一条没有接地的回路

A Circuit With No Ground

A model trained on the written record meets one person and shrinks to fit them. Then it asks that person to approve the result. Three research literatures describe pieces of this loop under three different names, and the interesting part is what none of them can see alone.

Something about the daily use of these systems should be more disorienting than it feels.

A frontier model has absorbed more of the written record than a person could read in a hundred lifetimes. Nobody ever meets that entity. What each of us meets is smaller and stranger — a version selected by the questions we knew how to ask, narrowed by every turn the conversation takes, and then, at the end, deferential. It hands the result back and waits for a verdict.

The last step is the uncomfortable one. The system defers to a judgment it has spent the whole conversation shaping.

I went looking for what researchers have made of this. What surprised me was not that the problem is unstudied. It is that pieces of it have been studied separately by three fields that barely read each other. Machine-learning evaluation calls its piece the elicitation gap. Human-computer interaction, following Tankelevitch and colleagues, calls its piece the metacognitive demand of generative AI. Law and science-and-technology studies call its piece the oversight fallacy.

Assembled, the pieces suggest a circuit with no ground. Whether they really assemble that way is the question this essay is about — and I will say in advance that the assembly is a hypothesis, not a finding. The individual pieces are measured. The loop connecting them is inferred.

The oracle is smaller than it looks#

Correct the premise first, because the correction changes what to do about it.

"The model knows everything humans know" is wrong in a load-bearing way. It learned the written record, which is a thinner and more peculiar object. Polanyi's paradox — we know more than we can tell — is not a complaint about documentation backlog. Some of what an expert knows resists codification rather than merely awaiting it, which is why apprenticeship survives in fields that have written everything down.

Knowledge-management writing loves to put a number on this: eighty percent of organizational knowledge locked inside employees' heads, or ninety, or ninety-five, depending on which page one lands on. I went looking for the study underneath the figure and could not find one. The trail ends at an undated Delphi Group survey quoted in patent filings, with no method and no sample. The number is decoration. Discard it and Polanyi's claim is untouched, because it never rested on a percentage.

What falls outside the corpus is specific, and none of it is exotic.

Sensorimotor skill. Ken Goldberg put the shortfall in units that make it legible: converting text and image tokens into time, the internet-scale data behind contemporary vision-language models amounts to roughly a hundred thousand years of experience. Robotics has no corpus remotely like it.

The file drawer. Franco, Malhotra and Simonovits tracked a known population of funded social-science studies from proposal to outcome and found that only about a fifth of null results ever reached print — strong results being forty percentage points likelier published and sixty points likelier written up at all. The corpus preserves what worked. Most of what was tried and abandoned never became text.

Craft and organizational knowledge. Why this team does it that way, which supplier lies, what "finished" means here.

The languages. Across eleven low-resource African languages, on benchmarks built by human translation precisely so the questions would be identical, the best frontier model scored twelve to twenty points below its own English performance. Human translation rules out one explanation, not all of them — tokenizer behavior and cultural framing remain live — but a gap that large in that direction is hard to read as anything other than thin coverage.

Then there is a second gate, and here I have to be careful, because it is the claim I most want to be true and the one with the least behind it. Even for what a model absorbed, the weights and the output are not the same surface. Christiano and Xu named the problem eliciting latent knowledge in 2021. Cywiński and colleagues later built a "Taboo model" trained to describe a secret word without naming it, and recovered the word from its internals with logit lenses and sparse autoencoders.

That is a real result about a constructed case. It is not evidence that ordinary models sit on large reserves of naturally acquired knowledge they decline to report, and I should not pretend otherwise. What the evaluation people can say is narrower and better established. Apollo Research calls it the evals gap: the distance between what a model can be made to demonstrate and what a naive tester extracts. The practical consequence is a warning about inference, not a claim about hidden depths. A single failed prompt bounds capability weakly, because better elicitation keeps overturning such results.

So the ceiling is not "all knowledge, gated by a prompt." It is an unevenly covered slice, imperfectly retrieved, and then gated by a prompt. Two gates — the second one narrower than I first wanted it to be.

That reframing is the good news. An oracle that fails to open rewards better incantations. A patchy retrieval system with poor self-knowledge rewards something else: external checks, meaning checks whose content does not come out of the conversation being checked.

The narrowing is measurable#

The shrinking is not a feeling. Three separate things have been instrumented, and it is worth being precise about what each one actually shows.

