Where Experts Come From

A few days ago, I got a message from my friend and partner in AI work, Dr. Linda Sommerville, and she pointed me to a great resource: a podcast interview with Simon Sinek and Ethan Mollick on AI and the future of work. I am glad she did because that got me down a rabbit hole of research on a very pertinent subject: AI and expertise, entry-level jobs, and what, truly, the impact of AI on work is. Here are some thoughts from that podcast (link and resources shared at the bottom of this blog) and my own contribution to the subject.

The conversation

Sinek opens by confessing that he generally avoids AI guests, because they arrive in one of two flavors, and he describes them with the weariness of a man who has sat through too many panels: “It’s the greatest thing or we’re all going to die.” What follows never goes to either place, which is rare enough to be worth an afternoon, and it contains one exchange I had to stop and listen to twice.

About twelve minutes in, Sinek asks the question every educator I know has quietly assumed the answer to, which is whether young people are simply better at this by virtue of having grown up alongside it. Mollick’s response is unusually flat for a conversation otherwise full of qualifications. “They’re not AI-Native,” he says. “You’re just talking to Claude. They’re conduits to Claude. Like, if you ask for a report, they’ll give you a beautiful report. They have no idea what’s in that report.”

Anyone who reads student work will recognize that with a small sinking feeling, because the document arrives polished and structured and correctly cited, and there is nobody home inside it; one follow-up question is usually enough to dissolve the whole thing. What makes the exchange interesting is not the diagnosis, which is familiar, but the conclusion Mollick draws from it, which inverts twenty years of educational technology orthodoxy. “I think this is a rare case where the more experienced you are, sometimes the older you are, the better you’re going to be at using AI if you decide to use it.”

We have spent two decades assuming that the young adapt while the old resist, and that our institutional task is helping faculty catch up with students. Mollick is describing a technology where the advantage runs the other direction, because the scarce skill is not operating the tool but knowing when the beautiful report is wrong, and that is a form of knowledge you cannot acquire quickly. He puts the mechanism precisely later on, when he says that an expert looking at flawed output can tell “not just, it’s wrong, but often it’s wrong because of a specific problem that you should have either specified better.” Recognizing the specific problem, rather than merely sensing that something is off, is what a career buys you.

The consequence arrives when the conversation turns to how anyone becomes that expert in the first place, and here Mollick produces what is both the funniest and the most alarming line in the hour. He describes the traditional path into a profession, in which juniors did repetitive work while their seniors reviewed it, and then explains why the arrangement has quietly broken down: “Every junior person knows less than ChatGPT. And every middle manager would rather delegate to the AI than a flawed human. And so everyone’s just doing AI work to each other.”

That last sentence is an entire organizational chart collapsing into a joke, and it earns the warning he attaches to it, which is that “the danger is that we lose the talent pipeline. There are solutions to it, but they’re going to require a fairly radical change in how we think about talent pipelines.” He does not say what the radical change is, and the conversation moves on, as conversations do. That unfinished sentence is what I have been chewing on for a week, because I run an institution whose entire purpose is the front end of that pipeline, and “fairly radical change” is the sort of phrase that stays comfortably abstract until you are the person who has to fund it.

Judgment was never a subject

Before we can rebuild a pipeline it seems worth asking what the pipeline was producing, because we discuss judgment as though it were a course we could add or a competency we could list in a job description. It has never behaved that way. Judgment settles on people the way sediment settles, over years of making decisions that could be wrong, having them corrected by someone who knew better, and caring enough about the correction to be changed by it. The junior analyst drafted the memo that came back covered in ink. The resident saw the patient first and then defended that assessment to an attending physician who had seen a thousand similar presentations and could tell within seconds that she was reaching.

Michael Polanyi described the mechanism in 1958, and I have not found a better formulation of it since: “An art which cannot be specified in detail cannot be transmitted by prescription, since no prescription for it exists. It can be passed on only by example from master to apprentice.” [1] What this means is that formation depends on an arrangement of people and work rather than on a body of transmissible content. Dismantle the arrangement and no curriculum will substitute for it.

