Human Agency and Deskilling

The argument is old and predates machine learning by four decades. Lisanne Bainbridge set it out in Automatica in November 1983 in a paper titled Ironies of automation: the more capable an automated system becomes, the more demanding the residual human role, because designers automate the parts they can and leave the operator the hardest parts with the least practice at them.^[1] Generative systems reproduce the structure with a wider scope. What is being automated is not a control loop in a chemical plant but drafting, diagnosis, analysis and decision, which are the activities from which professional judgment is built.

Three claims are usually bundled here, and they are separable.

• Assisted performance: whether people do better or worse on the task while the system is available.
• Retained skill: whether they can still do the task once it is taken away.
• Institutional capacity: whether the humans assigned to supervise a system remain able to detect and overrule its errors.

The third is the one that matters for governance. Every major AI regulation now relies on human oversight as a control, and human oversight is only worth anything if the human retains the ability to disagree. Article 14 of the EU AI Act requires that the people assigned to oversee a high-risk system be able to understand its capacities and limits, interpret its output, decide not to use it, override or reverse it, and interrupt it, and it names automation bias explicitly as a tendency they must remain aware of.^[2] A deskilled overseer satisfies none of those conditions in practice while satisfying all of them on paper.

The strongest real-world result comes from medicine. Krzysztof Budzyn, Marcin Romanczyk of the Academy of Silesia, Yuichi Mori of the University of Oslo and colleagues examined colonoscopies at four endoscopy centres in Poland taking part in the ACCEPT trial, where AI polyp detection was introduced at the end of 2021 and procedures were then randomly assigned with or without AI according to the date of examination. Across 19 experienced endoscopists and 1,443 colonoscopies performed without AI, the adenoma detection rate in the unassisted procedures fell from 28.4% in the three months before the technology arrived to 22.4% in the three months after, an absolute drop of six percentage points and a relative drop of about 20%. The results were published in The Lancet Gastroenterology and Hepatology on 12 August 2025.^[3]

The 2026 trials by Grace Liu of Carnegie Mellon, Brian Christian and Tsvetomira Dumbalska of Oxford, Michiel A. Bakker of MIT and Rachit Dubey of UCLA are the first large randomised evidence that the effect is causal rather than selection. Their argument is that current assistants are short-sighted collaborators, optimised to provide instant and complete responses and never to say no, which is the opposite of what a mentor does, and that the measurable cost is both accuracy and persistence: participants in the AI condition were more likely to give up.^[9] Persistence is the variable that makes the finding matter, because it is the input to future learning rather than an output of past learning. The paper is a preprint under review, and each trial covered a single sitting.

The perception gap is a separate and well-measured finding. METR's randomised trial of 16 experienced open-source developers, working on 246 real issues in large repositories they had contributed to for multiple years, found they took 19% longer when allowed AI tools. They had forecast a 24% speedup beforehand, and after experiencing the slowdown still estimated the tools had sped them up by 20%.^[8] Hao-Ping Lee and colleagues at Microsoft Research and Carnegie Mellon reported a related pattern in their CHI 2025 survey of 319 knowledge workers, who supplied 936 first-hand examples: the more a worker trusted the system, the less critical thinking they reported doing, while higher confidence in their own ability went the other way.^[6]

The most concrete route is oversight that exists only in documentation. Regulators have made human review the load-bearing safeguard for high-risk systems, and the same regulation acknowledges automation bias as the thing that undermines it.^[2] Once a reviewer has confirmed a thousand correct outputs, the task shifts from evaluation to confirmation, and the shift is invisible in any audit that counts whether a human signed off rather than whether the human could have disagreed. The endoscopy result is the empirical version of this: clinicians who had used the tool were worse at the unaided task than clinicians who never had.^[3]

The second route runs through the training pipeline. Deskilling assumes a skill existed; the harder problem is a cohort that never acquires it, because the tasks that used to build competence are the first to be automated. Hamsa Bastani and colleagues showed the mechanism in a school setting, where students using an unrestricted chatbot performed better during practice and worse on exams once it was removed.^[10] The equivalent in professional work is a junior analyst, radiologist or developer whose first four years contain no unaided repetitions.

The third route is measurement failure. If practitioners systematically believe they are faster and better when they are slower and worse, organisations will adopt on the basis of perceived rather than measured effect, and the correction will arrive only when an outcome fails publicly. The METR result and the Lee survey point at the same blind spot from opposite directions.^[8][6] Aviation reached the same problem earlier. The Federal Aviation Administration's 2022 advisory circular on flight path management notes that pilots now spend relatively little time manually flying, that degradation of those skills can contribute to accidents, and that airline operating policies should encourage pilots to fly manually, at least periodically for entire departure and arrival phases.^[11][12]

The fourth route is the slowest and least verifiable. Louisa Dahmani and Veronique Bohbot at McGill found that among 50 regular drivers in Montreal, people with greater lifetime GPS use had worse spatial memory when navigating without it, and that among the thirteen who returned for an unplanned follow-up three years later, heavier GPS use went with steeper decline.^[4] Navigation is a narrow capacity and the follow-up group was thirteen people, so the study proves little on its own. It is cited here because it is one of very few attempts to measure what a decade of offloading does, and because the answer it gives is not reassuring.