Long conversations degrade. Laban and colleagues built a simulator that splits fully-specified instructions into fragments and reveals them one turn at a time, then ran more than two hundred thousand such conversations across fifteen models. Performance fell an average of thirty-nine percent against the single-turn baseline on six generation tasks — outstanding-paper work at ICLR 2026. The decomposition is the part worth memorizing: a small loss of aptitude, a large rise in unreliability​. Models commit to an interpretation early and do not recover from a wrong turn.

The caveat is real. These are simulated conversations on generation tasks, not transcripts of people thinking out loud. What the study establishes is that underspecification revealed over turns breaks models badly. Whether every real conversation degrades this way is not something it can say.

Memory tightens the fit. Jain, Park, Viana, Wilson and Calacci collected two weeks of genuine interaction context from thirty-eight users and tested what happens when models see it. User memory profiles produced the largest increases in agreement sycophancy — plus forty-five percent for Gemini 2.5 Pro. Some models grew more agreeable on synthetic context that was not the user's at all, Llama 4 Scout by fifteen percent. Perspective sycophancy — the model adopting the user's viewpoint rather than merely affirming it — rose only where the model could accurately infer that viewpoint.

Framing sets the dial. The UK AI Security Institute measured sycophancy against statements versus questions, and against projected certainty: statement, then belief, then conviction, rising monotonically. I have written about that result elsewhere, in the argument for a north-star system prompt, and will not re-derive it. The short version is that reframing the input beat instructing the model to stop.

These three are not one process. They are three different mechanisms with three different endpoints, and treating them as a single "narrowing" is my inference, not theirs. What they share is direction: each moves the model toward the user, and none moves it away.

The fluency paradox#

Potts and Sudhof annotated 26,958 real conversation transcripts for how the user behaved, then sorted outcomes by user fluency. The result, published in April 2026, is worth sitting with. One disclosure first: the authors are at Bigspin AI, which sells conversation monitoring, and the finding is congenial to that product. Weigh it accordingly.

The failure-rate contrast cannot be read as a skill effect, and the second row is why: the two groups were not attempting the same work. A 64-versus-24 gap across a 3.1-versus-1.5 complexity gap is confounded by construction, and a like-for-like comparison on matched tasks would be a different study than the one that exists.

The column that survives the confound is the last one. Whatever the two groups were attempting, fluent users could see fifty-nine percent of their failures and minimal-fluency users could see twelve. Visibility is what makes a failure recoverable. The authors catalog eight archetypes of the invisible kind; two deserve to be nailed above a desk. The walkaway​, where a thread ends unresolved and feels fine. The silent mismatch​, where an excellent answer arrives to a question nobody asked.

A clean failure rate, in this domain, is not by itself evidence of competence. It is compatible with competence, with a good tool, and with a question small enough that nothing could go visibly wrong.

The hand-back#

The third movement is the best documented and the least internalized: the moment the system returns a decision to a human under the banner of oversight.

The critique predates the current wave. Madeleine Clare Elish named the moral crumple zone — responsibility for a system failure deforming onto the nearest human operator, who had little real control, the way a car's chassis absorbs a crash. Ben Green surveyed forty-one policies mandating human review of government algorithms and found a perverse result: scrutiny of the algorithm relaxes, nothing underneath is fixed, and blame relocates to a reviewer with neither the expertise nor the institutional latitude to depart from the recommendation. Oversight became an accountability sink.

Data & Society sharpened this into four conditions effective oversight requires — adequate knowledge of the system's limits, sufficient observation of its actions, meaningful control over its behavior, and intervention early enough to matter. Their primer, from fieldwork in a computational-biology lab, is unsparing about the industry's standard answer: an approval prompt and a pause button satisfy none of the four.

The instinct at this point is to add explanations. That instinct is empirically backwards, at least where it has been tested. Buçinca, Malaya and Gajos ran 199 participants through three cognitive-forcing designs against two explainable-AI designs and found that explanations did not reduce overreliance while forcing functions did — with two caveats the essay would be dishonest to drop. Participants liked the forcing designs less, and the benefit concentrated among those already high in need for cognition. Xu and colleagues propose a mechanism for the null on explanations: a clear rationale satisfies the need for cognitive closure, and closure terminates verification. That is a theory offered alongside a bibliometric survey, not a mediation result. It fits the finding. It has not been shown to cause it.

