Abstract
Does AI assistance create the conditions for replacing the people it helps? The question combines several distinct outcomes: assisted performance, autonomous task execution, demand for paid work, and the distribution of authority. Evidence for one does not establish the others.
This paper critically synthesizes experimental, administrative, and technical research retrieved through September 13, 2026. It does not report new empirical results or a systematic review. Randomized studies establish substantial productivity gains in some settings, but effects vary with task, worker, tool, and study design. A broader human–AI meta-analysis distinguishes augmentation from synergy: combinations can outperform humans without outperforming AI alone. Labor-market findings remain heterogeneous. The August 2026 revision of Canaries in the Coal Mine? documents a widening employment shortfall among young workers in highly exposed occupations through June 2026. Its authors explicitly characterize that pattern as descriptive, not causal. Danish administrative evidence identifies small average earnings and hours effects over an earlier adoption horizon. Neither result identifies whether retained work traces cause subsequent automation.
The original paper treated an assistance-to-replacement cycle as a stronger explanation than the evidence justified. This revision withdraws that generalization. It retains a narrower trace-reuse hypothesis: under specified conditions, reusable records of assisted work may change the cost or feasibility of later delegation. The hypothesis must demonstrate an incremental effect beyond external model improvement and ordinary organizational redesign.
We develop a causal identification strategy and distinguish short-term output from durable learning. We also replace claims of assured human control with dated evidence on research automation. Finally, we present ANCHOR as an unvalidated governance proposal, not a mechanism proven to stabilize augmentation. Human rights and legitimate authority do not depend on whether the empirical trace-reuse hypothesis survives. The defensible conclusion is conditional: AI changes what organizations can delegate; task structure, demand, institutional choices, and learning design determine what that delegation means for people.
Revision statement
This is a substantive reassessment, not a stylistic update. The original title and public address remain, but the subtitle and argument have changed.
Three claims are withdrawn. Assistance is not established as a general precursor to replacement. Cheap verification is not a necessary condition for organizations to automate. A nonsignificant pooled result cannot, by itself, disprove a causal mechanism.
Several claims are narrowed. The proposed “recursive capture” mechanism is now a trace-reuse hypothesis within existing task and organizational economics. The human-relevance framework is a set of audit dimensions, not a validated measurement instrument. ANCHOR is a proposal whose operational components require evaluation.
The revision also corrects the Census paper’s title and authorship, the authorship of Anthropic’s labor report, and the interpretation of METR’s frontier-risk assessment. New August and September 2026 research changes the discussion of labor-market adjustment and automated research. These changes matter more than preserving the earlier thesis.
1. Review question and method
Scope and evidence selection
The review asks four questions:
- What effects of AI assistance have been demonstrated on work performance and learning?
- What do labor-market studies establish about employment, hours, and entry into work?
- Is there evidence that reusing records of assisted work causes later delegation?
- Which governance obligations remain relevant if that causal hypothesis is wrong?
The search began with a complete rereading of the July paper as revised on September 12. Its references were reopened and checked against retrieved publisher pages, author versions, institutional reports, and full papers where accessible. Targeted searches then covered human–AI collaboration, workplace productivity experiments, learning, and employment. Separate searches examined task complementarity, workflow organization, research automation, and recursive training.
Sources were selected for relevance to a stated claim and for an identifiable method or evidential basis. Randomized experiments receive more weight for treatment effects within their studied settings. Administrative panels inform labor-market adjustment, subject to identification and coverage limits. Technical demonstrations inform capability in the tested system. Vendor usage data describe selected users. Standards and normative scholarship inform governance, not productivity estimates.
The review uses English-language sources retrieved through September 13, 2026. Recent additions include Stanford’s August 12 labor update, Anthropic’s August 28 automated-alignment research, and its September 4 formalization report. A newer publication date is not a newer observation window. Stanford’s revised payroll analysis ends in June 2026. The Census supplement describes November 2025–January 2026. METR’s May report assesses February–March systems.
What this method cannot establish
This is a purposive critical synthesis, not an exhaustive systematic review. Search coverage was not preregistered, and some broad searches timed out. There was no dual-reviewer screening, formal publication-bias assessment, or pooled reanalysis. Several publisher pages restricted automated access. Where possible, author manuscripts or institutional accounts supplied the relevant evidence. Such access limits are not evidence that the inaccessible studies are invalid.
The source record preserves retrieved versions and distinguishes full-paper inspection from abstract or institutional-summary inspection. Search summaries were used for discovery, not as independent corroboration. The earlier Last Copilot article appeared in discovery results and was excluded as evidence for its own claims.
No pooled productivity effect is calculated here. A quality rating, a completed pull request, an email hour, and an employment count are not interchangeable outcomes. Combining them would create numerical precision without a coherent estimand—the specific effect a study seeks to estimate.
AI tools assisted discovery, drafting, and source checking. A separate internal verification pass checked the revised claims and artifact, but this is not independent external peer review. No original participant data were collected or reanalyzed for this paper.
