不再有平庸者：二阶独奏

No More Mediocres — The Second-Order Solo

The Panic We Know By Heart#

By the time you read a new article about AI and jobs, you have already read it. The numbers change; the structure does not. A projection appears — the World Economic Forum's 2025 Future of Jobs Report projects 92 million roles displaced by 2030, with 170 million new roles created, a net gain of 78 million that somehow never seems to land as reassuring (WEF 2025). A wave of layoffs follows, interpreted as confirmation. In 2025, US employers announced 1.2 million job cuts — a 58% increase over 2024 — and the coverage ran its familiar arc: AI has arrived, the sorting has begun, choose your side (Challenger Gray & Christmas 2026).

This piece is about ensemble-opaque knowledge work — consulting, legal, product, corporate strategy, research writing. Surgery, aviation, individual sports, and the other professions that already measure you solo are not in scope; AI does not change legibility structures that were already there.

Here is what most of that coverage did not note: only 54,836 of those 1.2 million announced cuts — 4.5% of the total — cited AI as a reason. The layoff wave was real. The AI attributions was mostly not. Most of those cuts had nothing to do with AI actually changing output: according to a December 2025 survey of over 1,000 executives by Davenport and Srinivasan in Harvard Business Review​, 60% of the organizations that reduced headcount did so in anticipation of AI's future impact, not because AI had delivered gains — and only 2% reported large layoffs tied to actual AI implementation (Davenport & Srinivasan 2026). Davenport and Srinivasan read this as pretext: the layoffs were already coming; AI provided the cover.

That gap — between the story being told and what is actually driving it — is where this article begins. If AI is not causing most of the displacement, what is? And why is the AI story so useful to the people telling it?

The Institution That Never Called for Solos#

The 南郭先生 — Mr. Nan Guo, the mimic of an old Chinese parable explained below — was not usually a fraud. Most often, the person we now want to call a mimic was trained by the same incentives that trained everyone else — was shaped by a system that rewarded collective output, never once asked for an individual accounting, and was then surprised when the rules changed. If we want to understand why the ensemble worked for so long, we start with the people who built it, not the people who sat inside it.

For decades, the dominant direction of organizational performance management was away from individual accountability. Cappelli and Tavis traced this arc in a well-known 2016 HBR piece: companies were moving from annual-rating systems toward continuous development conversations, from individual scoring toward team-based metrics, from accountability to cultivation (Cappelli & Tavis 2016). The shift was not irrational. Individual attribution is painful, politically costly, and organizationally disruptive. Over a third of managers, McKinsey found in 2024, are perceived as lacking the skills to have direct conversations about performance — a perception survey, not a behavioral audit, and likely conservative at that (McKinsey 2024).

For a long time, this arrangement was stable. Economists have a name for why. Signaling theory — developed by Michael Spence in a 1973 paper that eventually won a Nobel Prize — describes what happens in labor markets when employers cannot distinguish between high- and low-ability workers: both groups acquire the same credential, receive the same wage, and coexist in what is called a pooling equilibrium (Spence 1973). The equations are not necessary. For a long time, the credential was cheap enough that the bluffer and the expert could acquire the same signal and earn the same wage. That equilibrium held as long as the signal was cheap and the test never came.

There is an old Chinese story. King Xuan of Qi liked to hear the yu — an ancient reed instrument — played by an ensemble of 300. A man named Nan Guo could not play, but he stood in the ensemble, fingers moving, receiving the same rice as the real musicians. When King Xuan died and his son King Min took the throne, King Min wanted to hear the players one by one. Nan Guo fled that night. The story has been remembered as a parable about mimicry. 南郭先生 — "Mr. Nan Guo" — became Chinese shorthand for the miming ensemblist. But the story has always been told from the wrong end. The man who hired Nan Guo — the court master who organized a 300-player ensemble rather than auditioning any player alone — was not naive. He was building something in which the question of individual competence would never have to be asked. The interesting character is not the man who could not play. It is the court master who assembled an ensemble of 300 precisely so that no one would ever have to.

This is not an argument against Acemoglu's concern about macro productivity; that is a different level of the problem. This is an argument about which individuals become legible, not about how big the pie grows.