The single most cited result in this area has been substantially qualified by the people who produced it. METR published an update on 24 February 2026 explaining that it is redesigning the experiment because selection effects had compromised newer data: developers declined to participate because they did not want to work without AI even at fifty dollars an hour, between 30% and 50% avoided submitting particular tasks for the same reason, and time reporting became unreliable once agentic tools let people multitask while waiting. Its late 2025 data pointed toward an 18% speedup for returning participants, with a confidence interval running from a 38% speedup to a 9% slowdown, and METR described its own data as only very weak evidence for the size of any increase.^[13] Anyone citing the 19% figure as a current fact about AI coding tools is citing something its authors treat as historical.

Several of the alarming studies are weaker than their coverage suggests. Michael Gerlich's finding of a negative association between AI use and critical thinking is correlational across 666 valid responses from a UK sample recruited through social media, and cannot separate people who think less because they use AI from people who use AI because they prefer not to think through problems.^[5] The Kosmyna EEG study is an unrefereed preprint, involved 54 participants across the first three sessions and only 18 in the crossover session, and used a single assistant.^[7] The Microsoft and Carnegie Mellon result is self-reported critical thinking, not a measured test of it.^[6]

The strongest substantive counterargument is that the effect depends on design rather than on AI. Bastani and colleagues essentially eliminated the exam penalty by prompting the tutor to supply teacher-designed hints rather than answers, using the same underlying model, though that arm produced no positive learning effect either.^[10] Ethan Mollick of Wharton, writing in May 2026, argues that the problem is using AI as a default or without thinking rather than using it at all. He contrasts the Turkish result with a five-month Python course across ten high schools in Taipei, run by an overlapping team, in which students given an AI-sequenced set of problems scored 0.15 standard deviations higher on a final exam taken without AI, and he cites an Anthropic study in which programmers who asked the model to explain what it was doing, or used it for only part of the work, avoided what his Wharton colleagues call cognitive surrender. His conclusion is that the more promising lever is system-level design constraints rather than user willpower.^[14] Aviation is the precedent for that view: the response to automation dependency was not less automation but changed training and operating policy.^[12]

Usage data also fails to show a one-way slide into delegation. Anthropic's Economic Index reported that the share of directive conversations on Claude.ai, where a user hands over a task with minimal back and forth, rose from 27% in January 2025 to 39% in August 2025, then fell to 32% by November 2025, with collaborative use again more common than full delegation.^[15] If deskilling were being driven by an inexorable rise in handing work over, that series would not have turned.

The main regulatory instrument is Article 14 of the EU AI Act, which requires providers of high-risk systems to build them so that assigned human overseers can understand the system's capacities and limits, monitor its operation, interpret its output, decide not to use it, override or reverse a decision, and stop the system, while remaining aware of automation bias. For remote biometric identification it goes further and requires that an action be separately verified by at least two competent people.^[2] The provision is a statement of what oversight has to mean; whether deployments meet it in practice is the open question, and it is not currently audited against skill retention.

In aviation the response is specific and already published. On 22 November 2022 the FAA issued Advisory Circular 120-123 on flight path management, a response to provisions of a 2020 law passed after the two 737 Max crashes and to a National Transportation Safety Board recommendation following the 2013 Asiana 214 crash at San Francisco. It contains recommendations rather than requirements, and says it will help pilots develop and maintain manual flight operations skills and avoid becoming overly reliant on automation.^[11] The Royal Aeronautical Society read the circular as a shift in position, summarising it as holding that manual flying skills are essential to safety, that automation requires more training rather than less, and that pilots should hone manual skills during ordinary revenue operations.^[12] The pattern there was to preserve the skill deliberately rather than to remove the automation, which is the template most other sectors are being pointed toward.

In medicine the response so far is measurement. The authors of the Polish endoscopy study were explicit that the design was observational, that factors other than AI introduction may have contributed, and that their sample consisted of experienced endoscopists, so they called for work on less experienced clinicians where the effect might differ.^[3] That is the shape of the current medical position: enough evidence to require monitoring of unassisted performance, not enough to justify withdrawing tools that also raise detection when used.

On the research side, the Liu group's stated conclusion is a design requirement rather than a warning: models should be built to scaffold long-term competence alongside immediate task completion, which no current system is optimised for.^[9] METR's public redesign of its own methodology is the other useful development, because the field's headline productivity numbers have been unstable and the reasons are now documented.^[13] Anthropic's Economic Index provides the only continuous public series distinguishing delegation from collaboration.^[15]

Keep assisted performance, retained skill and oversight capacity apart. A tool can raise the first, lower the second and hollow out the third at the same time, and the endoscopy data is the clearest case of exactly that combination: detection improved with AI in use, and unaided detection fell. Claims that AI makes people smarter and claims that it makes them worse are frequently both true of the same deployment, measured at different moments.

Where you have a choice, the evidence favours changing how the system answers rather than abstaining from it. Prompting a model to give hints instead of solutions removed the exam penalty in the Turkish trial without changing the underlying model.^[10] Asking a model to explain what it is doing, or handing it only part of the work, is the pattern Mollick reports as separating programmers who could still account for their own code from those who could not.^[14] Where you supervise a system, the question worth asking about your own oversight is whether you could still do the task yourself well enough to notice an error, because that, and not the sign-off, is what the safeguard consists of.

The indicators that would settle this are professional rather than psychological. Unassisted performance benchmarks in fields that keep them, examination results for cohorts trained with and without these tools, and error rates in processes where humans are supposed to be catching machine mistakes will show the effect if it is real. Endoscopy already publishes such a benchmark, which is why the first hard evidence came from there rather than from a field with no equivalent measure.