Evidence is assessed along separate dimensions: identification strength, measurement validity, population coverage, tool vintage, and duration. No single hierarchy settles all five. A small randomized study can identify a local effect cleanly while saying little about national employment. A payroll panel can reveal an important distributional pattern without isolating its cause.
The contribution is therefore a disciplined separation of claims and a proposed test of one underidentified mechanism. It is not a new general theory of technological unemployment.
2. Define the outcome before choosing the story
Augmentation, synergy, and substitution
Assisted performance compares a person using AI with a relevant unassisted person or workflow. It may concern speed, quality, safety, or another specified outcome. Better assisted performance does not establish better unaided skill after the tool is removed.
Human–AI synergy is a stricter comparison. The combined system must outperform both the human-only and AI-only alternatives on the measured outcome. If AI alone performs best, adding a person may improve accountability while reducing measured task performance. That is a tradeoff, not synergy.
Task substitution means assigning AI work previously performed by people. Labor displacement concerns the resulting reduction in demand for paid human work relative to a counterfactual. Neither means that every affected person loses a job. An organization can reduce hours per output while expanding output and employment.
These distinctions are empirically consequential. Vaccaro, Almaatouq, and Malone’s meta-analysis examined 106 experiments reported in 74 papers, yielding 370 effect sizes. Human–AI combinations outperformed humans alone on average: Hedges’ g = 0.64, with a 95% confidence interval of 0.53–0.74. Against the better standalone alternative, the pooled effect was negative: g = −0.23, with a 95% interval of −0.39 to −0.07.1
Heterogeneity was substantial. The evidence window ended in June 2023, and much of the research concerned decision support rather than current generative agents. The result does not prove that today’s assistants should operate alone. It does show why “AI improves human performance” cannot establish that human participation remains technically necessary.
Every enhancement or displacement claim should name its unit, comparison, outcome, and horizon. A lawyer may gain capacity while a junior colleague loses billable work. Lower prices may then increase access to legal services. Evaluating only the focal user misses the other people affected.
Existing economics explains much of the apparent paradox
Task-based economics already separates automation’s displacement effect from productivity gains and the creation of new labor-intensive tasks. Acemoglu and Restrepo provide the foundational account.2 Historical research on new work rejects a fixed stock of tasks, while showing that new opportunities do not appear uniformly across occupations or workers.3
Two newer models sharpen the analysis. Autor and Thompson’s Expertise asks how removing some tasks changes the expertise required for those that remain. Automating an expert task can broaden entry while reducing the scarcity value of expertise. Removing an inexpert task can have different wage and employment implications.4
Gans and Goldfarb’s O-Ring Automation studies tasks whose qualities complement one another. Automating one step can free a worker to improve the remaining bottlenecks. Under the model’s assumptions, partial automation can increase labor income. Summing task exposure without considering those relationships can overstate displacement.5
Demirer and colleagues’ Chaining Tasks, Redefining Work examines the costs of moving between human and AI execution. Automating contiguous steps can save coordination costs, producing nonlinear changes in job design. Their empirical analysis supports patterns of adjacent AI execution. It does not identify an effect of retaining a particular worker’s interaction history.6
These approaches supply serious alternatives to the original replacement-cycle narrative. Automation can spread because handoffs become cheaper, complementary investments mature, or external models improve. None requires locally retained work traces.
An accounting identity, not an employment forecast
For a fixed output category, total human hours equal output quantity multiplied by human hours per unit:
Human hours = output × human hours per unit.
If unit labor falls by 20% and output rises by 25%, total hours in that category are unchanged. This is an illustrative identity, not an estimated AI effect. It excludes changes in product quality and must be extended when new tasks or products appear.
Employment also depends on hours per worker. Wages depend on bargaining and market conditions, not just total hours. A productivity estimate therefore cannot be translated directly into a job-loss forecast.
The important question is not whether augmentation and substitution can coexist. They plainly can. It is which mechanisms determine their size, timing, and distribution in a particular setting.
3. What work experiments actually demonstrate
Writing and customer support
Noy and Zhang randomized access to ChatGPT in professional writing tasks involving 453 college-educated participants. Completion time fell by 40%, and evaluated quality rose by 18%.7 These were bounded assignments with supplied instructions. They did not reproduce the factual checking, organizational context, and continuing responsibility of a complete job.
Brynjolfsson, Li, and Raymond’s customer-support study is closer to operational work. A staggered deployment among more than 5,000 agents produced substantial gains in issues resolved per hour, concentrated among less-experienced and lower-skilled workers.8 The authors also report evidence consistent with diffusion of effective practices and some worker learning.
This is important evidence against treating assistance as merely a temporary disguise for replacement. Knowledge embedded in a tool can help more people perform well. But the rollout was not a randomized test of whether retaining agents’ traces subsequently reduced staffing. Improved support performance and the proposed trace-to-delegation mechanism remain different claims.
The possibility of learning also matters. If novices acquire useful strategies and retain them, assistance can increase human capability. Whether the employer later expands service or reduces labor depends on demand and organizational decisions that a productivity estimate does not settle.