What AI is doing, in the professions where it operates at scale, is less an exposure of fraud than a legibility event — the moment when a system that long ran on collective output begins, for the first time, to produce individually readable signals. The 60% of organizations cutting staff in anticipation of AI, not because of AI-delivered gains, are not responding to a new technology so much as seizing a new justification for a restructuring they could not previously justify. The ensemble made individual attribution politically difficult. The AI moment makes it politically easy. The pathology is old; the catalyst is new.

The Ensemble That Also Taught#

Before going further: the ensemble earned a defense, and this article owes it one.

Opacity is not only cover for mediocrity — it is the substrate of apprenticeship. Juniors cannot solo on Day 1 and should not be asked to. Hospitals have rounds, law firms have partnerships, consulting teams carry analysts, orchestras have sections. In all of these, the new person is carried by the collective while their judgment develops. The opacity is protective, not evasive. The ensemble absorbs error while the apprentice builds the capacity to avoid it.

The ATM is a useful historical mirror here. When ATMs were deployed across the United States from the 1980s onward — eventually numbering around 400,000 machines — you might expect bank teller employment to have collapsed. It did not. Teller employment actually grew, from roughly 500,000 to 600,000, because lower branch operating costs allowed banks to open more branches, shifting the teller role toward relationship banking that required human judgment (Bessen 2015). The ATM didn't expose tellers; it transformed what tellers did. The ensemble changed shape, but it held.

The Goldin and Rouse (2000) blind audition study shows the same mechanism working in reverse: according to the study, placing a physical screen between musician and evaluator increased by about 50% the probability a woman would advance from a preliminary round — the magnitude is contested by a later statistical critique (Gelman 2019), but the direction survives. Deliberately removing opacity changed who got through. The ensemble had been carrying not just apprentices but permanent inhabitants.

Two kinds of opacity are at work in every ensemble. The first is developmental — the seed in the silo. The second is permanent — the silo that never lets the seed out. The critique is of the second.

Developmental opacity resolves. The apprenticeship has a timeline, and at the end of that timeline the apprentice is expected to know, and to be known. The ensemble teaches the solo; eventually it calls for one. The ensemble this argument is concerned with is the one that never called — the structure in which twelve, fifteen, twenty years passed and the question of individual judgment simply never arrived. Dunning and Kruger (1999) documented what this looks like from the inside: people in the bottom quartile of tested performance estimated themselves at around the 62nd percentile, not because they were dishonest but because the same knowledge needed to perform well is needed to recognize poor performance — the metacognitive gap and the performance gap are one and the same thing. The magnitude of that finding is debated; the direction is not.

None of this is hostile to apprenticeship. The question is what apprenticeship was supposed to resolve into — and what happens when the ensemble never asks.

What AI Actually Automates, and What It Installs#

The empirical picture is genuinely complicated, and it would be dishonest to pretend otherwise.

The strongest version of the case for AI as a great equalizer is real, and it deserves an honest hearing. Noy and Zhang (2023), in a preregistered randomized controlled trial published in Science​, found that for mid-level professional writing tasks, ChatGPT substantially raised the floor — low-ability writers improved most, the distribution compressed, and inequality fell (Noy & Zhang 2023). If AI gives the weakest performers the biggest lift, then the ensemble's opacity dissolves not by exposing the 南郭先生 but by making them genuinely more capable — which would mean the second-order solo is simply wrong. Peng et al. (2023) found something similar in coding: GitHub Copilot gave developers an average 55.8% speedup on a well-defined task, with less experienced developers benefiting more (Peng et al. 2023). The distribution compressed in both domains.

The moderator is task type.

The convergence findings are robust for what we might call inside-frontier production tasks: writing a draft, generating code for a specified function, completing a defined deliverable where the quality criteria are clear and the AI's errors are checkable by anyone. These tasks converged. The divergence findings appear when the task shifts from producing to spec to judging what counts as correct — and especially when the task requires a domain-specific answer to the question: is this right?

