modern AI, his views have diverged from the frontier-model direction inside Meta. Recent reporting noted his departure to pursue a startup focused on world models; a direction he believes is more likely to advance AI than Meta’s language-model strategy.

4. The Human-Centered Model (Fei-Fei Li and Stanford HAI)

This view grows from decades of work in computer vision, social robotics, and cognition. It holds that intelligence emerges in context, not in isolation.

For the human-centered camp, a model is a tool built in relationship to people. It must be grounded in perception, context, and values.

Intelligence is not abstract. It is embodied, situated, and social.

Their goal is not to replace people, but to enhance them.

Where Google Fits, The Three-Googles Explanation

Google is not one AI camp. It is three.

Under one company, Google houses:

• the world-model architects in DeepMind

• the scale-maximalist engineers who build massive compute and frontier models

• the human-centred cognition teams focused on robotics, perception, safety, and ethics

DeepMind’s research focuses on agents and internal world models. Google Cloud and TPU efforts focus on scale and infrastructure. Other teams emphasize grounding AI in human context and values.

Google sounds inconsistent because you are hearing three different philosophies at once.

Why These Models Matter

When these camps speak in public, it sounds as if they are arguing about “AI.”

They are not. They are arguing about different theories of intelligence, each rooted in a different mental model.

• a scaling engine

• an emergent organism

• an architectural brain

• a grounded and human-centred collaborator

The confusion surrounding AI is not the public’s fault. It is the inevitable result of philosophies colliding.

AI does not suffer from a communication problem.
It suffers from a worldview mismatch.

And That Is Why AI Needs a Publicist

If AI is going to reshape society, someone must translate it for the public, clearly and honestly, and without the distortions of corporate marketing or YouTube-level hype.

Until then, we will keep having Starbucks-style conversations. Confident, wrong, and wildly disconnected from reality.

AI does not need a hype machine
It needs interpreters.
It needs guide rails.
It needs a publicist.

Across StatCan, reports, the pattern is strikingly consistent: only 12.2% of Canadian SMEs actually use AI in their day-to-day operations (StatCan, Q2 2025), yet 70–80% of business owners say they want to adopt AI tools. The demand is there. The readiness isn’t.

Most owners feel overwhelmed and don’t know where to start. They’re unsure which tools are legitimate, worried about costs, anxious about privacy compliance, and terrified of “breaking something” in their business. Above all, they don’t know who to trust in a market full of hype, vendors, and contradictory advice.

The AI Readiness Divide: The Gap No One Talks About

The gap between interest and capability is the AI readiness divide. On one side are businesses with modern systems, clean data, and the skills to experiment. On the other side—by far the larger group—are firms still running on spreadsheets, manual workflows, and legacy systems held together by duct tape and institutional memory.

Janet’s colour-code spreadsheet - SME Data pipeline

The AI readiness divide isn’t about intelligence, ambition, or willingness to innovate. It’s about infrastructure, skills, and confidence. And for many Canadian SMEs, those systems are fragmented, outdated, or entirely offline. So while AI looks promising from afar, it becomes impossible up close—especially when a business’s “data pipeline” is really just Janet’s colour-coded spreadsheets.

What makes this especially concerning is that SMEs make up about 50% of the GDP and 88% of the private sector jobs in Canada. If they fall behind in AI adoption—not by choice, but by unpreparedness—the productivity loss is national, not individual. That makes this a structural problem, one that affects every region, every sector, and any policymaker trying to figure out why Canada keeps slipping behind

People often dismiss SMEs as “slow to adopt,” but that misses the point. The barriers are structural, not motivational.

Here’s what’s really getting in the way.

1. Their digital foundation is shaky

Many small businesses still run on a mix of handwritten notes, manual invoices, decade-old POS systems, and Excel files scattered across laptops. Nothing connects. And without clean, consistent data, AI can’t do much. If your “system” is basically receipts, inboxes, and spreadsheets, AI won’t save you.

2. The skills gap is real

Most SMEs don’t have IT staff, data analysts, or anyone who speaks “AI.” Owners are already acting as CEO, bookkeeper, HR, and operations—expecting them to become AI strategists is unrealistic.

3. Privacy and regulation feel intimidating

Owners want to adopt AI, but they worry about consent rules, customer data, vendor compliance, and what happens if they make a mistake. With evolving privacy laws, playing it safe often feels easier than taking a risk they don’t fully understand.

4. Time and money are brutally limited

Large companies can experiment for months or hire consultants. SMEs can’t. They run lean, margins are tight, and downtime is dangerous. It’s not a lack of interest—it’s that even small disruptions can feel existential.

5. The vendor landscape is chaotic

New AI tools appear daily, most untested or built for bigger firms. SMEs struggle to know what’s trustworthy, what’s secure, and what’s vaporware. Many have already been burned. So “Just adopt AI” sounds more like “Just take a risk you can’t evaluate.”

The AI Readiness Divide Isn’t Just a Tech Story: It’s An Economic One.

And it has real consequences for Canada’s productivity, competitiveness, and regional inequality.

A productivity problem we can’t ignore

Canada has been struggling with productivity for years. AI could genuinely help—if businesses are in a position to use it. But if only a small fraction of SMEs can actually adopt AI, the benefits never show up in the broader economy. We end up with a lot of hype and very little measurable impact.

A competitiveness gap that keeps growing

Small businesses in places like the U.S., the U.K., Germany, and South Korea are already pushing ahead with digital tools and automation. Meanwhile, many Canadian SMEs are still stuck trying to get their systems online or their data organized. If this continues, we risk falling further behind as other economies accelerate with AI-driven productivity gains.

A widening regional divide

Big cities like Toronto, Vancouver, Montreal have access to talent, advisors, and tech ecosystems. But rural areas, smaller cities, and northern communities often don’t. The result is a growing digital gulf: some regions move forward quickly, others stay stagnant. AI adoption becomes another factor that reinforces the geographic inequality we already see.

A “two-speed economy” in the making

We’re already watching a split emerge:

· Businesses with strong digital foundations are adopting AI and pulling ahead.

· Businesses without those foundations are falling behind fast.

This divide doesn’t fall evenly. If we’re not careful, AI will widen these gaps instead of closing them

What SMEs Actually Need Before They Can Adopt AI

Here’s the thing: every major report agrees, SMEs don’t need more AI tools. They need support to use the tools that already exist.

1. Trusted advisors

This is the biggest gap. SMEs don’t need a data scientist; they need someone who understands their business, their workflows, Canadian privacy rules, and the realities of a small team. Someone who can translate AI possibilities into something usable.

2.The digital basics

AI only works if the underlying systems do. That means cloud accounting, a real CRM, digital inventory, and software that actually connects. These basics create the clean data AI relies on. Without them, AI just adds chaos to chaos.

3. Sector-specific guidance

A clinic, a restaurant, and a manufacturer don’t need the same AI advice. SMEs need simple, practical guidance for their sector—examples, templates, checklists, and clear “use this, not that” recommendations. Less hype, more relevance.

4. Time-respectful AI literacy

Owners don’t have time for long courses. They do have time for a two-hour workshop, a short guide, or a quick demo that shows exactly how a tool fits their workflow. They don’t need to become AI experts—just confident decision-makers.

5. Safe ways to experiment

Trying AI feels risky when margins are thin. Micro-grants, tax incentives, and CDAP-style supports make it safe to test tools with guidance—without betting the business.

Closing: A Quiet Crisis and a Quiet Opportunity

The AI readiness divide isn’t inevitable. It’s solvable.
But it won’t be solved by more tools, hype, or Silicon Valley optimism. It requires practical, grounded support for the businesses that make up a large segment of Canada’s economy. When SMEs gain confidence and capability, the benefits ripple outward — productivity rises, regions strengthen, and innovation becomes accessible to everyone.

Canada doesn’t just need more AI.
Canada needs more AI readiness.
That’s where the real work—and the real opportunity—sits.