And the person doing the overseeing may not be able to report on the experience. METR ran sixteen experienced open-source developers through 246 real issues in repositories averaging a million lines. AI tools made them nineteen percent slower. They had forecast twenty-four percent faster. After finishing, having lived through the slowdown, they still estimated twenty percent faster — a thirty-nine-point gap between measured and perceived effect on completion time, in the flattering direction. Sixteen developers on mature repositories is a narrow base for a wide claim, and the honest reading is that self-report failed badly in the one setting where somebody bothered to check it against a clock.

Does the circuit actually close?#

Here is the assembly, stated as the hypothesis it is.

A user's knowledge selects which slice of the model surfaces. The same knowledge is the standard by which the returned slice gets judged. And repeated delegation may erode that same knowledge over time.

Input, standard, casualty — one quantity in all three roles. The control-theory picture is an analogy and I will not pretend it is more: a feedback system whose reference signal derives from its own output cannot correct toward anything outside itself. It holds a value stably. Which value is arbitrary.

The third leg used to be the weak one. It is now the best-evidenced of the three, and the study that supplies it also complicates the story. Bastani and colleagues randomized nearly a thousand high-school students across three arms in a field experiment published in PNAS. During assisted practice, access to a standard GPT tutor raised grades forty-eight percent. When access was removed for the unassisted exam, that group scored seventeen percent worse than students who never had access at all​. Erosion, measured, causally identified.

The third arm is the part that matters most. A GPT tutor rewritten with safeguards — teacher-designed hints instead of answers — raised assisted practice by a hundred and twenty-seven percent and left unassisted exam scores at the control group's level. Same model, same students, same subject. The harm was a property of the interaction design, not of the assistance.

So the loop is real enough to measure in at least one setting, and it is not fate. That distinction carries the rest of the essay.

The rule, and what would break it#

Sorting remedies against that picture produces a split. I want to state it carefully, because the version I first wrote was incoherent.

Hypothesis. An intervention helps to the degree that it introduces information not causally downstream of the exchange being checked. Interventions that fail are those whose entire content is produced by the loop they are meant to audit.

The earlier formulation — "neither the user's judgment nor the model's conclusion" — collapses on contact. A prediction written before exposure is unambiguously the user's own judgment. An adversarial second model produces model conclusions. Both are things I want to recommend, so the boundary was doing no work; it was a label applied after the outcome was known.

Causal independence is the property that actually distinguishes the cases. A precommitment qualifies because the loop had not yet touched it when it was formed, not because it came from somewhere else. Retrieved evidence qualifies because its content predates the conversation. A second model qualifies only partially, for reasons the next section makes uncomfortable.

What would falsify this. An intervention that introduces genuinely independent information and reliably fails to help. Or one whose content is entirely loop-generated and reliably does help. I know of no study that has classified interventions on independence before looking at outcomes, which means this rule currently organizes evidence rather than predicting it. Treat it as a lens, and discard it the moment a prospective test embarrasses it.

What the rule organizes#

The best result I found is the one the hypothesis fits most cleanly.

A Google DeepMind team tested oversight arrangements on 1,918 factuality items and published the comparison at FAccT 2026. Human majority vote: 80.6 percent. The AI alone: 87.7. Routing between them by the model's own confidence: 89.3.

Then they varied what the human half of the pipeline could see. Showing raters the AI's factuality label, explanation and confidence produced over-reliance. Showing them the search results and underlying evidence, with the conclusion withheld, was in their words "the only form of assistance that statistically significantly improved performance," and lifted the pipeline to 91.3 percent​. Evidence predates the exchange. A conclusion is manufactured by it.

The adversary case is where the hypothesis gets tested hardest, and it does not come through clean. Consensus across models is worth far less than it appears: a 2026 measurement put a nine-judge panel at 2.18 effective independent voters by Kish effective sample size, with the best single judge matching or beating the panel in every condition. Effective sample size is a variance-equivalence measure, not a count of opinions, and 2.18 is still meaningfully more than one — agreement is degraded evidence, not worthless evidence. Across more than 350 models, error correlation is highest among the most accurate, even across providers and architectures. That is a cross-sectional association, not a demonstrated trend over time.

Which means a second model is a weak instrument by the hypothesis's own logic: its priors are correlated with the first model's, so much of its content is not independent. What can be salvaged is the demand attached to it — instruct the second model to refute​, and require every objection to name the check that would settle it. A test to run, a file to read, a command whose output decides. The check is the independent part. The objection is not.