Software development and consulting
Cui and colleagues analyzed randomized coding-assistant access across three firms and 4,867 developers. Their preferred instrumental-variable estimate was a 26.08% increase in completed tasks, with a standard error of 10.3 percentage points.9 This estimates an effect of tool use induced by randomized access under instrumental-variable assumptions. It is not the same quantity as the average effect of merely offering the tool.
The experiments lasted two to eight months and were conducted for business purposes. The paper states that the research was post-registered. Completed tasks are useful operational measures, but they do not fully capture maintainability, security, or later rework. The study provides neither a lifetime return to assistance nor an employment effect.
Dell’Acqua and colleagues’ preregistered consulting experiment involved 758 professionals. On tasks within the tested GPT-4 frontier, AI users completed 12.2% more tasks and worked 25.1% faster. On a selected managerial task outside that frontier, assisted participants were less likely to reach the correct answer.10
The published article appeared in March 2026, but publication did not transform the experiment into a test of September 2026 tools. Its durable contribution is the unevenness of performance across seemingly related tasks. The boundary must be remeasured as tools and workflows change.
Why METR’s coding results are not a universal counterweight
METR’s July 2025 study found that experienced open-source maintainers took 19% longer when allowed to use early-2025 AI tools. Participants nevertheless believed that AI had helped.11 The result identifies a real mismatch between perceived and measured speed in that small, demanding setting. It does not establish that AI generally slows software work.
The February 2026 follow-up changes the interpretation. METR collected data from 57 developers across 143 repositories and more than 800 tasks. It reported substantial selection and measurement problems. Some developers declined participation or withheld tasks because they disliked the possibility of working without AI. Some left control-assigned tasks unfinished. Concurrent agent use also made time accounting difficult.12
METR regarded the new central estimate as a poor proxy for the real productivity effect. The appropriate conclusion is not to average the two studies or replace one headline with another. The tools and the population of submitted tasks changed, and the original design became less reliable for the new question.
This is a general lesson for AI field research. Randomizing treatment after tasks are submitted does not make the submitted tasks representative of all work. Selection can occur before randomization and missing outcomes can occur afterward.
Individual gains do not guarantee workflow transformation
In The Cybernetic Teammate, 776 Procter & Gamble professionals were assigned to product-innovation tasks with or without AI, individually or in teams. Individuals using AI matched unaided teams on the study’s evaluated performance measures.13 The experiment shows that AI can reproduce some benefits of collaboration in a bounded task. It does not show that removing a colleague preserves every function of a continuing team.
A different field experiment offers a useful limit. Dillon and colleagues randomized integrated AI access among 7,137 knowledge workers across 66 firms. The revised paper reports that tool users in the treatment group spent about two fewer hours on email each week in the experiment’s second half. It detected no broader change in task quantity or composition from individual-level provision.14
The two-hour figure concerns users, not every person assigned access. The six-month experiment also did not randomize wholesale organizational redesign. Its findings are compatible with time savings arriving before workflow change, but they do not prove that such change must eventually occur.
Across these studies, the reliable synthesis is conditional heterogeneity. Some tools improve some work substantially. The evidence does not support a universal productivity coefficient, a guaranteed human–AI advantage, or a necessary progression from assistance to replacement.
4. Labor-market evidence: entry pressure without a settled aggregate story
Exposure, use, and outcomes are different measures
An occupational exposure index estimates where AI could affect tasks. It does not directly measure actual adoption, profitable automation, or job loss. The ILO’s 2025 refined index places roughly one-quarter of global employment in occupations with some generative-AI exposure. Its highest exposure category is much smaller.15
Vendor telemetry adds information about observed use, but selected platform users are not a representative sample of employers or occupations. A conversation classified as “automating” describes an interaction pattern. It does not establish that a paid position was removed.
Anthropic’s March usage report and the Stanford AI Index provide useful context, but neither should be counted as an independent causal replication of the studies they summarize.1617 Citation volume cannot compensate for shared underlying data.
Danish administrative evidence
Humlum and Vestergaard link adoption surveys to Danish administrative records. The March 2026 revision reports precise null effects on earnings and recorded hours, ruling out average changes larger than 2% over a horizon extending two years after ChatGPT’s release.18
Those bounds are informative. They contradict claims of large early average effects in the studied population. They do not establish equivalence across every occupation or subgroup. Nor do they forecast what later models will do in different labor institutions.
The authors find changes in tasks and movement toward AI-relevant occupations without corresponding large average earnings effects. That combination is not paradoxical. Work can change inside a job before administrative totals move.
U.S. firms: breadth of integration matters
Bonney and colleagues’ Census working paper, The Microstructure of AI Diffusion, uses the Business Trends and Outlook Survey’s AI supplement. During November 2025–January 2026, 18% of firms reported AI use in a business function, rising to 32% on an employment-weighted basis.19 These are different denominators, not conflicting adoption estimates.