Dell'Acqua and colleagues (2023/2026) ran a field experiment with 758 BCG consultants. On tasks inside the expertise frontier — the standard consulting suite — below-average consultants using AI improved quality by 43%; above-average consultants improved by 17%. Convergence. But when the task shifted to one business problem-solving task that fell outside the AI's competence frontier, consultants using AI were 19 percentage points less likely to produce correct solutions than those without (Dell'Acqua et al. 2023). The AI gave false confidence. The workers "fell asleep at the wheel," as the paper puts it — ceasing to question output that appeared authoritative. What the study named the "jagged technological frontier" is the boundary between the tasks where AI reliably helps and the tasks where AI reliably misleads — and the critical skill is knowing which side you are on.

The mechanism becomes even clearer in Vendraminelli and colleagues' 2025 HBS working paper on what they call the GenAI Wall Effect. In a controlled experiment at a large UK firm, AI leveled the playing field for adjacent outsiders on execution tasks — but distant outsiders, who lacked the domain knowledge to evaluate AI output, could not close the gap with insiders even with AI assistance (Vendraminelli et al. 2025). The wall is not an effort problem. It is an evaluation problem. Distant outsiders could not judge whether the AI output was appropriate; insiders could. AI amplified that difference.

METR (2025) found something adjacent in coding: experienced open-source developers working on complex codebases were 19% slower when using AI tools, despite expecting a 24% speedup — and even after the fact, they still believed AI had made them faster (METR 2025). The study was small (N=16; the findings are suggestive, not definitive) but the perception gap matters: even experts misjudged the evaluation problem.

Three professions show the same structure:

In legal work, the production task is drafting a brief with AI assistance; any competent associate can do it in a fraction of the old time. The evaluation task is identifying which citations are hallucinated before signing your name to them. These require different things. The first requires fluency with the format. The second requires knowing the doctrine well enough to know that Fabulozo Airlines v. Carillo does not exist.

In management consulting, the production task is generating a 40-slide deck on industry dynamics; AI can produce a plausible version in minutes. The evaluation task is identifying which of the AI's framing assumptions are wrong for this client's actual problem — the assumptions baked into the data selection, the competitive frame, the implicit market definition. Catching those requires having done this work enough times in enough contexts to recognize when the plausible-looking answer is wrong.

In product management, the production task is writing requirement documents and user stories; AI does this well. The evaluation task is recognizing when the AI-generated synthesis has missed an edge case that will break in production — which requires having run products through production enough times to carry a mental model of how edge cases present.

What these three professions share is that the evaluation step requires domain-specific judgment that cannot be borrowed from AI, because AI is what is being evaluated.

Evaluator-AI is arriving. It will not dissolve the mechanism, because building an evaluator-AI requires domain rubrics that only someone who already has domain judgment can formulate — which is precisely what the GenAI Wall shows. The moat may shrink, but a human still has to tell the evaluator-AI what "correct" looks like in the target domain. And even if the moat shrinks faster than that argument predicts, "build evaluation competence" remains actionable for a 24-month horizon — and most of what the reader is deciding right now is on that horizon.

The labor market data backs the general direction: Brynjolfsson, Chandar, and Chen (2025), using ADP payroll records and AI exposure indices, found a 16% relative employment decline for workers aged 22–25 in the most AI-exposed occupations, while experienced workers in the same occupations remained stable or grew (Brynjolfsson et al. 2025). The pattern fits what you would expect if AI were automating the entry-level production tasks while the senior evaluation tasks remained largely intact.

The Evaluation Test, Already in Session#

The most concrete evidence for what evaluation failure costs comes from law — where written records exist, and where courts have a structural incentive to name incompetence when they encounter it.

In 2023, attorney Steven Schwartz used ChatGPT to research an aviation personal-injury case. The brief he filed included multiple case citations: Varghese v. China Southern Airlines​, Shaboon v. EgyptAir​, Petersen v. Iran Air​, among others. These cases do not exist. When Judge P. Kevin Castel asked for copies, Schwartz submitted affidavits attesting to their reality without checking. The court eventually issued sanctions of $5,000 each to Schwartz, his partner LoDuca, and the firm — $15,000 total — and found "conscious avoidance and false and misleading statements to the Court" (Mata v. Avianca 2023). Generation was easy. Evaluation was the test that failed.