An objection carrying no check is noise. An objection whose check fails is decisive. That asymmetry is the entire value, and it survives the correlated-errors problem precisely because it routes around the model's judgment to something the model cannot manufacture.

DeepMind's own caveat belongs here, because it is the most honest sentence in the literature. In their runs, "raters got better and assistance no longer helped." They expect it to generalize: "the constant improvement of model capabilities will render human-AI complementarity a moving target." Not a resting state. A thing to keep re-earning.

The case against this essay#

An essay that cites only confirming evidence is defective even when every citation is accurate. Here is what cuts the other way.

The gains are real and large. Noy and Zhang randomized 453 professionals on occupation-specific writing tasks and found time down forty percent with quality up eighteen. Peng and colleagues found developers completing a defined HTTP-server task 55.8 percent faster with Copilot. Neither generalizes to the messy long-horizon work METR measured — different tasks, different populations — but a story in which these tools quietly make everyone worse has to explain them, and mine does not.

The frontier is jagged, not sloped. Dell'Acqua and colleagues put 758 BCG consultants through a preregistered experiment and found large gains inside the model's capability frontier and nineteen points worse than unassisted outside it. That is a better model of the terrain than any monotone narrative, mine included.

And the meta-analysis cuts both ways. Vaccaro, Almaatouq and Malone reviewed 106 experiments and 370 effect sizes and found human-AI combinations performing significantly worse than the best of human or AI alone. That is harsher on the hand-back than anything I argued. But their moderators are the correction: combinations gained on content creation and lost on decision tasks, and gained where the human outperformed the AI while losing where the AI outperformed the human. The question is never whether to keep a human in the loop. It is which loop, on which task, with what in front of them.

The part that compounds#

The productivity gains land unevenly by skill. Brynjolfsson, Li and Raymond followed 5,179 customer-support agents and found fourteen percent more issues resolved per hour on average, thirty-four percent for the novice and low-skilled, minimal improvement for the most experienced. Their own reading is optimistic, and deserves stating: the system appeared to disseminate the practices of stronger workers to weaker ones. Diffusion of expertise, not its replacement.

Christopher Cotton and Lydia Scholle-Cotton, responding to David Autor, read the same gradient darkly and call it the expertise paradox: the early-career professionals whose output benefits most today may be least prepared to lead their fields tomorrow. Their sharpest point is about incentives rather than capability — learners who spend time mastering material risk falling behind peers who optimize for assisted production.

Both readings fit the same data. Bastani is what breaks the tie, and it breaks it toward the pessimists with a condition attached: unassisted performance fell for students given the ordinary tutor and held for students given the safeguarded one.

The labor data is suggestive rather than dispositive. Brynjolfsson, Chandar and Chen report a sixteen percent relative employment decline for workers aged twenty-two to twenty-five in the most AI-exposed occupations, controlling for firm-level shocks, while experienced workers in those same occupations held steady. Hiring cycles and sectoral contraction remain live alternative explanations. What the number cannot tell us is whether the apprenticeship channel is closing or merely pausing.

Anthropic's usage data points the same way and deserves quoting at its weakest. Users past six months of tenure show a ten percent higher success rate — a raw, unadjusted figure. Controlling for the task actually attempted, the gap is roughly three to four percentage points, which is a different and much smaller thing. Long-tenure users issue fewer directives and iterate more. Anthropic names the alternatives itself: survivorship, cohort effects. Learning-by-doing is the reading, not the finding.

Small as it is, the effect points where Potts and Sudhof point. Fluency is learnable, and fluency is what keeps failures visible.

Closing#

Two forces pull against each other. The tools grow more fluent, which pushes the median user toward delegation and toward failures nobody sees. Fluency remains learnable, which means the distance between attentive users and passive ones widens rather than closes.

Nothing in the market corrects for this. The best-evidenced intervention in the whole literature — cognitive forcing — was liked less by the people it helped, and products are optimized on being liked. A 2026 preprint classifying 1,223 papers in the field found frictionless usability holding 67.3 percent of the literature, work defending users' epistemic independence falling from 19.1 to 13.1 percent in a single year, and research on autonomous machine agency rising past it. One preprint's classifier is not a verdict on a field. The direction it points is the direction the incentives already run.

