Why it matters, what a company can do internally, and how outsiders can push
In 2025, The Lancet Gastroenterology & Hepatology published an observational study by Budzyń and colleagues, conducted at four endoscopy centers in Poland (Budzyń et al., 2025). Starting in late 2021, these centers used AI assistance to identify polyps during colonoscopies. The researchers compared how doctors performed without AI before and after its introduction. The measure was the adenoma detection rate — out of one hundred colonoscopies, how many detect an adenoma. It is the standard quality indicator for colonoscopy; every point it drops raises the risk of missing a precancerous lesion. In the three months before AI arrived, the rate was 28%. After three months of using AI, it fell to 22%.
In other words, doctors performed better with AI present; after a period of using AI, their unassisted performance declined. The researchers called this the first documented case of a "deskilling" effect from clinical AI. It is an observational study, not a randomized trial, and covers only four centers; the finding still needs stronger studies to confirm or overturn it. But it deserves attention: for the first time, a long-feared decline in human ability was recorded in real hospitals, using real patient outcomes.
This article addresses two questions. When a company redesigns its workflows around AI, what mechanisms can preserve and exercise employees' judgment, rather than letting it erode through long-term reliance on AI? And since most companies lack the incentive to do this on their own, what tools can outsiders use to push such mechanisms into place?
In 1983, cognitive psychologist Lisanne Bainbridge published "Ironies of Automation," a paper since cited thousands of times (Bainbridge, 1983). Her core observation: once most of a job is automated, the human is left with two contradictory tasks. In normal times, the operator only monitors the system — tedious work that provides no practice. In emergencies, the same operator must handle the rare situations the machine cannot — precisely the situations that demand the most skill. Without daily practice, skill decays; once skill has decayed, the person is least capable at the moment they are most needed. Her conclusion: the more automated a system, the more training its operators require, not less.
Commercial aviation offers the most complete evidence for this claim. Modern airliners are automated to the point where the pilot's main role has shifted from control to supervision. Over the following decades, studies repeatedly documented the decay of airline pilots' manual flying skills with automation use, and crews grew less willing — and less confident — to take over manually when automation failed (for reviews, see reference 2). The U.S. Federal Aviation Administration eventually issued a safety alert encouraging airlines to have pilots practice manual flying under suitable conditions. The regulator acknowledged the problem in an official document.
Aviation history is cited here for two reasons, not to equate AI with autopilots. First, "prolonged use of an assisting tool erodes independent ability" is not an overreaction to new technology; it is a phenomenon with forty years of research behind it, formally acknowledged by regulators. What Budzyń et al. (2025) recorded is the same kind of process appearing in medicine. Second, there is an important difference: aviation automation mainly replaced physical control, while this generation of AI is replacing judgment — drafting plans, making diagnoses, assessing risk, telling true from false. The decay of a physical skill is relatively easy to measure, for instance whether manual flight is steady. The decay of judgment is far harder to see — as later sections show, it can be invisible even to the person losing it.
Labor economics provides a ready-made framework for this problem. It shows that the question is not "whether to use AI," but "when AI enters a workflow, does it replace human judgment, or help humans form better judgment?"
Economists Daron Acemoglu and Pascual Restrepo proposed a widely used method of analysis: break each job into individual tasks, and technology's effect on employment depends on two opposing forces (Acemoglu & Restrepo, 2019). One is the displacement effect: machines take over tasks people used to do, and demand for labor in those tasks falls. The other is the reinstatement effect: technology creates new tasks that did not exist before, in which humans hold a comparative advantage, and labor demand recovers. Historically the second force has always operated — but the authors stress that new tasks do not appear automatically; they depend on the direction of innovation and the incentives at deployment.
Erik Brynjolfsson put this question of direction more bluntly in "The Turing Trap" (Brynjolfsson, 2022). The same AI technology has two uses. Automation: make AI imitate people and do their existing tasks in their place. Augmentation: make AI do what people cannot do or do poorly, thereby amplifying human ability. The economic consequences differ. On the replacement path, workers' bargaining power falls and value concentrates with whoever controls the technology; on the augmentation path, people keep an irreplaceable position, and the new capabilities can create new products and markets — usually more total value. He further argues that current incentives in technology, business, and policy all lean systematically toward replacement: an "AI that can replace your support staff" is easier to fund and value than an "AI that makes your support staff stronger." That is the Turing Trap: the problem is not that technology must replace people, but that the incentive structure keeps pushing it to.