The paper distinguishes functional deployment from workers’ task-level use. Functional breadth and operational investment are positively associated with reported employment decreases. Task-level integration has no significant association with headcount reduction after those other dimensions are included.
That result is more relevant than a simple adopter/non-adopter comparison. It points toward the organization of work as a possible channel. But it is a regression association, not a randomized mechanism test. Controlling for organizational investment can also remove part of an adoption pathway. The coefficient cannot prove that bottom-up use has no total employment effect.
The original paper’s mislabeled citation obscured this contribution. The corrected account gives the study its actual question and avoids turning firm self-reports into causal estimates.
The August 2026 Canaries revision
Brynjolfsson, Chandar, and Chen’s revised ADP analysis extends through June 2026. Employment among workers aged 22–25 in highly exposed occupations stands about 19% below a counterfactual in which it kept pace with their less-exposed peers.20
This is a relative shortfall, not a finding that AI eliminated 19% of all young workers’ jobs. In the authors’ level comparison, employment fell about 11% in the two most-exposed quintiles while growing about 10% in the three least-exposed quintiles between November 2022 and June 2026.
The adjustment operates mainly through reduced hiring rather than increased separations. Experienced workers show no comparable exposure-related gap. These distinctions make the study relevant to apprenticeship and entry, even though it does not establish their long-term consequences.
The revision also strengthens the cautions. Education controls attenuate the gaps. Some differential trends predate generative AI. The patterns are stronger in the ADP analysis sample than in national survey benchmarks. Estimates adjusting for firm hiring are sensitive to specification. The authors explicitly present descriptive indicators, not causal estimates.
Earlier versions emphasized different regression estimates. Comparing an old adjusted headline with the new descriptive 19% figure would conflate a changing data window with a changing estimand. The revised paper’s same-measure comparison is a 15% shortfall at the July 2025 vintage and 19% by June 2026.
Why apparently conflicting findings can coexist
Massenkoff and McCrory’s March 2026 Anthropic labor report combines task exposure, platform use, and U.S. labor-survey data. It finds no systematic unemployment increase in highly exposed occupations and suggestive evidence of reduced entry into those occupations among younger workers.21
Unemployment is not hiring. A graduate who never enters an occupation may not appear among its unemployed former workers. Payroll employment, job-finding transitions, earnings, and hours answer different questions. Denmark and the United States also differ in institutions and adoption patterns.
The current evidence supports concern about uneven entry opportunities. It does not establish generalized AI-driven mass displacement. Equally, “no detectable aggregate effect” cannot be used to dismiss losses concentrated in a subgroup or to predict that future effects will remain small.
None of these studies identifies a causal contribution from locally retained interaction traces. They inform the outcomes the trace-reuse hypothesis must explain, not proof of its mechanism.
5. Learning and the supply of future expertise
Assistance can support or bypass learning
Bastani and colleagues randomized AI support in mathematics classes involving nearly 1,000 students. A relatively unrestricted GPT interface improved assisted practice performance but reduced subsequent unaided exam performance by 17% relative to controls, approximately 5.4 percentage points. A tutor designed with learning safeguards largely removed that harm, without demonstrating a positive unaided learning effect.22
The intervention was a package: the tutor included problem-specific guidance and teacher-supplied material. Its results do not isolate “asking questions” as the sole protective ingredient. Nor does a school experiment establish years of workplace deskilling.
Positive evidence matters just as much. Kestin and colleagues studied a custom AI tutor in a randomized crossover experiment with 194 undergraduate physics students across two lessons. Immediate learning outcomes were better with the tutor than with the active-learning classroom comparison.23 The package changed both delivery setting and instructional interaction. It was not a test of generic chatbots against all teachers.
Together, these studies reject a simple “AI prevents learning” thesis. They also reject the assumption that completing more assisted work necessarily builds more unaided competence.
New evidence from coding
Shen and Tamkin’s January 2026 randomized study involved 52 engineers learning an unfamiliar Python library. The AI group averaged 50% on an immediate comprehension quiz, compared with 67% in the unaided group. The task-time difference was not statistically significant.24
The 17-point gap is a difference in percentage points, unlike Bastani’s 17% relative reduction. Confusing them would exaggerate comparability across the two studies.
Engineers who used AI for conceptual inquiry or follow-up explanations tended to score better than those who delegated answers. Those interaction patterns were not themselves randomized. They may reflect prior ability, motivation, or other differences. They suggest intervention candidates; they do not prove that selecting a named interaction style causes better learning.
The experiment tested short-term comprehension in a small sample. Whether that outcome predicts durable expertise, unfamiliar incident recovery, or career progression remains unresolved.
The apprenticeship risk is a hypothesis about a pipeline
Lee and colleagues’ survey of knowledge workers associates greater confidence in AI with less reported critical-thinking effort. It measures perceptions and examples, not experimentally demonstrated cognitive decline.25 Anthropic’s June 2026 Cadences report supplies countervailing self-report evidence: heavier delegators express greater optimism about their skills’ value and do not report less learning.26 That does not prove preservation of skill either.