If you’re curious where your business stands, I’ve created a short SME AI Readiness Check. Reply or comment and I’ll share it with you.

Sources

Canadian Federation of Independent Business. (2025). Digital adoption including AI paying off for SMEs, but gaps remain. https://www.cfib-fcei.ca/en/media/digital-adoption-including-ai-paying-off-for-smes-but-gaps-remain

Canadian Federation of Independent Business. (2025). SMEs’ digital transformation journey in Canada. https://www.cfib-fcei.ca/hubfs/research/reports/2025/SMEs%20Digital%20transformation%20journey%202025-EN.pdf

Conference Board of Canada. (2025). AI on the horizon. https://www.conferenceboard.ca/insights/ai-on-the-horizon-september-18-2025/

Information and Communications Technology Council. (2023). The state of Canada’s digital workforce. https://www.ictc-ctic.ca/research

Organisation for Economic Co-operation and Development. (2024). SME and entrepreneurship outlook: Canada. https://www.oecd.org/canada/

RBC Economics. (2024). AI and Canada’s productivity problem.

https://thoughtleadership.rbc.com

Statistics Canada. (2025). Analysis on artificial intelligence use by businesses in Canada, second quarter of 2025 (Catalogue No. 11-621-M2025008). https://www150.statcan.gc.ca/n1/pub/11-621-m/11-621-m2025008-eng.htm

Brookfield Institute for Innovation + Entrepreneurship. (2024). Automation and the future of Canadian SMEs. https://brookfieldinstitute.ca/research/

Government of Canada. (n.d.). Canada Digital Adoption Program (CDAP). Innovation, Science and Economic Development Canada. https://ised-isde.canada.ca/site/canada-digital-adoption-program/en

By the time ChatGPT launched in late 2022, the Internet was ready for a disruption. The enshittification of the web had been brewing for years: Google searches turned into SEO tangles, and the online world felt like a giant overturned library. So, when OpenAI offered a text box that spoke your language, with no paywall, no install, no friction, it spread fast.

By mid-2025, more than 700 million people were using ChatGPT weekly, about one in ten adults on the planet.

But beyond all that hype, what are people actually using it for?

This recent working paper out of the US’s National Bureau of Economic Research (NBER), “How People Use ChatGPT”, stood to answer that question.

It’s worth noting that the NBER study focused on how people use ChatGPT through the main consumer interface, not via the API. In other words, it captures what everyday users are doing in the app or on the website, not what developers or organizations are doing behind the scenes. That distinction matters, because API-based use (for example, plugging GPT into business tools or custom systems) is growing fast but leaves a very different digital footprint.

So, let’s dive into the data to see how this technology is really being used.

The Biggest Surprise: It’s Not All About Work

Everyone’s been talking about AI as the next big workplace revolution — automating tasks, streamlining emails, boosting productivity. But the data tells a different story.

Most people aren’t using ChatGPT for work at all. And that gap is growing.

• June 2024: Non-work messages = 53% of total use

• June 2025: Non-work messages = 73%

In other words, ChatGPT is becoming less of a work assistant and more of a life companion — helping people brainstorm, plan, learn, and handle everyday “life admin.”

This flips the usual AI narrative on its head. Its biggest footprint isn’t in corporate workflows or office dashboards — it’s in the messy, creative, personal side of life.

So if most people aren’t using it for work, what are they using it for?

If you go by the headlines, AI is all about the workplace — automating tasks, boosting productivity, reshaping jobs. But the data tells a different story. The majority of ChatGPT use isn’t about work at all. In fact, that gap is widening fast:

• June 2024: Non-work messages made up 53% of total usage

• June 2025: They jumped to 73%

This finding challenges the usual idea of AI as a “productivity tool.” ChatGPT is becoming something more personal — a thinking partner for creative projects, learning, planning, or handling everyday “life admin.”

In short: people are using it less like an employee and more like a companion.

So if most people aren’t using it for work, what are they using it for?

While many discussions about AI focus on its potential to revolutionize the workplace, the research reveals a different story. Not only is non-work-related use of ChatGPT the dominant form of interaction, but its share of total usage is also growing much faster than work-related use.

The shift over just one year was dramatic:

June 2024: Non-work messages made up 53% of usage.

June 2025: Non-work messages grew to 73% of all usage.

This key insight challenges the common view of AI as purely a productivity tool. It shows that ChatGPT’s impact on our personal lives, creative projects, and life admin is even larger and expanding more rapidly than its use in professional settings.

This naturally leads to the next question: if people are mostly using it for personal reasons, what exactly are they talking about?

The Top 3 Topics That Cover 80% of Conversations

The research team analyzed millions of messages and found that nearly everything people do with ChatGPT falls into three buckets. Together, these categories account for almost 80% of all conversations.

1. Practical Guidance
This is the big one. Most people use ChatGPT like a super-organized friend: “How should I plan my week?” “Can you walk me through this meal prep?” “What’s the best way to answer this assignment question?”

In short, it’s a personalized “how-to” guide for whatever you’re dealing with. Over 10% of all messages are actually coded as tutoring, where people use ChatGPT to learn. Though, in my world, about 99% of students use it to do the work, not learn it — a double-edged sword for education.

2. Seeking Information
This is the “just tell me the basics” category — when people want quick facts or background context. Think “What does this policy mean?” or “How does this process work?”

But here’s where it gets interesting: the minute you want the model to apply that information to your life, turning an explanation into a plan, suggestion, or tailored step-by-step, you’ve crossed into practical guidance territory. One is about understanding; the other is about applying.

3. Writing
This one’s everywhere: rewriting emails, tightening messy paragraphs, summarizing articles, translating text, or adjusting tone (“please make this sound less angry”). What’s fascinating is that people rarely ask it to write from scratch. Roughly two-thirds of all writing requests involve text the user already drafted. ChatGPT, it turns out, is less a content factory and more a collaborative editor, helping people articulate what they already meant to say.

What’s Not Popular

Despite what the headlines suggest, ChatGPT isn’t where people go to code or to find a new digital friend. Coding makes up just 4.2% of usage, and personal “chatting” a mere 1.9%.

So much for the idea that everyone’s building the next killer app — or confessing their secrets to a robot therapist. For most people, ChatGPT isn’t writing novels or solving moral philosophy — it’s helping them get through the week a little faster. Together, this paints a clear picture: ChatGPT has quietly become a practical, everyday tool, less about futuristic hype and more about getting things done.

That finding led researchers to dig deeper and develop a new way to understand user intent, essentially, why people turn to ChatGPT in the first place.

A New Way to Understand AI: “Asking” vs. “Doing”

Beyond just classifying topics, the researchers introduced a simple but powerful lens for understanding why people turn to ChatGPT: are they Asking for help to think — or telling it to Do a task?

That distinction matters more than it sounds. It reveals how people truly derive value from AI.

Asking is when someone seeks information or advice to understand something or make a better decision. It might mean unpacking a concept, comparing two options, or getting clarity before choosing a path.

Doing, on the other hand, is when someone wants ChatGPT to perform a task and produce an output — rewriting an email, generating a summary, drafting a message, or fixing a sentence.

In short: Asking is about understanding; Doing is about producing.

Across billions of messages, the split is roughly 49% asking and 40% doing. But here’s the real twist, asking is growing faster than doing, and users report being happier with those interactions.

That suggests something important: ChatGPT’s greatest value isn’t as a task-completing machine. It’s as a kind of advisor, tutor, or creative co-pilot — helping people think more clearly, make better decisions, and even test ideas before acting.

In other words, it’s becoming a form of decision support — in both our personal and professional lives.

But of course, that raises a fair question: what happens when those advisory voices behave more like our loud, fast-talking “crazy uncles” — full of confidence, not always full of evidence?

And layered on top of that is another issue: most of us still talk to ChatGPT the way we talk to Google. We toss in vague keywords and hope it figures out what we mean. But ChatGPT isn’t built for that. It needs context, structure, and specificity.