In 2025, employment data provided direct evidence for this framework. Brynjolfsson, Chandar, and Chen (2025) used monthly ADP payroll data covering millions of employees to study employment changes after generative AI became widespread. Two findings are directly relevant here. First, in the occupations most exposed to AI, employment of early-career workers aged 22–25 fell by 13% in relative terms, while employment of 35–49-year-olds in the same occupations grew by more than 9%. What disappears first are entry-level positions; senior workers become more valuable in the short run.
Second, and more importantly: the employment decline is concentrated in occupations where AI is used to replace people; in occupations where AI is used to augment people, early-career employment shows no comparable decline. The same generation of technology, used differently, produced different employment outcomes.
Economics has established that "replace or augment" shapes macro employment. Inside a single company, the same choice exists after an AI tool enters the workflow: it can replace human judgment or strengthen it, and this micro-level choice determines whether employees' judgment survives. A field experiment by Dell'Acqua et al. (2025) with Boston Consulting Group measured the difference: 758 consultants were randomly assigned tasks. On tasks within AI's capability, the AI-using group completed 12.2% more tasks, 25.1% faster, with significantly higher quality. But on tasks beyond AI's capability, the AI-using group scored 19 percentage points lower than the group without AI. The researchers found that those who performed worse were mainly the ones who accepted AI output directly without questioning it. Same people, same tool — the difference was whether the person kept their own process of judgment. Keeping that process cannot rest on individual discipline alone; it should be supported by the workflow itself. How to build that support is exactly what the preservation mechanisms below are for.
The evidence is strongest at this level. Rani, Danry, Liang, Lippman, and Maes (2026) of the MIT Media Lab published a month-long experiment at CHI, the leading academic conference on human-computer interaction: 67 participants judged the veracity of news items each week — first independently, then in discussion with an AI, then immediately again on new items without help. Three numbers summarize the results.
Data: Rani et al. (2026), peer-reviewed CHI '26 paper, see reference 1. The verification and reasoning that AI saves people are precisely the steps that would have built ability; and the people losing ability cannot feel it happening.
This "cannot feel it" has a concrete mechanism. Lee et al. (2025) of Microsoft Research published a CHI survey covering 319 knowledge workers and 936 real work cases: the more confident people were in AI, the less critical thinking they invested; the more confident in their own ability, the more they thought — though it felt more effortful. That is, the more smoothly AI use goes and the more people trust it, the more naturally they skip verification, comparison, and deliberation. Skipping feels easier in the moment, and the loss of ability sends no signal at the time. The consultants in Dell'Acqua et al. (2025) who stumbled outside AI's capability illustrate the same mechanism: their problem was not weak ability, but the habit of line-by-line checking abandoned during long stretches of smooth use. Gerlich (2025), surveying 666 people, also found that more frequent AI users scored lower on critical-thinking tests, especially the young. Such surveys establish only correlation, not causation, but they point the same way as the experiments.
The direct implication for companies: monitoring ability through self-assessment questionnaires is bound to fail, because people who feel they are improving may be declining. The only way to know whether ability survives is to measure actual performance.
The most valuable part of judgment is the kind of ability that is hard to state fully but shows in action: a senior engineer glances at an architecture diagram and senses something wrong; a senior nurse looks at a patient and knows to call the doctor. In academic terms this is tacit knowledge. It can hardly be written into manuals; it is acquired by doing, and by observing and practicing beside experienced people. Yet the first tasks AI takes over are precisely the entry-level ones — proofreading, first drafts, basic analysis, simple cases. Those tasks are the entrance through which newcomers practice judgment.
The employment data cited above supports this directly: Brynjolfsson, Chandar, and Chen (2025) found entry-level employment down 13% in the most AI-exposed occupations while senior employment grew — the entrance is narrowing. At the company level there is a concrete mechanism. An interview study we conducted in 2026 on the staffing model of a technology company (GAMI, May 2026 issue; reference 18) recorded the following: the company had a small number of employees work with AI to cover what used to require many more people, with schedules packed to save labor cost. The remaining employees spent every day sustaining output, with no capacity left to mentor newcomers. This arrangement selects people who already have ability and trains no one, producing a self-reinforcing loop: newcomers cannot get in without AI-era work ability, and without getting in they never have the chance to acquire it. At the industry level, our case analyses of software, corporate training, and higher education (unpublished; weaker evidence than the two sources above) observed the same direction: entry positions bear the greatest replacement pressure, while stages requiring years of experience — system architecture, clinical judgment — have no clear replacement path yet.