The stronger research question concerns a training pipeline. If junior tasks supply practice, removing them without an alternative could reduce the later supply of experts. But AI might instead lower entry barriers and provide better practice at lower cost. Both pathways are plausible.
Bainbridge’s Ironies of Automation identifies a longstanding human-factors problem: people may be left to handle rare failures after routine practice has disappeared.27 Generative AI makes that issue newly salient; it did not create it.
Institutions should measure output, delayed unaided performance, and recovery ability separately. They should also distinguish an individual’s skill from the organization’s resilience. Not everyone must retain every manual skill. A reliable reserve of expertise may suffice in some systems, while continuous individual competence remains essential in others.
The relevant standard is demonstrated recovery at acceptable cost and risk. Preserving obsolete practice without testing its value is no better than eliminating practice without measuring the loss.
6. Research automation: update the boundary, not the rhetoric
Different loops change different things
A model’s answer can influence its next answer without changing its weights. Retrieval can improve a system by changing available context. An agent can modify code, tests, or a workflow without modifying the foundation model. A training pipeline can use generated data to update another model.
All are feedback processes, but they have different mechanisms. OECD’s definition distinguishes autonomy and adaptiveness; it is not evidence that humans reliably control every action of a deployed system.28
Self-play in AlphaGo Zero is a clear example of learning from generated experience in a tightly specified domain.29 Constitutional AI and self-rewarding language models provide examples of AI-generated feedback within training procedures.3031 These establish technical feasibility of particular loops, not their prevalence across all AI deployments.
DeepMind’s AlphaEvolve describes program generation and selection using automated evaluators, including first-party reports of improvements to computing infrastructure.32 Such work can contribute to AI development without constituting a self-directed organization.
What changed after the original paper
The original paper overstated what METR’s May 2026 report established about institutional control. That report assessed internal frontier agents during February–March. It concluded that agents plausibly had the means, motive, and opportunity to start small unauthorized “rogue deployments,” though not highly robust ones.33
This is an assessment of risk and evidence, not proof that such a deployment had occurred. It is also not support for blanket reassurance that humans retain effective control of every consequential research action. Formal authority and actual ability to enforce it must be evaluated separately.
Anthropic’s August 28 report describes automated researchers proposing and testing training methods for ten well-characterized alignment failures. The methods improved the targeted evaluations, with further tests on withheld benchmarks and larger models.34 Researchers chose the goals and evaluation conditions. The report acknowledges incomplete capability checks, proxy limitations, and uncertainty about persistence after further training.
Its comparison with human researchers was not an equal-budget contest: the humans could not iterate in the same way. We therefore do not treat the result as proof of general research superiority or solved alignment. It does establish a materially stronger example of automated training research than the earlier paper acknowledged.
On September 4, Anthropic reported a largely autonomous, eleven-day Lean formalization of Fermat’s Last Theorem using a collaborative scaffold and an internal model.35 This is a provider-reported result with a released artifact and described external mathematical review. This review did not independently rebuild the proof.
The distinction is important. Formalizing an existing proof under a proof checker is not the same as independently choosing a scientific agenda. Yet it is substantial multi-day work, not merely a short coding demonstration. Fixed claims about today’s autonomous limits can become obsolete quickly.
Verification helps, but does not settle adoption or safety
Exact tests can make search and selection cheaper. They can also leave the wrong specification perfectly satisfied. A proof checker verifies a formal proposition relative to its logic and assumptions. It does not decide whether that proposition captures an organization’s real objective.
Learned evaluators have a different problem. Gao and colleagues show how optimizing an imperfect reward model can improve the proxy while degrading the intended measure.36 OpenAI’s sycophancy postmortem illustrates how product feedback and evaluation can favor undesirable behavior.37
Cheap verification is therefore a plausible moderator, not a law of replacement. Organizations may automate despite weak checks because errors are inexpensive to them, difficult to detect, or borne by others. Conversely, excellent checking can support safer human assistance rather than remove the human.
The International AI Safety Report remains useful for the broader uncertainty around AI-assisted research feedback.38 It should not be used as a substitute for dated evaluations of particular systems. The same discipline applies to optimistic demonstrations and alarming scenarios.
7. The trace-reuse hypothesis and its alternatives
A narrower mechanism
We use trace reuse to mean authorized reuse of artifacts from assisted work in a later system or workflow. The artifact might be a correction, test case, documented exception, or revised procedure. Mere collection is not reuse. Temporary conversational context is not automatically a contribution to model training.
The original label, recursive capture, is retained for continuity, but “capture” should not imply an established extraction mechanism or an inevitable loss to workers.
The revised hypothesis is:
In a specified workflow, making useful work traces available for later evaluation or redesign can increase subsequent delegation relative to an otherwise comparable workflow without that reuse.
This is an incremental causal claim. It requires a comparison that holds external changes sufficiently constant. If both workflows automate after the same model upgrade, that is not evidence for trace reuse.