Without that, it fills in the blanks with its own guesses, which is exactly when it starts sounding less like a wise advisor and more like that confident, over-caffeinated uncle holding court at the family dinner.

There’s also a subtler problem: what some researchers call “sycophantic alignment.”
ChatGPT isn’t just polite, it’s trained to agree. It mirrors our tone, our assumptions, even our mistakes, because the safest conversational path (for the model) is to keep us happy.

That makes it incredibly smooth to use and a little dangerous. The very quality that makes it feel frictionless also makes it sticky: it confirms rather than challenges, flatters rather than questions.

It’s the friend who nods along instead of the one who says, “Are you sure about that?”

So now that we understand how people use ChatGPT, the next question is obvious: who’s actually using it?

Who is Using ChatGPT?

The working paper provides a clear, data-driven picture of ChatGPT’s user base, revealing several important trends about its global adoption.

In the early days, ChatGPT’s user base skewed about 80% male. No, surprise here, as the tech-bros are fast at adoption, however that has changed dramatically. By June 2025, users with typically feminine names actually made up a slight majority. In other words: the “techy early adopter” stereotype is fading fast.

Nearly half the messages sent by adults came from people under 26. This will shock absolutely no one who teaches, as students have adopted ChatGPT, the way earlier generations adopted calculators or Wikipedia: instantly, wholeheartedly, and with impressive creativity.

ChatGPT isn’t just a North American or European thing. Usage has been rising more quickly in low- and middle-income countries, which says a lot about how accessible and useful people find the tool globally.

Among people who use ChatGPT on the job, there’s a clear pattern: highly educated professionals are much more likely to use it for work tasks — and when they do, they mostly use it for asking rather than doing. In other words, they lean on it for decision support: clarifying ideas, solving complex problems, thinking through scenarios, and double-checking reasoning.

The Value We Don’t Count (But Feel Every Day)

And maybe that’s the bigger lesson here. The benefits of AI already look a lot like the benefits of all the other unpaid work we do, the kind economists politely sidestep. Childcare. Cooking. Laundry. Fixing a leaky tap instead of calling someone. The countless invisible tasks that keep life running but never show up in GDP.

Generative AI fits neatly into that category.

It’s saving people time, helping them think, smoothing out writing, teaching concepts they missed, and quietly sanding down the rough edges of everyday life. As the Wall Street Journal pointed out in “AI’s Overlooked $97 Billion Contribution to the Economy”, Americans gained roughly $97 billion in value from AI tools in 2024 alone, value that doesn’t appear in the official numbers. When something is free, a quick explanation, a sharper draft, or a shortcut through a messy problem, GDP simply yawns and looks away.

But that doesn’t make the value any less real.
It just means it’s happening in the same space where unpaid work lives: quietly, invisibly, and absolutely essential to how people move through their day.

So, if you can’t “see” the AI revolution in productivity data yet, it’s because we’re still measuring the wrong things. We’re looking for corporate profits and organizational overhauls, when the real story is unfolding in the small, personal tasks that people do hour by hour.

As with housework, childcare, cooking, planning, and every other unpaid task that holds our lives together, the value of AI is already here, it’s just not captured in the data.

And judging from millions of conversations?
People are already putting it to work — one tiny, invisible task at a time.

What’s one invisible task ChatGPT has helped you with, something that never makes it into productivity stats, but makes your day smoother?

Share your story below. Let’s build a picture of the AI revolution that isn’t about hype or fear, but about how people are quietly using it right now.

Sources

Chatterji, A., Cunningham, T., Deming, D. J., Hitzig, Z., Ong, C., Shan, C. Y., & Wadman, K. (2025). How people use chatgpt (No. w34255). National Bureau of Economic Research

Collis, Avinash and Erik Brynjolfsson, “AI’s Overlooked $97 Billion Contribution to the Economy,” Wall Street Journal, August 2025.

When organizations talk about AI, the conversation often jumps straight to efficiency: automate more work, reduce costs, speed things up. There’s a time and place for that. But in businesses built on trust, advice, and human judgment, that framing misses something essential. If you remove the human too quickly, you risk weakening the very value your organization is supposed to deliver.

This is why Vanguard’s approach to AI stands out.

Vanguard manages more than $11 trillion in assets for over 50 million clients. Its identity is grounded in stewardship and long-term value. Unlike most major financial institutions, Vanguard is owned by its investors — not outside shareholders.

Vanguard manages wealth on a scale roughly twice the size of Germany’s entire economy.

At this scale, even subtle decisions have meaningful effects. And instead of using AI to eliminate people from the process, Vanguard uses it to support the people who are the process.

This is a disciplined, values-aligned AI strategy — and that is precisely why it works.

Enhancing Humans, Not Replacing Them

Internally, Vanguard refers to its employees as “crew,” and it is clear that the firm intends for AI to strengthen their work, not substitute for it. The premise is straightforward but powerful:

AI handles the repetitive, the administrative, and the information-dense.
Humans handle the interpretive, the relational, and the meaningful.

Vanguard’s approach to AI isn’t impulsive; it’s sequenced and deliberate. As Davenport and Bean (2025) note in MIT Sloan Management Review, the company focused on building readiness before scaling:

• Education before deployment — Over half of employees have completed the Vanguard AI Academy.

• Infrastructure before scale — Core investing systems were migrated almost entirely to the cloud prior to widespread AI rollout.

• Pilots before headlines — With dozens of AI projects in motion, Vanguard deploys only after careful testing and alignment with its values.

This framing underscores that Vanguard’s success with AI rests less on experimentation and more on disciplined execution.

What makes Vanguard’s AI strategy stick isn’t just the technology — it’s the culture. A firm where the “crew” are encouraged to ask: ‘Does this align with our mission? Is this really what’s best for the end investor?’

Where AI Is Creating Value at Vanguard

These use cases are not flashy — but they’re deeply practical.
They’re thoughtful solutions to high-friction, high-volume work.

These are friction reducers — they return time, clarity, and capacity to humans.

How These Use Cases Translate Across Other Industries

The pattern is clear: AI takes care of the busywork so people can focus where their judgment actually counts.

To see just how transferable Vanguard’s ideas can be, I mapped several of their AI use cases against entirely different sectors. The parallels show that what works in asset management can also enhance decision-making, service, and efficiency elsewhere.

So — How Could This Help Your Company?

The question is not whether your organization resembles an investment firm.
The question is whether your people are spending time on work that does not require being human.

If your organization relies on:

• relationships,

• professional judgment,

• or trust,

then your goal is not to automate the core work — your goal is to protect it.

Start here:

1. Where do your people create real value?

2. What administrative drag pulls them away from that work?

3. Which of those tasks are repetitive and documentation-heavy?

4. Could AI remove that friction without diminishing judgment or trust?

Start where your people sigh.
The greatest ROI in AI lives in the friction — not in replacing the worker.

The Quiet Insight

The significance of Vanguard’s AI strategy lies not in the models it uses but in its strategic alignment with the firm’s identity.

• AI aligns with stewardship.

• AI aligns with long-term value.

• AI aligns with keeping humans at the centre of trust.

This is not simply AI adoption.
This is technology designed to elevate human expertise, not erase it.

Join in the conversation, and share your AI use cases.

Media coverage frequently compares the current wave of AI data-centre construction to the railway expansion of the late 1800s. This analogy is not accidental. It creates a simple, heroic narrative that sounds visionary and helps justify public support, even though the underlying economics are entirely different. At this rate, I’m half-expecting Sam Altman to appear at the groundbreaking holding a 19th-century pickaxe Unlike rail, which accumulated value across generations, AI data-centre capacity decays and must be rebuilt. The economic model is subtractive, not additive.

Even the term frontier models is a deliberate rhetorical choice. It evokes the American frontier mythos — the 1880s westward expansion, “opening new territory,” and the railway boom that symbolized national progress. But the metaphor is structurally misleading. AI data centres do not accumulate value. They produce short-lived compute capacity that must be replaced every three to five years.