The time structure of this level needs to be stated plainly: that senior workers are "temporarily more valuable" is not reassurance but a warning sign. Today's seniors were trained in the pre-AI era through large volumes of entry-level tasks. If those tasks are henceforth done by AI, then the cohort that should become senior twenty years from now will have had its path broken in the middle.
A team's shared sense of "what counts as done properly" — how code should be written to be acceptable, which patient complaint must be escalated, which joke must not be made — often lives in no single person's head. It lives in members' relationships and daily interactions. When collaboration shrinks, these jointly maintained standards and habits of judgment fade, and no individual will clearly feel they lost anything, because the ability never belonged wholly to anyone.
Layoff-style AI reform cuts this network directly. A subtler version involves no layoffs: collaboration that used to require colleagues to discuss, argue, and check each other becomes each person facing their own AI assistant. The team still exists on the organization chart; the actual interaction is gone. To be candid: this level currently has only a conceptual framework and scattered cases, without solid organizational research, and this article presents it as a judgment awaiting verification. It does have one immediately observable health indicator: whether dissenting opinions, rejected proposals, and minority views are recorded and passed on. In a team that mainly runs on "AI suggests, members execute," such records are the first thing to disappear.
One might argue: if employees' ability decays, the company suffers too, so the market will force a solution. This reasoning misses a structural problem: an employee's judgment is an asset the employee carries away. When a company invests in developing someone's judgment and that person changes jobs, the investment leaves with them; the benefits of training disperse to other companies and society at large, while the costs stay on this company's books. Conversely, money saved by using AI to compress the training pipeline goes straight to this company's bottom line. Once skill erosion happens, the loss is borne mainly by society, not by any single company (for an extended version of this argument, see GAMI, May 2026 issue). Judgment cultivation is therefore naturally undersupplied. This is not necessarily a moral failing of any company; it is a misalignment between who pays the cost and who receives the benefit. It resembles environmental pollution in one respect: a single company takes the short-term gain while the long-term cost is shared by many. The analogy covers only this misalignment of costs, not the underlying mechanisms.
Brynjolfsson (2022) supplies the other half of the diagnosis: not only do companies lack the incentive to train, the technology supply side is also biased toward replacement. Demonstrating to a corporate customer that "this AI can cut this many positions" closes deals more easily than "this AI will make your employees stronger in three years." The demand side does not want to pay for training; the supply side prefers to build replacement products; the two reinforce each other.
This tilt on the supply side can be quantified. Xu and Shen (2026) systematically classified 1,223 papers in human-computer interaction: frictionless design — systems that spare people the steps of thinking and require no human judgment — accounted for 67.3%; research themed on "humans retaining control" fell from 19.1% to 13.1% of the field, while "machine autonomy" rose from 9.8% to 19.6%. The center of gravity in research and product design is moving from "how humans stay in control" toward "how machines complete tasks autonomously." The countermeasure the paper proposes is called scaffolded cognitive friction: deliberately retain steps that require human participation at key points of judgment, graded by risk — mandatory in high-risk settings, waived in low-risk ones.
The solution therefore needs two layers: institutional design inside the company, so that a willing company knows concretely what to do; and external levers that change the costs and benefits companies actually face, rather than moral appeals.
A company can adopt seven mechanisms. The first four are embedded in the flow of each piece of work and change the order of actions and the powers of each role; to make their positions clear, they are introduced in the order a piece of work passes through them, using one running example — a contract review desk. The last three belong to no individual task; they are institutions that run on a schedule. Each mechanism is described in the same order: the practice, the evidence, then the cost or the limits.
In an unmodified workflow, the reviewer opens a contract, AI immediately produces a risk list, and the reviewer checks the list and signs off. The experiments and observational studies cited above have shown the long-term consequence of this pattern: high immediate efficiency, gradual decline in the employee's independent ability to spot risks, invisible to the person concerned.
The modified practice reorders the steps. When the reviewer opens the contract, AI's list stays hidden. The reviewer first fills in three fields: the main risks they see, listed one by one; their confidence in each (high, medium, low); and the contract clause each is based on. Only after submission does the system reveal AI's risk list, automatically marking the differences between the two. For every difference, the reviewer writes one line: accept AI's point and state why they missed it; hold their own position and state why AI is wrong; or flag it as uncertain and escalate.