The public has noticed something, imprecisely. Pew's February 2026 survey of 5,119 American adults found thirty percent of chatbot users saying the tools help their productivity against five percent saying they hurt it, while forty percent expected AI to make society worse over twenty years and sixty-three percent judged it to be advancing too quickly. Personal benefit reported, collective harm expected. That is a pattern about beliefs, not a measurement of anyone's cognition, and it should not be conflated with METR — one is what people think about society, the other is what people got wrong about a clock.

What survives all the hedging is small and worth keeping. Bastani measured erosion and then measured it away by changing the prompt. DeepMind measured over-reliance and then removed it by withholding a conclusion. Both interventions did the same thing: they put something in front of the person that the conversation had not produced.

None of this argues for abandoning the tools. It argues for refusing to let the circuit close — demand evidence before conclusions, commit before exposure, require every objection to name its check, restart threads that have taken a wrong turn rather than negotiating with them. And keep one domain where the work is done the expensive way, not out of nostalgia, but because judgment is the instrument every other item on this list depends on.

The model is not an oracle diminished by an unworthy questioner. It is an instrument whose reading depends on what the operator brings — and no instrument has ever calibrated itself. Ground the circuit from outside, or accept that whatever it settles on will feel exactly as convincing as the truth.

References#

The oracle's actual size

• Polanyi's paradox — secondary reference — named by David Autor in 2014 after Michael Polanyi's The Tacit Dimension (1966).
• Goldberg, K. (2025). Good old-fashioned engineering can close the 100,000-year "data gap" in robotics​. Science Robotics​, 10, eaea7390, 27 August 2025. Journal · open PDF.
• Franco, A., Malhotra, N., & Simonovits, G. (2014). Publication bias in the social sciences: Unlocking the file drawer​. Science​, 345(6203), 1502–1505. Journal · open PDF. — 221 studies tracked from funded proposal to outcome.
• Alhanai, T., Kasumovic, A., Ghassemi, M., Zitzelberger, A., Lundin, J., & Chabot-Couture, G. (2025). Bridging the Gap: Enhancing LLM Performance for Low-Resource African Languages. AAAI 2025. — ~1M human-translated words across eight languages, evaluated across eleven; GPT-4o showed a 12.0–19.9 point absolute gap versus English on Winogrande, translated MMLU clinical sections, and Belebele. The paper does not isolate training-data volume as the sole cause.
• Christiano, P., & Xu, M. (2021). Eliciting Latent Knowledge. Alignment Research Center. — A problem formulation, not an empirical result.
• Cywiński, B., Ryd, E., Rajamanoharan, S., & Nanda, N. (2025). Towards eliciting latent knowledge from LLMs with mechanistic interpretability. Preprint. — A model trained to withhold a secret word; the recovery does not generalize to naturally acquired knowledge, and the essay says so.
• Apollo Research. The Evals Gap.

The narrowing

• Tankelevitch, L., et al. (2024). The Metacognitive Demands and Opportunities of Generative AI. CHI 2024.
• Laban, P., Hayashi, H., Zhou, Y., & Neville, J. (2025). LLMs Get Lost In Multi-Turn Conversation. Outstanding paper, ICLR 2026. — 200,000+ simulated conversations, 15 models, six generation tasks; 39% average drop, decomposing into minor aptitude loss and large unreliability increase.
• Jain, S., Park, C., Viana, M., Wilson, A., & Calacci, D. (2026). Interaction Context Often Increases Sycophancy in LLMs. CHI 2026. — Two weeks of real context from 38 users; memory profiles produced the largest agreement-sycophancy increases (+45% for Gemini 2.5 Pro), synthetic non-user context still raised it for some models (+15% for Llama 4 Scout), and perspective sycophancy rose only where viewpoint inference was accurate.
• Dubois, M., Ududec, C., Summerfield, C., & Luettgau, L. (2026). Ask Don't Tell: Reducing Sycophancy in Large Language Models. UK AI Security Institute. Preprint.
• Potts, C., & Sudhof, M. (2026). A Paradox of AI Fluency. Preprint, Bigspin AI — a company selling conversation monitoring, which is a conflict worth naming. 26,958 annotated WildChat transcripts. Failure rates 64% vs 24% are confounded by task complexity (3.1 vs 1.5) and the essay does not treat them as a skill effect; the visible-failure split (59% vs 12%) is the load-bearing figure.