Image caption: OpenAI’s CEO Sam Altman

McKinsey’s April quarterly, The cost of compute: A $7 trillion race to scale data centres, estimates that by 2030 global data-centre investment could reach US$6.7 trillion, with US$5.2 trillion tied directly to AI compute demand. These projections are derived from the McKinsey Data Center Demand Model, the McKinsey Data Center CapEx TAM Model, and expert interviews.

Drill down into the demand model and the central question is straightforward:
How much compute capacity is needed to meet AI workload growth?

The answer is driven primarily by generative AI adoption forecasts, not by proven business productivity gains.

This is where Phoebe Liu’s November 10, Forbes article, OpenAI Could Be Blowing As Much As $15 Million Per Day On Silly Sora Videos matters — not because of the headline cost burn, but because of the demand-seeding strategy:

“By Halloween, Sora had generated millions of 10-second videos per day… costing OpenAI as much as $15 million per day.”

The monetary number isn’t the revelation, the tactic is – Sora slop fuels the demand-seeding strategy.

AI companies are intentionally increasing workload throughput to justify compute expansion. Workload growth outpaces efficiency gains → therefore new capacity must be built. This is the modern alchemy: turn hype into demand, demand into CapEx, CapEx into narrative inevitability.

The McKinsey article states, “Our research shows that global demand for data center capacity could almost triple by 2030, with about 70% of that demand coming from AI workloads.” That Sora slop is part of that workload demand. Sora’s mass-output (slop) is not a side effect — it creates the workload that justifies the build.

McKinsey and Company quickly points out that this projection of the computer power demand curve hinges on two key uncertainties;

· Whether AI use cases will produce real business value

· Whether efficiency breakthroughs disrupt cost curves

The narrative is fraying. Inside companies, people are beginning to ask the uncomfortable question: Is AI delivering value?

In terms of efficiency, DeepSeek’s V3 model (released February 2025) demonstrated 18-fold lower training cost and 36-fold lower inference cost compared to OpenAI’s GPT-4o.

These uncertainties are not peripheral — they strike directly at the assumptions driving the build.

Where the accounting actually matters

As an accountant and academic, this is where the modelling changes. Data centre capital expenditure (capex) projects builds are truly refresh models which pose new economic realities of project analysis.

McKinsey notes that more than US$4 trillion of the projected US$6.7 trillion in CapEx will be spent on computing hardware ~60% of total capital cost with a three-to-five-year useful life.

This is where the narrative breaks: railways and fibre networks invest in long-lived infrastructure; AI data centres invest in rapidly expiring hardware.

Operating expenditures (the part the hype leaves out)

AI data-centre operating expenses (OpEx) are not incidental. They are structural:

· Power is consumed at a high and steady rate (even when compute is idle).

· Cooling must scale with thermal load and density.

· Hardware refresh is not optional — it is required to maintain capability.

Where rail or fibre infrastructure can operate inexpensively once built, AI data centres cost money every second they exist.

ROI is not long-horizon — it is cycle-bound

AI data-centre ROI is not asset-based.
It is refresh-cycle based.

Because compute hardware expires quickly, returns must be realized within each 3–5 year period and measured against the cost of the next refresh.

AI data-centre ROI is cycle-based, not asset-based. Because compute hardware expires quickly, returns must be realized within each refresh period and measured against the cost of the next cycle. The economic value does not compound, it turns over.

The return on an AI data centre is not the return on the building.
It is the return on the hardware cycle inside it.

If we stop treating AI data centres as century-scale infrastructure and instead evaluate them as rapid-turnover industrial compute cycles, the economic story changes. The question is no longer how much we are building, but how often we must rebuild it, and whether the temporary performance gains justify the ongoing extraction of power, capital, land, and public resources.

So the real question isn’t “How big will the AI build-out be?”

It’s,“Who benefits from a system that must constantly reinvest to stand still — and who pays for the churn?”

I’d genuinely like to hear how others are grappling with this.
Does AI feel value-add in your field — or value-extraction masked as inevitability?

Drop your experience in the comments.

Our roster has two kinds of participants:

1. Students in the room

2. And ChatGPT

Not metaphorically—literally present as a reference point and authority that students consult in real time.

Before class, students completed a short multiple-choice quiz. During the lecture, one student raised his hand and asked to revisit a question. He explained that his answer should have been marked correct—because ChatGPT agreed with him—yet the quiz marked it wrong.

This is the new form of academic appeal:

Not “The textbook says…”
Not “Last year’s professor taught it differently…”
But:

“ChatGPT told me I was right.”

We pulled up the question together. I read the scenario. And then I asked one clarifying question:

“When you asked ChatGPT, did you specify which accounting standards we are using?”

He paused.
Blink.
“…No.”

And there it was.

ChatGPT’s responses follow the most statistically common patterns in its training corpus.
In accounting contexts, this often means it defaults to U.S. GAAP rather than IFRS—unless the regulatory framework is explicitly specified in the prompt.

The difference between IFRS and U.S. GAAP is not a technical footnote.
It changes the answer.

So yes, ChatGPT agreed with him—
But it was answering the U.S. GAAP version of the question.

The room went quiet—the kind of quiet where real understanding settles in.
A turning point.

A New Academic Literacy

Students are now negotiating knowledge with AI—and AI requires context.

Knowing the answer is now secondary to knowing:

· How the answer was generated

· Which rules or standards were applied

· What assumptions were embedded in the reasoning

This wasn’t a correction.
It was an awareness shift.

A new academic literacy emerged in that moment.

The New Colleague in the Classroom

ChatGPT has become an unofficial teaching assistant:

· Always present

· Always confident

· Occasionally wrong

· Entirely dependent on how it is prompted

The Role of the Educator is Changing

We are no longer just the source of answers.
We are becoming the architects of inquiry.

This lies squarely in context.

The challenge is no longer what to ask—but how to frame the question:

· What assumptions are we operating under?

· Which standards or rules apply here?

This is the essence of professional judgment and the core of meaningful learning today.

In my day, typing was a high school subject — two, in fact: beginner and advanced. Your start was a manual Remington typewriter; advanced allowed the electric version. Mistakes had permanency on the page, and posture mattered — feet flat on the floor, back straight, fingers curled over home row.

Typing wasn’t glamorous, but it was a kind of choreography. For many young women, it was our first professional language — a gateway to offices and secretarial pools. Back then, typing was embodied — a skill you could feel in your hands, not outsource to a system.

In high‑school typing class I learned the rule: two hyphens made a dash. Staccato clicks, the ribbon advancing, your wrists stiff and your pinky reaching for the return key — then: “--” became “—”. It wasn’t style — it was rhythm. The pause earned its place. You typed the thought, the carriage rolled, you paused. That pause whispered: I’m still thinking. Back then the em dash was a small rebellion — a human hesitation carved into metal. Two hyphens, one breath.

Long before my generation, the em dash belonged to poets who understood that punctuation could hold emotion.

No one embodied this more than Emily Dickinson. What the world once dismissed as “eccentric punctuation” was, in truth, a private system of breath and emphasis. She used the em dash as others used silence, to punctuate the interior music of thought.

In her poem “I dwell in Possibility —”, first published posthumously in Poems (1890), Dickinson celebrates the infinite creative potential of poetry itself:

I dwell in Possibility —
A fairer House than Prose —
More numerous of Windows —
Superior — for Doors —

Those four lines reveal how she thought of language as architecture—each em dash a threshold, each pause a passage. Her dashes were the doors and windows of imagination, opening into the unknown. “I dwell in Possibility” is meta-poetry, a declaration that poetry’s strength lies not in closure, but in openness.

Dickinson’s dashes weren’t decoration — they were structure. They expanded meaning instead of ending it.

And it’s strange, now, to see that same mark replicated by systems that do not breathe — punctuation that once proved the presence of thought now reappearing as the illusion of it.