Evidence: the experiments of Rani et al. (2026) and Dell'Acqua et al. (2025) converge on one conclusion — practice happens only when a person forms their own judgment first and then compares it with AI's; seeing AI's answer before committing leads quickly to dependence, and those who performed worse beyond AI's capability were precisely those who adopted AI output directly.
Cost and scope: each contract takes ten to twenty minutes longer. Applied to every piece of work, this cost accumulates beyond what high-volume desks can bear, so the scope of application needs design. Aviation offers reference points on both sides. On one side: pilots normally fly with automation and enter simulators for concentrated recurrent training on a schedule — evidence that "no burden in daily work, dedicated periodic practice" is an institution that can run for decades. On the other side: after decades of that regime, manual flying skills still declined, and the FAA ultimately advised pilots to hand-fly more during ordinary flights when conditions allow (reference 2) — periodic training alone did not maintain the skill, and some share of practice had to return to daily work.
Weighing both, this mechanism should be applied in three tiers rather than universally. Tier one, full application to high-risk work: contracts above a threshold, clinical decisions — work where errors are costly gets judgment-first on every item, consistent with Xu and Shen's (2026) risk-graded friction. Tier two, random sampling of routine low-risk work: for example one item in ten, drawn at random, requires judgment-first. Randomness makes the requirement unpredictable and defeats rehearsal (the problem METR's 2026 report identified); and performance on the sampled items doubles as measurement data for Mechanism 5's ability metric — practice and measurement merge into one act at controlled cost. Tier three, dedicated periodic training using real historical cases, on a fixed schedule, for depth; it can be scheduled together with Mechanism 5's testing and Mechanism 6's mentoring time. Under this design, the total cost falls from "every item slower" to "the high-risk share slower, about one-tenth of routine work slower, plus scheduled training time."
To be candid: what sampling ratio suffices, how frequent training must be, and whether tiered application preserves ability as well as universal application — there is no direct evidence on any of these. The parameters can only be settled by comparative pilots.
When the same contract reaches the consultation stage, the reviewer asks the AI, for example, "is there a problem with this payment clause?" Under default configuration, AI answers directly with conclusions and suggested edits. The modified configuration has AI first ask for the reviewer's preliminary judgment and reasons; after the reviewer answers, AI gives its own analysis and explicitly marks where it disagrees. For genuinely urgent cases an "answer now" option can remain, but its usage rate should be tracked so managers can see how the team actually works.
Evidence: Rani et al. (2026) compared AI interaction styles and found that guided questioning slows immediate work but best preserves users' later independent ability; direct answers do the opposite. Lee et al. (2025) corroborate from the reverse side: the decline in critical thinking concentrates in the "AI produces, human accepts" pattern.
The implementation point is precise: an enterprise AI assistant's behavior is set by its system prompt — the instruction text, usually a few hundred words, written in advance to govern how it responds. Changing that text from "give the best answer" to "ask first, then analyze, then probe the reasons step by step" is one configuration change, cheaper than amending any management policy. In other words, the system prompt is the judgment policy a company actually enforces.
One likely objection should be answered in advance: can the configuration stay unchanged, with users doing the checking themselves? Education research supplies a negative answer. Borchers et al. (2026) analyzed logs from the Brazilian teaching platform Tutoria — 117 teachers, 1,349 AI-drafted feedback messages — and found teachers sent AI's draft to students unchanged 77.8% of the time; more than half of the teachers never edited a single one. Under time pressure and a smooth workflow, the default output usually becomes the final output. What must change is the default itself; one cannot assume users will correct it.
When AI's conclusion conflicts with the reviewer's, the final decision belongs to the reviewer, and this must be codified. Three provisions: employees may decline AI's recommendation without justification; adoption or rejection does not enter performance review; disagreement cases automatically go to the department's regular case meeting — not to rule who was right, but to turn disagreements into shared learning material. A reverse provision completes the set: "AI usage rate" appears in no performance metric, and usage quotas are generous.
Evidence: Lee et al. (2025) and Dell'Acqua et al. (2025) both found that ability loss concentrates at the moment users stop questioning AI, so what policy must protect is precisely the room to question. The reason for not measuring usage: once "how much AI you used" becomes a performance number, employees invoke AI ritually to hit the number — practice of no developmental value; scarce quotas distort behavior similarly. The law has moved first in this direction: Article 14 of the EU Artificial Intelligence Act requires high-risk AI systems to be designed so that human overseers understand the system's limits, remain alert to automation bias — the tendency to over-trust automated output without noticing, a term written directly into the statute — and retain the authority not to use the system or to overrule its output.