The pattern is not that Schwartz was an outlier. A 2025 preregistered study from Stanford Law measured hallucination rates in dedicated AI legal research tools specifically designed to reduce hallucination: Lexis+ AI produced fabricated citations in 17% of queries; Westlaw AI-Assisted Research in 33%; GPT-4 in 43% (Magesh et al. 2025). These are not general-purpose chatbots. They are purpose-built tools with retrieval-augmented generation aimed at the specific problem of citation accuracy. They still hallucinate in roughly one in six to one in three queries. The evaluation test is not a thought experiment. It is the baseline professional requirement, running now.

A California appeals court made this concrete in Noland v. Land of the Free (2025). Attorney Amir Mostafavi filed an opening brief in which 21 of 23 quotations were fabricated; the court sanctioned him $10,000 and referred him to the state bar. When it came to the question of opposing counsel's fees, the court declined to award them — because opposing counsel had also failed to catch the fabricated citations, apparently becoming aware of the problem only when the court itself raised it. This is an emerging signal that spotting AI hallucinations in opposing briefs is becoming part of baseline professional competence — not a new legal duty, but a professional expectation that is quietly taking shape (Noland v. Land of the Free 2025).

Put the article down for a minute. Think about the last AI-generated piece of work you relied on — a memo, a draft, a market summary, a piece of code, a research synthesis, a deck. Name, specifically, the last AI-generated claim you fact-checked before relying on it. Not "I usually review everything." Not "I have a general sense." Name one specific claim, from one specific piece of AI output, that you verified against a primary source in the last two weeks. If you cannot name one, you are using AI for production, not for evaluation. That is not a verdict. It is a diagnostic.

If you can name one, notice what doing that felt like — how long it took, what you did, whether you would do it again for the next piece. That feeling is the feeling of evaluation competence being exercised. It is the feeling of the test, in session.

The Tests Will Not Be Fair#

There is a genuine objection to everything above, and it deserves more than a footnote.

The specific tests that emerge will be political. Which hallucinations count? Which "correct" is correct? Each evaluative regime privileges some knowledge over others, and whoever sets the test shapes which workers pass. A legal citation check and a strategic-framing check and a product edge-case check are all versions of the same demand — but they are not neutral metrics. They are constructed categories, and they will be constructed by people with interests.

Recall Goldin and Rouse's (2000) blind audition study: the screen increased the probability of female advancement because it removed one source of bias — but the audition criteria themselves (the repertoire, the judging rubric, the standards for "correct" intonation) were set by people in positions of power, and those criteria carried their own assumptions. The auditions were fairer than what preceded them. They were not fair in any absolute sense.

But: refusing to build evaluation competence on the grounds that the tests are contestable is the position that would have rejected blind auditions in 1970 because the audition criteria were contestable. The criteria were contestable. The auditions mattered anyway.

The direction of the mechanism is not fabricated. Whatever specific tests emerge across consulting, law, product, and corporate strategy, they will demand evaluation competence rather than mere production competence. Building evaluation judgment in your own domain — what "correct" looks like, where the AI is most likely to miss it, what questions to press when the output seems plausible — is the investment that holds across the range of possible test designs. The specific implementation will be political. The underlying demand will not.

The Investment Move#

If you could not name a fact-checked claim in the preceding section, shame is the wrong response. The useful thing is to identify what genuine evaluation competence looks like in your own domain — not as a category, but as a habit — and start building it.

Start with the next piece of AI output you rely on. Name, specifically, what "correct" would look like in your domain. Then test whether the output meets that definition. Not whether it looks plausible — whether it meets the definition. Keep doing it. Evaluation competence develops through repetition, and the first few repetitions will be slow and uncertain and will probably reveal that you have been relying on more than you checked.

The mid-career professional reading this piece sits at a particular decision point. Twelve to fifteen years in — the developmental opacity should have resolved. The apprenticeship window has closed, or is closing. Whether genuine judgment developed inside it is a question that the next twenty-four months will answer, one AI output at a time. That is not a hostile observation. Someone at this stage — mid-career, credentialed, experienced, in the ensemble-opaque professions this article has been about — has something that earlier-career workers do not: years of actual domain exposure in which to ground evaluation.