A handful of professions understood this and built accordingly. Medicine designed residency, aviation designed type ratings and simulator hours, conservatories designed the studio lesson, and each of those fields spent real money to preserve a structure in which novices work under observation on consequential tasks. Those are the fields best insulated today, which suggests the problem is expensive rather than mysterious. The rest of us inherited something far more casual, a junior tier that produced value while it quietly produced people, doing the second thing without anyone writing it down or defending it in a budget meeting.

The difficulty now is not that AI can do the junior work. It is that the junior work was doing two jobs while we were only ever paying for one.

The half of the argument that nobody repeats

The part of Mollick’s pipeline warning that gets quoted is the part about learning, that beginners developed competence by doing repetitive work and will now stop. The part that almost never travels alongside it is that managers used the very same repetitive work to find out who was any good. One low-value activity carried two functions: the memo taught the analyst something about how arguments hold together, and it told the partner whether this particular analyst could think. Remove the memo and you lose the formation, which is the crisis everyone is discussing, and you also lose the observation, which is a different crisis and, for those of us who issue degrees, the more immediate one.

Economists have been mapping this second territory for fifty years, and we in higher education should borrow their vocabulary rather than reinventing it badly. Michael Spence and Kenneth Arrow argued in 1973 that education functions partly as a signal, a way of conveying to an employer something about a person the employer cannot observe directly. [2] Joseph Altonji and Charles Pierret later showed what happens after the hire, demonstrating that employers begin by leaning heavily on the credential and then shift their weight onto observed productivity as they accumulate direct evidence over the first decade of a career. [3] The diploma makes a claim; the early years are where that claim gets tested and gradually replaced by actual knowledge of the person.

Both halves of that arrangement are now degrading, and they are degrading for the same reason, which is what makes this different from the perennial complaint that grades are noisy. Our transcripts lose their discriminating power the moment a submitted assignment can be produced to a high standard either by a student with real judgment or by a student with a good subscription, since on the page those two are indistinguishable. Simultaneously, the entry-level position that used to correct our imperfect signal is being quietly withdrawn, so there is no second look coming. The credential got noisier and the instrument that used to clean it up went away in the same few years, and I have not seen anyone in my sector name that as a single problem rather than two unrelated inconveniences.

The obvious objection is that screens never disappear, they migrate, and the objection is correct, which is precisely why it worries me. Employers who can no longer read talent from junior work will substitute other things: work trials, process telemetry, referrals, and above all the prestige of whatever name is printed on the diploma. Those substitutes track family income and personal network considerably more tightly than a two-year junior post ever did. A first job is a mediocre meritocratic instrument, full of favoritism and luck, but it is a better one than knowing somebody’s uncle, and replacing the apprenticeship screen with the reputation screen is therefore a regressive change dressed as a neutral efficiency. In an economy like ours it will fall hardest on exactly the students my institution exists to serve, and it will arrive without any announcement, because nobody ever publishes a memo saying that the ladder has been removed.

What the evidence actually shows, and where it goes quiet

The strongest data on this is American, and it is suggestive rather than conclusive. The Stanford Digital Economy Lab, working inside ADP payroll records covering millions of workers, finds employment for people aged 22 to 25 in the most AI-exposed occupations running roughly 19 percent below where it would sit had it tracked their less-exposed peers, with the adjustment occurring through reduced hiring rather than layoffs. [4] That mechanism is worth dwelling on, because a hiring freeze produces no photograph and no press release; the door simply fails to open, and the people it fails to open for have no way of knowing that anything happened to them.

There is a serious challenge to this reading. Yuriy Iscenko and Daniel Curto Millet examined 238 million job postings and argue that postings in AI-exposed fields peaked in early 2022, months before ChatGPT was released, which fits the timing of interest rate rises moving through rate-sensitive sectors better than it fits a technological explanation. [5] I cannot dismiss that. What the monetary story handles less comfortably is why the decline concentrates in occupations where AI substitutes for human tasks while employment holds up where it complements them, since a rate shock has no particular reason to sort itself along that line.