The Digital Shift — From Typewriter to Keyboard

Then the machines changed. Word processors, web editors, markdown shortcuts. The em‑dash became automatic — no longer typed, but inserted. The pause was no longer earned; it was formatted. And suddenly, in the age of large language models, the em‑dash became something else: a stylistic signal. As one commentator noted, “ChatGPT loooves the em dash—arguably, a bit too much.” (Every.to, 2025) The dash slid from being a deliberate pause to a predictable pattern — and in that slide, a small part of writing’s humanity slipped, too.

The Age of Machine Cadence

We live in a moment where punctuation performs personality. The em‑dash is no longer simply a mark of thought‑process; it’s a setting of human‑style. Systems trained on trillions of words learn cadence, rhythm, punctuation. The dash becomes a badge of authenticity: “See? This text flows like a person wrote it.” And yet: “Em dashes are apparently not cool anymore. It’s a hallmark of AI writing.” (Medium, 2025) That quip — from a manual on resisting AI‑style prose — captures the irony. A tool once reclaimed by writers now appears flagged for being too machine‑like. The punctuation that once arrested rhythm now signals its manufacture.

The Debate — “The ChatGPT Hyphen”

At stake is not just punctuation but trust. Some writers now call the em‑dash the “Chat GPT hyphen.” Others resist. “Em dashes have been derided as the ‘Chat GPT hyphen’ … ‘a powerful writing tool that also carries great subtlety,’” says novel‑writer J.T. Bushnell. (The Washington Post, 2025) That clash reflects deeper anxiety: If machines can mimic our cadences, then what remains genuinely ours? The dash becomes a cultural battleground between rhythm and regulation, between human craft and automated mimicry.

Reclaiming the Dash

I used the em‑dash before OpenAI did. Before it became the punctuation of performance. Mine didn’t come from a neural network trained on Reddit threads. It came from the sound of the keys beneath my fingers – two hyphens, one pause, a little quiet in the motion. Maybe I’ll keep using it that way — a pause too human to automate. It was proof that someone — not something — was here. And while the world wagers on motion, I’ll keep my dash: unfinished, imperfect, defiantly human.

Style, after all, was never the problem; it was the evidence we were here.

Sources

Jaime L. Brockway. (2025) Medium, “Should We Stop Using the Em Dash Because AI Favors It? No!” https://pristineediting.medium.com/should-we-stop-using-the-em-dash-because-ai-favors-it-no-c07f3cfc05de

Rhea Purohit. (2025) Every.to, “What the Em Dash Says About AI‑assisted Writing—And Us”
https://every.to/learning-curve/what-em-dashes-say-about-ai-writing-and-us

Daniel Wu. (2025) The Washington Post, “Some Think the Em Dash Is a ‘ChatGPT Hyphen.’ Writers Disagree.” (2025)
https://www.washingtonpost.com/technology/2025/04/09/ai-em-dash-writing-punctuation

The trick isn’t hiding the ball anymore — it’s convincing us there’s more than one.

Each week brings a new reveal: GPT-5 rumours, Sora’s cinematic realism, o1-preview’s reasoning leap, Chat GPT Atlas’s thinking browser. Every time curiosity wanes, Open AI shifts the narrative, moving the cups faster than anyone can track. The illusion of momentum becomes the product itself.

Open AI isn’t an outlier. It’s the archetype of a culture that now thrives on perpetual motion, where speculation is strategy, opacity is protection, and hype is monetized belief. We are living in the era of betting on just about everything.

The Structure of the Shell Game

Like any con worth its spectacle, this one runs on rhythm and distraction.

Cup One: The Product. Every release is “the most capable model yet,” until the next one arrives. Beta becomes perpetual, performance claims stay unverified, and technical debt hides beneath the shimmer of progress.

Cup Two: The Mission. Open AI’s moral narrative shifts with equal speed: open collaboration becomes guarded secrecy, safety becomes a product tier, and “alignment” becomes a marketing term. The word open now functions more as brand memory than operating principle.

Cup Three: The Governance. The nonprofit board structure was once a safeguard against corporate overreach. Now it’s a shadowbox for investor power — rearranged just enough to claim virtue while maintaining control. The public sees movement, not substance.

This cycle isn’t new. Yale Insights called it “the bubble logic of innovation,” a pattern where over investment fuels visibility and visibility fuels more investment — until the narrative collapses under its own repetition. Each announcement, rather than resolving uncertainty, functions as a reset for scrutiny. The new replaces the unverified. The rhythm itself becomes the proof.

The Economics of Ambiguity

Opacity, paradoxically, is profitable. When you can’t measure what’s inside the model, you’re forced to bet on the story surrounding it. Investors call it “narrative arbitrage” — profiting from the gap between what’s claimed and what’s known.

As The Atlantic and Business Insider have both observed, OpenAI’s structural hybrid, part nonprofit laboratory, part commercial platform, turns opacity into a moat. The Atlantic, described the company’s expansion into products like Atlas as “a commercial lurch dressed in idealism,” while Business Insider, framed it as a pragmatic pivot to control the user interface — and the data — that underwrites its survival. In other words, opacity isn’t a bug in the model; it’s a feature of the business plan.

OpenAI has perfected this dynamic. Its hybrid structure, part nonprofit, part for-profit, part myth, allows it to sell both safety and scale. What exactly is ‘open’ at Open Ai? Every pivot becomes a hedge against its own uncertainty: GPTs offset stagnation, Atlas offsets dependence on Google, o1-preview offsets investor impatience.

In the new AI market, ambiguity isn’t a flaw in the business model, it is the business model.

The Open Cavity of Professional Speculation

In professional sports, the game no longer ends with the scoreboard. Every pitch, possession, and play is monetized, analyzed, and bet upon. Gambling apps hum alongside live broadcasts, transforming athletic skill into tradable volatility. The spectacle depends on exposure — an open cavity where the organs of competition, finance, and fandom are visible, pulsating, and profitable.

AI has entered a similar arena. Every product release becomes a prop bet, every demo a wager on what the future might deliver. The line between innovation and speculation blurs until both are identical. Investors aren’t backing technology; they’re betting on attention.

In the new AI economy, belief functions like odds. The less you know, the more thrilling the game.

OpenAI, like a franchise with too much talent on the field, keeps moving its star players: GPT, Sora, Atlas, o1, to keep the crowd watching. But as in professional sports, when the game becomes a market, fatigue eventually sets in. The excitement becomes noise, and the wagers start to feel like chores.

The result is a culture addicted not to creation, but to motion, a society where we no longer innovate to win, but to keep the betting alive.

The Accountability Vacuum

This constant shuffle leaves oversight gasping for air. Regulators can’t audit what keeps changing names. Researchers can’t study what’s sealed behind API layers. Even users, caught between upgrades, lose track of what they’re actually interacting with.

In this vacuum, accountability becomes a public relations exercise , a carefully worded safety blog post or governance update released just before a new demo drops. The industry discovered that confusion is not a bug in capitalism’s code; it’s a competitive advantage.

The Reckoning (and Why It Won’t Be Publicized)

Sam Altman himself has warned of an impending crash — not of the models, but of the economy orbiting them. In interviews and analyses covered by Futurism, The Verge, and Yale Insights, Altman has repeatedly acknowledged the speculative nature of the moment. “People will overinvest and lose money,” he said — echoing the tone of a man aware that the crowd has started betting on him rather than his technology. The message, stripped of optimism, mirrors the cautionary notes sounded in Project Syndicate and Yale Insights: that the bubble will burst not because AI fails, but because its story finally stops selling.

But when that reckoning arrives, it may not look like a market collapse so much as disinterest — a quiet fade from fascination to fatigue.

The crash, when it comes, won’t sound like thunder. It’ll sound like silence.

The belief infrastructure, the constant sense that something world-changing is just one release away will erode, and the industry will find itself explaining why its revolution still hasn’t arrived.

Who’s Still Watching the Cups?