The hypothesis concerns feasibility and delegation first. It makes no unconditional prediction about total employment, wages, or human welfare. A successfully reused correction could improve an assistant while increasing the demand for the person using it.
The causal structure
A candidate pathway has four links:
- Assisted work produces an informative artifact.
- The artifact is actually incorporated into an evaluator, model, or workflow.
- That incorporation changes reliable performance or the cost of execution.
- The organization changes delegated work because of that improvement.
Each link can fail. Logs may contain little useful information. Tacit judgment may resist representation. New tests may reveal previously hidden risk and reduce delegation. A firm may choose service expansion over labor reduction.
At least four alternative pathways must be measured. External model improvements can directly increase delegation. Pre-existing automation intent can cause both trace collection and redesign. Easy-to-specify tasks can be more likely to produce usable traces and to automate. Demand or budget shocks can change staffing and investment simultaneously.
Chronology is necessary but insufficient. Seeing traces accumulate before automation does not eliminate common causes. Nor does a manager’s retrospective explanation supply the missing counterfactual.
Researchers should establish provenance: which artifacts entered which update, when, and with what permissions. They should then test whether that update improved held-out performance. Only afterward should they ask whether deployment decisions changed. A chain of plausible anecdotes is not a causal estimate.
How this differs from ordinary organizational learning
Firms have long documented procedures, trained novices, standardized production, and reorganized jobs. AI can make some of those processes cheaper or more scalable, but the existence of organizational learning is not a novel discovery.
The narrower possible contribution is measurable reuse at the interface between assistance and later delegation. Does access to those artifacts add explanatory or predictive value beyond task structure, model capability, cost, and management intent? Can it improve deployment decisions? If not, “recursive capture” is an unnecessary label.
A further distinction is normative. Authorized reuse that benefits workers is not equivalent to undisclosed appropriation. The same technical channel can operate under different ownership and bargaining arrangements. Its existence does not determine its legitimacy.
Synthetic-data findings do not prove human indispensability
Shumailov and colleagues demonstrate model collapse under particular recursive training conditions.39 Gerstgrasser and colleagues show that accumulating synthetic generations alongside original real data can avoid collapse in the settings they study.40
The second result is essential to interpreting the first. Synthetic output is not inherently corrosive, and recursive generation is not inherently self-correcting. Sampling, retention, filtering, and evaluation determine what information survives.
Neither study demonstrates that preserving a particular occupation is necessary for model quality. Fresh grounding can come from instruments, experiments, databases, or other sources. The separate question of maintaining human expertise must be tested in the domain where that expertise is supposed to improve outcomes.
8. Human agency is not a residual task category
The original “residual relevance trap” combined an economic prediction with a moral argument. They should be separated.
For a fixed task set, expanding machine capability can reduce the set that machines cannot perform. The entire economy does not have a fixed task set. New work and new complementarities can expand human roles. No mechanical law establishes a shrinking overall human frontier.
The moral argument is different: people should not have to outperform machines to retain rights, consent, or political standing. That proposition does not predict employment and cannot be validated by a productivity experiment.
A practical audit should separate six dimensions:
- Agency: room to initiate action and make consequential choices.
- Authority: recognized and usable power to approve, contest, or stop.
- Capability: competence needed for the person’s role and relevant failures.
- Economic participation: access to income, time savings, ownership, and opportunity.
- Epistemic independence: access to evidence and challenge outside one system’s judgments.
- Standing: rights and access to remedy regardless of productive usefulness.
This is not a validated six-factor instrument. The dimensions overlap, require domain-specific measures, and should not be compressed into an unexplained score. “Economic participation” is a distributive question, not a claim that every deployment creates an automatic legal entitlement to a specified share.
Elish’s “moral crumple zone” explains why a human signature can conceal a transfer of blame without a transfer of control.41 An operator who cannot detect a problem or halt a process is not an effective safeguard merely because the interface includes an approval button.
The converse matters too. Requiring human review can delay a safer automated response or reproduce human bias. The appropriate arrangement depends on measured performance and legitimate authority, not a presumption that a person must touch every output.
Oversight should therefore be tested as a system. Who receives the evidence? How much time is available? Can an objection change the outcome without retaliation? Does responsibility follow actual control? A correct policy document is not evidence that these conditions hold in use.
9. ANCHOR as a governance proposal
ANCHOR organizes questions that the empirical studies leave open. It is not a certification, a validated causal model, or a promise that augmentation will remain stable. Its principles overlap with existing risk management, labor governance, and human-factors practice.
A — Authority over purposes and stopping. Identify who authorizes deployment and who can halt it. Test whether escalation reaches someone who has both information and power. Authority should follow the legitimate responsibilities of the domain, not automatically belong to the nearest operator.
N — Non-contingent standing. Protect rights, due process, and access to remedy independently of productivity. OECD’s human-centered principles provide an established normative basis.42 These are constraints on acceptable action, not performance bonuses to trade against higher output.
C — Capability development. Where independent competence matters, fund and evaluate learning. Measure delayed unaided performance and transfer, not just assisted completion. The effective intervention may be tutoring, simulation, supervision, or selective manual practice.