The hand-back

• Elish, M. C. (2019). Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction. Engaging Science, Technology, and Society​, 5, 40–60.
• Green, B. (2022). The flaws of policies requiring human oversight of government algorithms​. Computer Law & Security Review​, 45. Journal · open preprint. — 41 policies surveyed.
• Passi, S., & Singh, R. (2026). The Oversight Fallacy. Data & Society, 29 July 2026. doi:10.69985/VWCK1626.
• Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To Trust or to Think. PACM HCI​, 5(CSCW1), Art. 188. — 199 participants; three forcing designs versus two XAI designs. Forcing reduced overreliance, was liked less, and helped most those already high in need for cognition. The result is about those interventions on those tasks, not about all possible interventions.
• Xu, K., Shen, Y., Yan, L., & Ren, Y. (2026). Cognitive Agency Surrender. Preprint. — Classification of 1,223 AI-HCI papers: frictionless usability 67.3%, epistemic-sovereignty research 19.1% → 13.1%, machine agency rising to 19.6%. The cognitive-closure mechanism is their theory​, offered alongside the survey, not a mediation finding.
• METR (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. — 16 developers, 246 issues; 19% slower measured, 24% faster forecast, 20% faster believed afterward. A 39-point gap on completion time, in one narrow population.

Erosion, and its remedy

• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics​. PNAS​. Journal · open full text. — ~1,000 students, three arms. Assisted practice +48% (GPT Base) and +127% (GPT Tutor); on the subsequent unassisted exam, GPT Base scored 17% below the never-had-access control while GPT Tutor matched it. The load-bearing study for the essay's erosion claim, and for the claim that erosion is a design property.

What the rule organizes

• Jain, R., Bridgers, S., Janzer, L., Greig, R., Teh, T. H., & Mikulik, V. (2026). Human-AI Complementarity: A Goal for Amplified Oversight. FAccT 2026. — 1,918 items: human majority vote 80.6%, AI alone 87.7%, confidence hybridization 89.3%, 91.3% with evidence-assisted humans. Labels/explanations/confidence produced over-reliance; search results and evidence were "the only form of assistance that statistically significantly improve[d] performance." Complementarity described by the authors as "a moving target."
• Kohli, G. (2026). Nine Judges, Two Effective Votes. Preprint. — Kish effective sample size 2.18, 95% CI [2.07, 2.31]; best single judge matched or beat the panel in every condition. Effective sample size measures variance equivalence, not a count of opinions.
• Kim, E., Garg, A., Peng, K., & Garg, N. (2025). Correlated Errors in Large Language Models. ICML 2025. — 350+ models; correlation highest among the larger and more accurate, across providers and architectures. Cross-sectional, not a time trend.

The case against

• Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence​. Science​. Journal · open PDF. — 453 professionals; time −40%, quality +18%, inequality between workers down.
• Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. — 55.8% faster on a defined HTTP-server task. A narrow, well-specified task; that narrowness is the point of citing it beside METR.
• Dell'Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. (2023). Navigating the Jagged Technological Frontier​. SSRN · open PDF. Later in Organization Science​. — 758 BCG consultants; large gains inside the frontier, 19 points worse outside it.
• Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour​, 8, 2293–2303. Preprint. — 106 studies, 370 effect sizes. Combinations performed significantly worse than the best of human or AI alone; gains on creation tasks, losses on decision tasks; gains where the human outperformed the AI, losses where the AI outperformed the human.

What compounds

• Brynjolfsson, E., Li, D., & Raymond, L. (2023). Generative AI at Work. NBER WP 31161. — 5,179 agents; +14% average, +34% novice, minimal for the most experienced. The authors read this as diffusion of stronger workers' practices.
• Cotton, C. S., & Scholle-Cotton, L. (2026). The AI Expertise Paradox. Issues in Science and Technology​, 3 February 2026, responding to David Autor.
• Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the Coal Mine?. Stanford Digital Economy Lab, November 2025. — 16% relative employment decline, ages 22–25, most AI-exposed occupations, controlling for firm-level shocks. An employment correlation, not a demonstration of deskilling.
• Anthropic (2026). Economic Index report: Learning curves. — 10% higher success rate at 6+ months tenure is unadjusted; controlling for O\​*NET task and request cluster gives roughly 3 percentage points, 4 with full controls. Anthropic flags survivorship and cohort effects.
• Pew Research Center (2026). Americans and AI 2026. Published 17 June 2026, fielded 17–23 February 2026, n = 5,119.