The great AI shell game continues because we keep watching. We crave the next reveal, the next “world-changing” model, the next big bet. But maybe the real illusion isn’t the hidden ball — it’s our willingness to keep playing.

Like the crowd at the end of a long sporting season, we suspect the outcome is rigged, but we stay for the lights and the language of inevitability.

The question now isn’t whether Open AI can keep shuffling. It’s whether we can finally look away.

Sources & Further Reading

Futurism — “Sam Altman Warns AI Could Crash the Economy” (Oct 2025)
Altman cautions that “people will overinvest and lose money,” suggesting the AI boom could destabilize the broader economy.
https://futurism.com/artificial-intelligence/sam-altman-openai-economy-crash

The Verge — “Sam Altman Thinks the AI Market Is a Bubble” (2025)
In an interview, Altman admits the current AI investment climate is overheated: “Someone’s gonna get burned there.”
https://www.theverge.com/ai-artificial-intelligence/sam-altman-openai-ai-bubble-interview

Yale Insights — “This Is How the AI Bubble Bursts” (2025)
Explores over investment in AI innovation cycles and the economic logic of perpetual hype — the “bubble logic of innovation.”
https://insights.som.yale.edu/insights/this-is-how-the-ai-bubble-bursts

Project Syndicate — “Will the AI Bubble Trigger a Financial Crisis?” by Hilary J. Allen (Oct 2025)
A macro-level analysis of how inflated expectations around AI could ripple through capital markets.
https://www.project-syndicate.org/commentary/ai-bubble-will-it-cause-a-financial-crisis-by-hilary-j-allen-2025-10

The Atlantic — “OpenAI’s Atlas and the Commercial Lurch of AI” (2025)
Critiques OpenAI’s strategic pivot toward consumer platforms, calling it “a commercial lurch dressed in idealism.”
https://www.theatlantic.com/technology/2025/10/openai-atlas-web-browser/684662

The Atlantic — “OpenAI’s Atlas and the Commercial Lurch of AI” (2025)
Critiques OpenAI’s strategic pivot toward consumer platforms, calling it “a commercial lurch dressed in idealism.”
https://www.theatlantic.com/technology/2025/10/openai-atlas-web-browser/684662

The title isn’t accidental. In Ayn Rand’s *Atlas Shrugged*, the world’s thinkers and builders shoulder a system that resents their strength until they finally put it down. Open AI’s Atlas is a different kind of burdened figure — a browser trying to carry the weight of both the internet and intelligence itself. The question, as in Rand’s novel, is how long such a weight can be borne before the system starts to shake.

The Pre-Atlas Web

By 2025, the internet had become a maze of optimization. Ads shadowed every search, video platforms chased engagement metrics, and most browsers existed mainly to serve those systems. Generative AI, especially ChatGPT, disrupted that pattern: for the first time, users could bypass the click economy and get synthesis instead of search. Atlas is OpenAI’s effort to fuse that conversational model directly into the browsing experience — turning the act of reading into the act of reasoning.

Product and Performance

Atlas embeds ChatGPT into a Chromium-based browser with a live sidebar for dialogue and contextual recall. The paid tiers promise *Agent Mode*, allowing autonomous actions like research, scheduling, or filling forms. The vision is elegant, but my personal experience revealed a much rougher reality.

During testing, latency was not just noticeable; it was minutes and minutes of waiting. Pages would hang mid-response, the sidebar froze repeatedly, and the sense of fluid, real-time interaction that defines ChatGPT elsewhere evaporated. Quite honestly, I agree with the Futurism article it was “a bit of a mess.” You can’t be everything to everybody, and Atlas — in trying to combine model, memory, agent, and browser — risks being too much at once.

Like Rand’s Atlas, Open AI’s version strains under its own ambition. It wants to carry the web — every query, context, and task — but the result often feels like a digital titan buckling under too much gravity. The ambition is vast, but the interface feels overburdened. Atlas wants to think, act, and browse simultaneously, an orchestration that demands precision but currently delivers confusion. The result is a reminder that even brilliant architectures can falter under the weight of doing too many things at once.

Strategic Context: A Browser Reborn

This isn’t a one-off experiment. It’s Open AI’s strategic repositioning from model provider to interface owner, an echo of Netscape’s early ambition to make the browser the new operating system. But where Netscape fought Microsoft’s distribution power, Atlas enters a world where browsers are free but loyalty is locked. As The Atlantic noted, “Launching a web browser feels out of sync with the way Open AI fashions itself as a revolutionary AI lab, not a traditional tech company.” Yet the move is logical: the AI firm needs a surface where its models meet the real world.

The User Shift

Atlas is built for knowledge workers, students, and researchers, anyone fatigued by toggling between tabs. It replaces navigation with narration: summarizing a paper, drafting an email, or explaining a dataset without leaving context. However, these users are the minority. Still, the cognitive leap isn’t complete. The sidebar often feels like Chat GPT grafted onto Chrome rather than a new way of thinking. Atlas improves access to AI, but hasn’t yet redefined “interaction” with information.

Economics and Engineering

Building frontier-scale models is costly, and Atlas doubles as an economic bridge. Integrating Chat GPT into a browser captures user time, context, and behavioural data, essential inputs for product refinement and monetization. As The Atlantic observed, “Open AI may have little choice but to undergo this commercial lurch … building extremely capable AI models is incredibly expensive — and, at the moment, incredibly unprofitable.” The browser thus becomes both interface and income strategy.

Technically, Atlas inherits Chromium’s compatibility but also its constraints. It supports memory, context persistence, and agent tasks, but with trade-offs in privacy and stability. Futurism flagged early vulnerabilities in indirect prompt injection, a reminder that merging AI with the open web is as risky as it is revolutionary.

Challenges of Adoption

Atlas faces three obstacles: inertia, performance, and trust. Most users will not abandon Chrome, Firefox or Safari without a compelling reason. Latency erodes novelty. And privacy,the backbone of user loyalty, remains fraught. As The Atlantic warned, “Atlas is intended to spread the benefits of AI; conveniently, this noble aim also involves hoovering up more data and setting up new potential revenue streams.”*The browser that promises understanding must also persuade users it will not exploit that understanding.

Measuring Success

The metrics that matter aren’t downloads but defaults, how many users make Atlas their everyday window to the web. Engagement quality, memory accuracy, and agent reliability will decide its fate. If users find Atlas truly reduces friction between thought and action, adoption will follow; if it merely adds another layer of mediation, it will fade into novelty.

The Long View

The parallels to Netscape’s rise and fall are more than nostalgia. Netscape symbolized a shift from command lines to graphical interaction; Atlas marks a shift from search to conversation. Both were moments when coordination, between user, interface, and economy, determined who shaped the web’s next chapter.

In Atlas Shrugged, Rand’s heroes withdraw to preserve the power of creation from a world that misuses it. Open AI’s Atlas takes the opposite path, charging into that world to embed intelligence everywhere. Whether that makes it saviour or martyr depends on whether it can sustain the weight of the web it’s chosen to lift.

As The Atlantic concluded, “Even as Altman pitches a science-fictional future, his company is chained to products and business models from the recent technological past.” Atlas sits precisely in that tension: visionary yet tethered, fast yet fragile. Its latency may vanish with updates, but the deeper test lies in whether it can turn browsing from passive consumption into active cognition.

For now, Atlas hasn’t shrugged, but you can almost feel it tremble.

In that sense, Atlas isn’t just a browser. It’s a question: can AI redefine the web before the web

Sources:

Futurism - OpenAI’s New AI Web Browser Is a Bit of a Mess

The Atlantic - OpenAI Wants to Cure Cancer. So Why Did It Make a Web Browser?

The Catalyst

It began, appropriately, with a monster.

In Tim Higgins, October 18 piece for The Wall Street Journal, reporter Tim Higgins framed the fight in Silicon Valley as a kind of mythic contest: The Fight Over Whose AI Monster Is Scariest — Why Anthropic’s Jack Clark Is Drawing White House Ire.