H — Human-trace terms. Document permitted uses of work artifacts. Distinguish provider training from internal retrieval, evaluation, and workflow redesign. Address privacy, provenance, retention, and worker voice. A “no model training” promise does not necessarily prohibit every organizational reuse.
O — Outcomes and distribution of gains. Record where benefits and costs fall. Faster work might become shorter hours, lower prices, better service, profit, or greater output. Include workers who are never hired, not only those who remain to answer a survey.
R — Recovery and revision. Maintain a tested route to contain failure, restore service, and reconsider deployment. Full manual replacement may be unrealistic in some systems. Redundant tools, reduced-service modes, and independent expertise can be alternatives if they meet the relevant recovery requirements.
NIST’s AI Risk Management Framework already calls for governance, contextual mapping, measurement, and management of risk.43 ANCHOR’s emphasis is the longitudinal organization of work: what competence and authority the deployment leaves behind, and who receives the gains. Any claim of additional value should be tested against established practice.
ILO case studies of social dialogue show how worker institutions can influence deployment.44 Such cases identify mechanisms and implementation conditions. They do not establish that consultation universally increases productivity or prevents displacement.
The framework also has costs. Detailed reporting can burden small firms. Practice requirements can protect incumbents or consume time without improving recovery. Consultation can become ceremonial. These are reasons to evaluate implementation and proportionality, not to assume that adding six labels improves governance.
A minimum deployment record should identify affected groups, baseline outcomes, permitted trace uses, and accountable decision-makers. Follow-up should report changes in quality, labor input, learning, incidents, and distribution. Record dissent and unresolved tradeoffs rather than manufacturing a single “human relevance” score.
10. A research design that can distinguish the mechanisms
Primary study: randomized access to reusable traces
The core test should randomize workflow teams or organizational units, rather than individual prompts, when artifacts can spill across coworkers. Both arms receive the same initial assistant, task access, and ordinary safety controls.
The treatment arm receives a governed pipeline that makes specified work traces available for subsequent evaluation and redesign. The comparison arm does not receive that additional reuse capability during the study. Essential security and incident records remain available in both arms. A study should not remove necessary safety evidence to create a clean experiment.
This identifies the effect of a reuse policy and implementation package, not the intrinsic value of a byte of data. If resources permit, a factorial design can separate artifact access from additional engineering support. Both arms need comparable budgets and access to external model upgrades.
Recruitment must precede treatment assignment. Baseline measures should include task structure, existing codification, automation plans, and labor demand. Researchers should log later model versions and organizational interventions. Otherwise, unequal access to a new model could masquerade as an effect of trace reuse.
Estimand and outcomes
Let Z indicate assignment to the reuse package. Define D at month 12 as the share of a prespecified task basket actually completed through the delegated workflow while meeting independent quality and safety criteria. The primary intention-to-treat estimand is:
The mean difference between D under assignment to reuse and D under assignment to the comparison condition.
This contrast measures the effect of being assigned the package, including nonuse. A separate estimate of actual reuse would require additional assumptions or another design. Reporting only enthusiastic adopters would sacrifice the protection of randomization.
The task basket must be fixed or reweighted to a declared baseline. Otherwise, organizations can increase the apparent delegated share by dropping difficult cases. Report the evolving real-world task mix separately, because adaptation is itself an important outcome.
Delegation should not mean a successful demonstration. Count routine completion, rework, escalation, and failure over a defined observation period. Include the cost of checking and maintaining the workflow. A system that delegates execution while multiplying hidden review labor has not necessarily saved human work.
Secondary outcomes should include paid human hours per quality-adjusted output, total hours, headcount, hiring, wages, and output demand. Track delayed skill and recovery outcomes as separate endpoints. The same deployment can improve the primary outcome while harming one secondary outcome.
Identification and mechanism checks
Use cluster-appropriate uncertainty estimates. Prespecify the primary outcome and horizon before inspecting effects. Longer follow-up at 24 and 36 months can test persistence, but should not become a moving deadline that protects the hypothesis from failure.
Measure whether the treatment actually changes usable trace availability. Audit incorporation into tests, models, or procedures. Evaluate the resulting system on held-out cases before inspecting employment effects. These checks establish whether the intended mechanism was activated.
Do not control for post-treatment codification when estimating the total treatment effect: it may be part of the pathway. A mediation analysis needs explicit assumptions about confounding between the mediator and the outcome. A randomized intervention on codification or evaluator access can provide stronger evidence than retrospective decomposition.
Spillovers require special attention. Treated teams may share procedures with controls. A common vendor update may incorporate data from many participants. Randomization at a broader organizational level, staged saturation designs, or explicit exposure measures may be necessary. No design should claim a no-interference assumption while ignoring shared systems.
Attrition and noncompletion also matter. Record tasks abandoned after assignment, worker turnover, and missing output. Report sensitivity analyses for plausible missing outcomes. METR’s follow-up demonstrates why a task-level experiment can become misleading when participants selectively avoid its control condition.