For Southeast Asia there is no study of comparable rigor, and that absence troubles me, because the Philippines may have more riding on this question than any country on earth. Our information technology and business process sector employed 1.89 million people in 2025 on revenues of 40.3 billion dollars, equivalent to roughly 8 percent of GDP and around 64 percent of all services exports, with close to 89 percent of that employment concentrated in contact centre work. [6] An IMF working paper mapping AI exposure across Philippine occupations found the encouraging result that most highly exposed occupations here are also highly complementary, meaning the technology is more likely to assist than replace. The residual category, high exposure combined with low complementarity, covers about 14 percent of the workforce, and business process work sits at its centre, with 73 percent of the sector’s workers classified as contact centre information clerks. [7]

I am not going to convert those figures into a prophecy, and I have a specific reason for the reluctance. In 2016 the analyst firm HFS Research projected that automation would reduce low-skilled outsourcing headcount in this exact sector through 2022. [8] Across precisely that window our headcount rose from 1.15 million to 1.57 million, which means the last confident forecast about this industry was wrong in direction rather than merely in magnitude, and anyone writing about it now, myself very much included, ought to keep that failure in view.

What I find harder to argue with is the industry’s own arithmetic. In July 2026 IBPAP revised its roadmap and named agentic AI among the reasons, lowering a 2028 target that had stood at 2.5 million workers to a range running from 2.14 million in the best case down to 1.85 million in the downside case. [9] Set that downside figure beside the 1.89 million already working in the sector during 2025 and the shape of it becomes clear, because the pessimistic scenario is not slower growth but fewer people than we employ today, and the roadmap describes those remaining workers as AI-enabled, which is a diplomatic way of saying not the same jobs.

The June 2026 Labour Force Survey supplies the other end of the picture: youth employment fell from 90.6 percent to 86.5 percent year on year, the youth labour force grew by roughly 600,000 while about 246,000 found work, and National Statistician Claire Dennis Mapa attributed the shift primarily to a sharp rise in participation, particularly among fresh graduates. [10] A single quarter is not a trend and I would be embarrassed to build a case on one release, but when the national statistician points at fresh graduates in an economy whose largest export sector is overwhelmingly entry-level work that the IMF has flagged as our most displaceable category, the configuration is worth saying aloud rather than filing.

What Sinek is worried about, and where I would push Mollick further

The most human passage in the interview has nothing to do with labour economics. Sinek explains that he has no interest in producing a synthetic version of himself, because “I like people knowing that when they see me and they think it’s me, it really is me,” and when the conversation turns to art he says what he values in the piece on his wall is “knowing that a person conceived of it. I like knowing that a person made it.” On learning he is disarmingly simple: “I enjoy learning the same way a painter likes painting and a musician likes playing music and composing.” It is from there, rather than from any economic anxiety, that he names the fear underneath the whole hour: “My concern is that thinking, the ability to think is the sacrifice here.”

Mollick does not offer reassurance, which I appreciated. He tells him instead that shortcuts have measurable consequences, that “if you shortcut that through AI giving you the answer, is you learn nothing. We have enough experiments to show that,” and that the task facing educators is “making people essentially lift mental weights” in a world newly full of ways to avoid lifting anything. He is also hopeful about the tool in its proper role, observing that “personalized education is now an actual possibility” and predicting that “we’ll do more in-class assignments. Outside of class, we’ll use AI tutors. We’ll figure it out.”

This is the one point where I would push him further than he goes. Mollick’s observation that experience now raises the value of AI use is correct and useful, but it contains a corollary he does not follow, which is that the advantage he is describing has an expiry date. The experienced professionals who can spot the specific problem in a flawed output acquired that capacity under working conditions we are in the process of removing, which means the current generation of good AI users is not evidence that the system reproduces itself. It is evidence that it worked until recently. An advantage held by people formed under an arrangement that no longer exists is a depreciating asset, and the interesting question is not how those professionals should use AI today but where their replacements are supposed to come from.