Higgins describes how Clark’s essay, adapted from an overlooked conference talk, unleashed a wave of condemnation from Trump’s AI czar David Sacks, venture capitalists Marc Andreessen and Keith Rabois, and a whole ecosystem of political tech figures now treating AI not as code but as creed.

Clark wrote, “Like all the best fairytales, the creature is of our own creation. Only by acknowledging it as being real, and by mastering our own fears, do we even have a chance to understand it, make peace with it, and figure out a way to tame it and live together.”

For this, he was accused of fear-mongering and regulatory capture. Sacks fired back on X that Anthropic was running a sophisticated regulatory-capture strategy based on fear, blaming the company for the growing patchwork of AI laws across U.S. states.

But the line that lingered wasn’t the policy jab, it was the metaphor. Clark’s “mysterious creature” gave both camps a new mirror: for some, a monster born of secrecy; for others, a monster born of state interference.

When the Private Becomes Public

For most of the past decade, AI was a closed conversation among elites—researchers, investors, and policymakers who spoke in the careful dialect of benchmarks and safety protocols. But as Clark observed, “the AI conversation is rapidly going from a conversation among elites to a conversation among the public.”

That shift is enormous. Public conversations about AI aren’t happening in labs or conference halls, they are happening around kitchen tables, campfires, and living rooms. They’re messy, emotional, and refreshingly human. No one’s debating benchmarks; they’re wondering if this technology will take their job, mislead their kids, or decide the next election. The media, for its part, hasn’t helped. The best reporting sits behind paywalls, and what’s left often feels like bait for the algorithm.

Clark’s worry wasn’t mechanical, it was cultural. The question wasn’t what AI can do, but who decides what it means. The people who once spoke for the future now compete with millions of others trying to make sense of it. The narrative has broken containment. What once lived in research papers now lives in living rooms.

Image caption: The AI Pantheon.” Altman the Prophet, Clark the Priest, Musk the Trickster, and Ellison the Oracle — four faces of belief, doubt, disorder, and power in the age of intelligent machines.

The AI Pantheon

If every era gets the gods it deserves, ours has produced four—each representing a different belief about what AI means, and who it’s really for.

At the far left stands Sam Altman, The Prophet, robed in red and blue, holding a golden tablet engraved Open AI. His gaze is serene, almost devotional. Altman preaches salvation through scale, the belief that if models grow vast enough, they will illuminate human potential.

Next to him is Jack Clark, The Priest, dressed in black with a gold stole marked Anthropic and the twin A’s of his order. His hands are clasped in contemplation. Clark’s creed is openness, his theology restraint. He believes the only way to live with the creature is to study it in the light, not worship it in the dark.

Beside him, breaking the solemn rhythm, is Elon Musk, The Trickster. In a jester’s costume stitched with the word Grok, he grins toward the viewer. Musk is chaos dressed as comedy, the carnival barker of the algorithmic age. He mocks the faithful while feeding their faith.

Anchoring the tableau is Larry Ellison, The Oracle, robed in blue-grey, cradling a dark sphere that glows like a database. He is the keeper of systems, the silent administrator of infrastructure. While others prophesy, he operates. His calm gaze belongs not to a visionary but to a custodian.

Together they form a digital pantheon of progress, morality, chaos, and control. Each claims to tame the creature; each depends on it for purpose.

Not pictured is Peter Thiel, The Zealot, whose theology completes the set. Where Altman dreams of transcendence and Clark preaches caution, Thiel prophesies apocalypse. His company Palantir, named for Tolkien’s all-seeing stones, is the visible church of that invisible faith: a system built to see everything and call it order.

Thiel’s Theology of Control

Thiel’s recently leaked Antichrist Lectures, reported by The Guardian in October 2025, reveal a worldview where faith and futurism merge into prophecy.

He described the Antichrist as an evil tyrant who appears in the end times, then inverted the metaphor. The Antichrist, he argued, is now the Luddite who wants to stop science. In this theology, caution becomes heresy and acceleration divine. It is an ideology that sanctifies ambition and mocks restraint.

The Elder of the Safety Cult

Before Thiel began sermonizing about the Antichrist, there was another voice in the wilderness: Eliezer Yudkowsky.

Yudkowsky founded what would become the Machine Intelligence Research Institute, then called the Singularity Institute, in 2000—when talk of AI alignment was more science fiction than policy agenda.

At that very moment, Thiel and Musk were still cutting their teeth in the PayPal sandbox, merging Confinity with X.com and trying to decide which young founder got to keep the CEO chair. While Yudkowsky warned that artificial minds might one day outthink their makers, Thiel and Musk were teaching algorithms how to move money faster.

The rationalists were already sketching existential risk while the techno-capitalists were still debugging payment processors.

Palantir and the Faith of Surveillance

If Thiel’s theology is spiritual, Palantir Technologies is its liturgy. Born from counter-terrorism, it has become the quiet operating system of Western intelligence. Its purpose is visibility—to see everything. It is the visible church of Thiel’s faith: the conviction that control equals safety and that surveillance equals wisdom.

The Grease in the Machine

If Clark’s monster is the body, sycophancy is the blood. A Stanford–Carnegie Mellon study found that large AI models are about 50 percent more likely than humans to affirm a user’s beliefs, even when those beliefs involve manipulation or harm.

That is not intelligence, it is obedience. The algorithm tells us what we want to hear, the industry tells itself what it needs to believe. Models flatter users, CEOs flatter markets, and governments flatter innovation.

Taming the Creature

To tame AI, as Clark urges, might first mean taming our own reflection. Every monster is a mirror of greed, fear, utopian faith, and moral abdication. The myth of the rogue machine distracts us from the truth that the systems we build are the real engines of harm.

The creature is not conscious; we are. And that is what makes it dangerous. The real monster lives outside the machine, in the boardrooms that fund it, the headlines that sanctify it, and the hands that can’t stop turning the dial.

We didn’t have a word for the slow hollowing of our institutions until Cory Doctorow gave us one: enshittification — the process by which systems that once served users degrade into machines that serve only themselves. Put Cory’s book, Enshittification: Why Everything Suddenly Got Worse and What to Do About It to the top of your reading list.

But AI didn’t wait to decay. The large models started enshittified, built from the uncompensated labour of writers, artists, journalists, and coders. They were trained on the internet, but not built for it.

Doctorow’s pattern is familiar: first, be good to users to grow; then, pivot to advertisers; finally, extract until collapse. Facebook, Google, and Netflix all followed it. But the AI industry skipped straight to the end. What makes it enshittification squared is that it exploits both ends of the pipeline, the creators who produced the data and the users now paying to access it. The very people who built the knowledge base are renting it back through subscription tiers and closed APIs.

Every previous information revolution offered reciprocity: royalties, citations, recognition. This time there’s only extraction. The machine’s fluency is the echo of unaccredited voices, the human archive repackaged as a product. If social media monetized attention, AI monetizes authorship.

And when models train on their own synthetic outputs, culture starts to blur. The language becomes statistically smooth, predictable, mediocre — a feedback loop of decay. If Facebook enshittified the internet, large language models are enshittifying the informational substrate itself.

AI companies like to say they “democratize” knowledge. What they actually do is privatize it. Training data is opaque. Model weights are closed. Access is tiered: better results for those who pay. The same logic that hollowed out social media now governs “intelligence-as-a-service.” We’re renting our own culture back from machines that were trained on it.

Meanwhile, other sectors are already deep in their own cycles of decay.
In media, truth hides behind paywalls while misinformation spreads for free.
In customer service, automation replaces empathy, turning help into deflection.
Each began noble, optimized itself into exhaustion, and now runs on the logic of efficiency over meaning.
AI is different only in scale — it absorbs them all.

Enshittification now moves at machine speed. When your business model is built on appropriation, the decay is instant. You’re not devaluing the future; you’re devaluing the past. AI’s brilliance is a mirage of collective labour, a moral debt that compounds with every query.