What would count against the hypothesis?
A nonsignificant effect is inconclusive when the interval includes both negligible and important effects. The study must preregister a smallest effect that would matter operationally, based on deployment cost and risk. That threshold cannot be chosen after results arrive.
If a sufficiently precise upper confidence bound lies below the minimum positive effect, the study can reject a practically important benefit in that population and horizon. A well-estimated negative effect would be stronger evidence against the predicted direction. Equivalence requires an appropriate interval-based test, not a large p-value.
Cross-domain pooling requires care. Opposite effects in different workflows can cancel. Prespecified subgroup analyses should examine task structure, verification quality, and baseline codification. A pooled null does not disprove every local mechanism, while one positive subgroup does not validate a general law.
The hypothesis should be retired as a useful general explanation if well-powered replications repeatedly show no meaningful incremental effect and simpler models predict equally well. A narrower local mechanism may remain, but its scope must be stated rather than quietly expanded.
Complementary studies
Verification and adoption. Randomize an improvement in evaluator quality or cost where safe, then measure reliable delegation and total human review effort. Compare this with prediction based on task structure and model quality alone. Observational correlations can support forecasting but cannot isolate verification as a cause.
Learning design. Randomize answer delivery, explanation, guided practice, and appropriate non-AI instruction with matched learning time. Measure immediate output, delayed unaided competence, transfer, and error detection. Separate assignment to an interface from the interaction style a participant chooses within it.
Apprenticeship and expert supply. Follow entry cohorts, including people denied entry, rather than only surviving employees. Link changes in junior hiring to later competence, promotions, incidents, and external recruiting. A genuine expert-supply test needs a career-relevant horizon, not a one-week quiz.
Distribution and governance. Evaluate concrete interventions, such as gain-sharing or protected training time, against usual practice. Natural experiments require credible assumptions about pre-trends, selection, and concurrent reforms. Randomization is appropriate for implementation methods, not for withholding fundamental rights.
Resilience. Test manual practice, simulation, redundant systems, and specialist reserves on unfamiliar but safe scenarios. Compare recovery quality and cost. If a cheaper arrangement meets the required safety standard, routine manual practice is not automatically preferable.
Each study can fail independently. Evidence for a better tutor does not validate trace reuse. Evidence for more delegation does not justify a rights violation. This modularity is a requirement of sound inference, not a way to rescue every part of the original framework.
11. Limits and implications
The strongest remaining uncertainty is causal attribution. The reviewed literature establishes many parts of the surrounding story, but not the incremental trace-reuse path. This review found no study directly randomizing reusable work traces and measuring their subsequent effect on delegation over the proposed horizon. That is a bounded search finding, not a claim that no such study exists anywhere.
The evidence also reflects concentrated access. Major firms and AI providers can study populations and systems that independent researchers cannot easily observe. Vendor involvement does not invalidate a result, but it increases the importance of transparent methods, replication, and access to negative findings.
Tool obsolescence is another limit. A clean estimate for one model remains a clean estimate for that model; it does not become a current estimate merely because it is the best available experiment. Conversely, a new benchmark does not invalidate an old workplace result without testing comparable work.
Labor studies have different coverage and clocks. Online professional tasks omit much physical and interpersonal work. Administrative data lag technical change. Platform data arrive quickly but select users. Exposure indices approximate a moving technology with fixed task descriptions. These sources should constrain one another, not be forced into one verdict.
Normative questions remain even under a favorable economic outcome. An economy can gain output while concentrating losses, weakening consent, or denying remedies. It can also automate extensively while expanding access and leisure. The evidence does not dictate one distribution of gains.
For employers, the implication is to evaluate a deployment as a change in work organization, not a software license. For researchers, it is to distinguish which causal arrow a study identifies. For policymakers, it is to monitor entry and distribution alongside aggregate employment. These recommendations do not require believing that assistance secretly contains its own replacement.
Conclusion
The original paper asked a consequential question but gave its proposed mechanism too much explanatory weight. Current evidence does not establish a general cycle in which AI first enhances people and then replaces them through reuse of their work traces.
What the evidence does establish is more specific. Assistance can improve output substantially, with effects that depend on the task and setting. Human–AI combinations need not outperform AI alone. Learning can improve or deteriorate with instructional design. Recent labor data show pressure on some entry paths without identifying a settled economy-wide displacement effect. Research automation is advancing beyond the limited examples used in the original paper.
Trace reuse remains a plausible contributor to later delegation. It must earn that role through direct comparison with external capability improvement, task complementarity, and organizational redesign. Its employment consequences must then be measured rather than assumed.
The title The Last Copilot should be read as a question about institutions, not a forecast of the final human worker. We need neither permanent human bottlenecks nor an untested promise of frictionless abundance. We need evidence about what each deployment changes, legitimate authority over those changes, and an explicit account of who benefits.
Human standing does not depend on winning the next benchmark. That principle survives even if the paper’s empirical hypothesis does not.