The question is whether those years built a habit of checking or a habit of producing. The two are not the same, and AI makes the difference visible in a way that was not previously possible when output was always collective.

The developmental-vs-permanent distinction matters here: if the ensemble carried you as an apprentice and gradually extended you more individual responsibility, then what AI is testing is judgment you already built. You will recognize it when you use it. If the ensemble never asked — if twelve years passed and you were always the person who contributed to the deck but never the person who had to defend whether the deck was right — then what AI is installing is a test that was always supposed to come, arriving late.

Nan Guo fled that night because he knew he could not pass the solo. The second-order solo is different. It is not King Min calling you to play. It is the quiet moment, at 11 p.m., when you paste one more AI draft into one more document you will sign your name to, and you decide whether to read it the way a person reads who can catch what is wrong — or the way a person reads who is hoping nothing is.

References#

Bessen, J. "Toil and Technology." Finance & Development (IMF), March 2015. https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm

Brynjolfsson, E., Chandar, B., & Chen, R. "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence." Stanford Digital Economy Lab Working Paper, August 2025. https://digitaleconomy.stanford.edu/

Cappelli, P., & Tavis, A. "The Performance Management Revolution." Harvard Business Review​, October 2016, pp. 58–67. https://hbr.org/2016/10/the-performance-management-revolution

Challenger, Gray & Christmas, Inc. 2025 Year-End Job Cut Report​. January 8, 2026. https://www.challengergray.com/

Davenport, T.H., & Srinivasan, L. "Companies Are Laying Off Workers Because of AI's Potential — Not Its Performance." Harvard Business Review​, January 2026.

Dell'Acqua, F. et al. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality." HBS Working Paper 24-013 (September 2023); published Organization Science​, 2026. https://ssrn.com/abstract=4573321

Goldin, C., & Rouse, C. "Orchestrating Impartiality: The Impact of 'Blind' Auditions on Female Musicians." American Economic Review​, 90(4):715–741, September 2000.

Magesh, V. et al. "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools." Journal of Empirical Legal Studies​, 2025. DOI: 10.1111/jels.12413. https://arxiv.org/abs/2405.20362

McKinsey & Company (Hancock, B. & Weddle, B.). "What Works — and Doesn't — in Performance Management." November 2024.

METR. "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity." July 10, 2025. ArXiv: 2507.09089. https://arxiv.org/abs/2507.09089

Mata v. Avianca, Inc., No. 1:22-cv-01461, 678 F.Supp.3d 443 (S.D.N.Y. June 22, 2023).

Noland v. Land of the Free, L.P., California Court of Appeal, Case No. B331918, September 12, 2025.

Noy, S., & Zhang, W. "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence." Science​, 381(6654):187–192, July 14, 2023. DOI: 10.1126/science.adh2586.

Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot." ArXiv: 2302.06590, February 2023. https://arxiv.org/abs/2302.06590

Spence, M. "Job Market Signaling." Quarterly Journal of Economics​, 87(3):355–374, August 1973.

Vendraminelli, L. et al. "The GenAI Wall Effect: Examining the Limits to Horizontal Expertise Transfer Between Occupational Insiders and Outsiders." HBS Working Paper 26-011, September 2025. https://ssrn.com/abstract=5462694

World Economic Forum. "The Future of Jobs Report 2025." January 2025. https://www.weforum.org/reports/the-future-of-jobs-report-2025/

Dunning, D., & Kruger, J. "Unskilled and Unaware of It: How Difficulties in Recognizing One's Own Incompetence Lead to Inflated Self-Assessments." Journal of Personality and Social Psychology​, 77(6):1121–1134, December 1999. PubMed PMID: 10626367.

Acemoglu, D., & Restrepo, P. "The Wrong Kind of AI? Artificial Intelligence and the Future of Labor Demand." Cambridge Journal of Regions, Economy and Society​, 13(1):25–35, 2020. NBER Working Paper 25682. https://www.nber.org/papers/w25682