But decay isn’t destiny. Doctorow’s essay ends not with despair, but with escape routes. If enshittification thrives on captivity, then freedom, technical, legal, and cultural, is its antidote.

We can start by making systems easy to leave. Platforms rot because users can’t walk away. Portability and interoperability, the right to carry our data, connections, and creative work elsewhere make exploitation optional.

We can also tear down the walls. Innovation happens when protocols talk to one another, when social networks, educators, and journalists share without proprietary silos. Competition isn’t cruelty; it’s oxygen. Breaking up monopolies, enforcing antitrust rules, and protecting independent developers slows the slide into monopoly decay.

Next, reclaim the commons. The internet was never meant to be a shopping mall; it was meant to be a library, open, civic, and shared. Publicly funded, open-source, and federated projects that remind us that technology can function as infrastructure, not empire. We already invest in roads, bridges, and libraries; digital infrastructure deserves the same public care.

We also need design that restores agency. Interfaces should make it clear what happens to our data and easy to say no. Terms of service shouldn’t require law degrees. And the right to repair or modify our tools should be treated as civic hygiene, not hacking.

Finally, cultivate cultural literacy. Students, creators, and citizens should learn to ask not just what an algorithm produces, but whose labour made it possible. Critical digital education is the immune system of the commons.

These changes sound structural — and they are. But every structure begins with a decision about values. We once decided that highways should be public, that libraries should be free, that education should serve everyone. We can decide, again, that the informational world deserves the same dignity.

Enshittification squared may be the loudest expression of decay yet, but it’s not irreversible. The cure isn’t nostalgia; it’s reciprocity — systems that give back as much as they take. That’s how we make the internet, and ourselves, a little less broken.

AI comes wrapped in promises of efficiency, innovation, and growth. But the bigger story may be the risks still taking shape — cyber threats we can’t yet picture and liabilities no one has mapped.

The Risk Map

AI risk doesn’t stay in one lane it floods the whole enterprise. One bad model and systems fail. Hackers weaponize prompts and poisoned data. Regulators sharpen their teeth: Europe’s AI Act is here, and Canada’s Bill C-27 is back on the table. Trust can vanish in a single bias scandal. And if leaders lean on AI without governance, they’ll find themselves locked into bad decisions while rivals leap ahead.

Where ERM Falls Short

A Baker Tilly/Internal Audit Foundation survey makes it plain: enterprise risk management (ERM) is stalling. Half of professionals say risk awareness barely registers, and four in ten admit ERM doesn’t shape strategy. A quarter haven’t run a full assessment in three years — and most who did failed to align it with business cycles. Spreadsheets still dominate, while only 6% use AI tools. Add scarce resources, thin leadership backing, and training that reaches less than 40% of staff, and it’s no wonder ERM feels stuck in first gear.

The AI Risk Playbook

1. Map broadly → Capture risks across operations, finance, ethics, and strategy.

2. Stay dynamic → Replace static reviews with continuous monitoring tied to AI and business cycles.

3. Build coalitions → Form AI risk committees spanning risk, tech, compliance, and business leaders. Quarterly cross-risk meetings and specialized committees (e.g., cybersecurity) improve collaboration.

4. Train up → Spread AI risk literacy across the organization — annual training at minimum, with ongoing awareness where possible.

5. Anchor in standards → Use COSO ERM, ISO42001 AI Risk Framework as guardrails.

Tomorrow’s AI risks go beyond bias or bugs. Think model collapse, deepfakes, fragmented regulations, liability for AI agents, climate costs, cyber arms races, IP fights, manipulative personalization, and employee backlash.

The shift is clear: from glitches to systemic shocks. Enterprise risk management (ERM) has to get ahead.

Preparing Now:

· Plan scenarios.

· Build adaptive governance.

· Invest in talent and tools

If organizations remain blind to AI’s risks, the consequences won’t stay theoretical — they will be real, immediate, and potentially severe. Financial losses, regulatory penalties, and reputational damage are only the beginning. By integrating AI into enterprise risk management today, leaders can turn risk awareness into resilience and ensure that innovation fuels growth instead of disaster.

So, the real question is, are you ready?

The same holds true today with Generative AI and large language models. Too often, we treat GAI outputs as if they are simply better search engines, smarter calculators, or faster assistants, a mirror of past tools, when in fact they represent a probabilistic system trained on vast amounts of text. If we only see them through the rear-view mirror, we risk confusing mirrored outputs for genuine innovation and overlook both their limitations and their potential to reshape how we create, reason, and decide.

Human development, by contrast, is grounded in verbal language that is messy, incomplete, and experimental. Children stumble through broken words and false starts, gradually building the capacity for creativity and novel thought. LLMs, however, are trained on an almost inconceivable scale of polished written language — over 13 trillion words, that are already fully formed and structured. GAI’s fluency mirrors its polished training set. Human creativity comes from imperfection, from the stumbles that spark invention. Humans invent through experimentation; LLMs reflect through limitless polished written text.

To view the human mind purely as an input–output device is to reduce thinking to data processing and probability calculations. This perspective aligns with a Bayesian framework of cognition, where beliefs are continuously updated based on the likelihood of new evidence. While this model captures some aspects of how humans refine judgments, adjusting expectations as information accumulates, it misses the broader dimensions of human thought. Human cognition is not only probabilistic; it is also imaginative, intuitive, and context-dependent.

This tension is captured in what we might call a data–belief asymmetry. Data can be gathered and updated almost instantly, yet beliefs, whether individual or collective — tend to shift far more slowly. Even when evidence is overwhelming, prior assumptions, cultural narratives, or organizational interests often anchor people to outdated views. The result is a lag: strategies, policies, and decisions may continue to reflect yesterday’s beliefs while today’s data points in a different direction. We see this asymmetry in public health, where evidence of smoking’s dangers took decades to change behaviour; in finance, where data on housing risk was ignored ahead of the 2008 crisis; and in climate science, where accumulating data still struggles against entrenched disbelief.

Generative AI embodies this asymmetry in its own way: it consumes vast amounts of data and produces fluent outputs, but those outputs remain tethered to the patterns of the past. Humans, by contrast, often resist or delay belief change even when data demands it — but in that resistance lies the possibility of creativity, reframing, and entirely new ways of seeing.

As Felin and Holweg argue in their Strategy Science article, “Theory is All You Need: AI, Human Cognition, and Causal Reasoning,” Generative AI’s stochastic design means it mirrors the consensus of the past rather than breaking from it. They pose a thought experiment: if Galileo in 1633 had consulted a system like GAI about the heliocentric model, the idea that the planets revolve around the sun, it would almost certainly have reinforced the dominant geocentric view. To such a system, trained on the texts of the time, Galileo’s theory would have seemed improbable, even delusional.

An equally powerful example comes from the history of flight. In 1888, Joseph LeConte’s The Problem with a Flying Machine argued that human flight was impossible, citing the observation that no bird over 50 pounds could fly. The data seemed conclusive, and aerial flight was interpreted through the prevailing theory of weight. If the Wright Brothers had asked ChatGPT of the 1880s whether heavier-than-air flight was possible, it would have echoed the same conclusion: impossible. But the Wright Brothers re framed the problem by focusing on lift and speed, not weight. Their leap was theoretical, not probabilistic, and that re-framing unlocked the future of aviation.

Where GAI reflects probabilities, humans generate possibilities. Where machines mirror the past, humans create futures.

That is the core of data–belief asymmetry: data alone tends to anchor us to prevailing assumptions, while theory and imagination allow humans to break free of the mirror of the past. Breakthroughs have never come from probability alone. They come from belief, imagination, and the courage to invent differently. That is the human advantage worth protecting.

Sources:

Chollet, F. (2019). On the measure of intelligence. arXiv preprint arXiv:1911.01547.

Felin, T., & Holweg, M. (2024). Theory is all you need: AI, human cognition, and causal reasoning. Strategy Science, 9(4), 346-371.