The second place I would press is gentler. “We’ll figure it out” is a reasonable thing for a professor to say about a system, but it is what institutions tend to say when no one in particular owns a problem, and the problem here has an owner whether or not the owner has volunteered. Mollick frames the pipeline collapse largely as an organizational matter, something firms will have to redesign. I think a great deal of it lands on universities instead, and lands on us without our having asked for it or budgeted for it, because if the workplace no longer supplies the first years of formation then the only institution positioned to supply them is the one the graduate has just left.

The irony, restated

In 1983 a British psychologist named Lisanne Bainbridge published a five-page paper on industrial automation that mentions nothing resembling artificial intelligence and describes our situation better than most of what has been written this year. She observed that automating a process does not remove the human being but changes what the human being is for, and that the person left monitoring the machine gradually loses the skill monitoring depends on, a finding she summarized in a sentence that has not been improved on: “a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one.” Her second observation belongs beside Mollick’s, since she noted that when something goes wrong and a person must take over, the situation is by definition unusual, so the operator “needs to be more rather than less skilled, and less rather than more loaded, than average.” [11]

We are moving a generation into verification roles on the assumption that verification is the cheap part of the work. Producing the work was how the capacity to check it got built.

None of this is settled, and I want to resist the fatalism that writing about AI tends to slide toward. Daron Acemoglu and Simon Johnson have argued at length that the broad prosperity of earlier industrial transformations was not delivered by the technologies themselves but won afterward through institutions, bargaining and law, and that the considerably worse outcome was available throughout. [12] Mollick makes a compressed version of the same argument, observing that whether a technology augments people or displaces them is not a property of the technology but of what the surrounding institutions decide to require of it. Universities are institutions that decide what to require, which is close to the whole of what we do.

I am curious to hear your thoughts on this as well.

References

[1] Michael Polanyi, Personal Knowledge: Towards a Post-Critical Philosophy (Chicago: University of Chicago Press, 1958), 53.

[2] Michael Spence, “Job Market Signaling,” Quarterly Journal of Economics 87, no. 3 (1973): 355-374; Kenneth J. Arrow, “Higher Education as a Filter,” Journal of Public Economics 2, no. 3 (1973): 193-216.

[3] Joseph G. Altonji and Charles R. Pierret, “Employer Learning and Statistical Discrimination,” Quarterly Journal of Economics 116, no. 1 (2001): 313-350.

[4] Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab, August 2026 update.

[5] Yuriy Iscenko and Daniel Curto Millet, Looking for the Ladder: Is AI Impacting Entry-Level Jobs?, Economic Innovation Group, January 2026.

[6] IBPAP, 2026 Industry Overview: The Philippine IT-BPM Sector; GDP and services export shares from ASEAN+3 Macroeconomic Research Office, Can the Philippines IT-BPM Industry Stay Ahead Amid the AI Wave?

[7] Micholo Cucio and Tristan Hennig, Artificial Intelligence and the Philippine Labor Market: Mapping Occupational Exposure and Complementarity, IMF Working Paper WP/25/43, February 2025.

[8] HFS Research, cited in Statista, “Change in Low-Skilled IT/BPO Service Worker Numbers Due to Automation and AI from 2016 to 2022, by Select Country,” published August 30, 2017.

[9] “AI, Global Competition Force Philippines’ IT-BPM Industry to Cut Targets,” BusinessWorld, July 15, 2026.

[10] Philippine Statistics Authority, June 2026 Labour Force Survey, released August 6, 2026.

[11] Lisanne Bainbridge, “Ironies of Automation,” Automatica 19, no. 6 (1983): 775-779.

[12] Daron Acemoglu and Simon Johnson, Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity (New York: PublicAffairs, 2023).

Podcast

Simon Sinek and Ethan Mollick, “The AI Skills Nobody is Teaching (And Everyone Needs),” A Bit of Optimism, season 6, episode 17, June 16, 2026. Ethan Mollick is Ralph J. Roberts Distinguished Faculty Scholar and Associate Professor of Management at the Wharton School, University of Pennsylvania, and co-director of Wharton’s Generative AI Labs. See also Ethan Mollick, Co-Intelligence: Living and Working with AI (New York: Portfolio, 2024).