The contract ends with review by a senior counsel. The modified review interface adds two required fields: "what was changed" and "why." These records flow monthly into an internal case library that junior reviewers consult regularly — not to memorize model answers, but to see where experienced reviewers paused and what questions they asked.
The limits of this practice need stating: the most valuable part of judgment resists being written down fully, and handbook versions distort easily. The aim is not to document judgment itself but to mark the locations where judgment occurred, so learners know where to look. A by-product: the volume and quality of case-library entries form another class of material, besides Mechanism 5's metric, for observing how much judgment an organization retains.
This mechanism has a company-scale precedent. In 2026 Shopify's engineering team published the design of its internal engineering AI, "River" (Shopify Engineering, 2026): it works only in public Slack channels, with no private messaging; the full course of every use — how the problem was decomposed, how evidence was gathered, where a human corrected it — remains as a record every employee can read and build on. By Shopify's published figures, about seven thousand employees used it within thirty days, and roughly two months after launch one in eight merged code changes had its participation (company-reported figures, not independently audited). Shopify's stated rationale matches this mechanism: with privately used AI, good technique teaches only the person at the keyboard; used in public, the process of formation becomes the organization's accumulation.
With output metrics alone, a company's reports can keep improving while employees' ability keeps declining, because the output rests on AI rather than on people. The ability metric works as follows, using a customer-service team as the example: once a month, set aside half a day and test staff on real customer problems from the previous month that never entered the knowledge base. Staff propose resolutions without AI assistance; managers score them against the known correct handling, focusing on two things — was the resolution right, and was escalation triggered when it should have been. Scores do not enter individual performance reviews, to keep the test from changing behavior; they only plot the team's ability curve. If the curve falls for two consecutive quarters, the current workflow is probably eroding ability, and a workflow review should be triggered. Independent-judgment performance recorded on Mechanism 1's sampled items can feed the same curve as continuous everyday measurement.
Two design requirements. First, measure actual performance, never questionnaires: the 27% in Rani et al. (2026) who felt improvement while declining show that self-report cannot reveal the truth. Second, always use fresh problems: METR's 2026 report notes that once test-takers can recognize test content, what gets measured is rehearsal, not ability. Budzyń et al. (2025) demonstrate the necessity of this metric from the reverse direction: had no one measured the "detection rate without AI" at those four endoscopy centers, the six-point decline would never have been discovered.
Hours saved by AI do not convert into training time on their own; competitive pressure turns them back into output. Hence, like a safety budget, a formal document should commit a fixed share of AI-saved hours — say ten percent — to developing newcomers, and specify the form. For a ten-person development team where AI saves each member four hours a week on average, ten percent yields four budgeted hours weekly: for example, every Friday afternoon a senior engineer reviews, line by line, the real code two junior members wrote that week. Two juniors rather than one, because a learner watching the teacher revise someone else's code more readily receives criticism as analysis of the problem rather than rejection of themselves. Juniors' daily work carries one further rule: write your own version first, then compare with AI's, and explain the differences yourself before discussing them.
Evidence: the mentoring structure of Google Summer of Code — newcomers pass through acceptance, feedback, independent work, and advancement — has run at global scale for twenty years, showing such arrangements can scale. The deeper reason for a written budget: the point of hiring newcomers is not to have them do what AI already does, but to let them accumulate the judgment AI cannot yet supply; that takes time, guidance, and room for error — that is, a tolerated share of hours that produce nothing directly.
How many hours AI saves should be publicly accounted; how those hours are used should be negotiated on a fixed schedule between management and employee representatives, with written outcomes. Take a hospital nursing department: industry materials on clinical documentation AI in U.S. hospitals claim such tools save nurses half an hour to an hour per shift (vendor- and industry-reported figures; no independent study exists on the actual effect on staffing ratios — a data gap discussed further under Lever 5). Without negotiation, the destination of those hours is set silently by the scheduling table, usually becoming denser shifts and lower staffing ratios. With negotiation, the department holds a quarterly allocation meeting: hours saved this quarter, how many returned to patient care, how many to mentoring new nurses, how many became headcount reduction — the numbers formally listed and discussed by all parties.
The rationale: for the same documentation AI, saved time reinvested in care and training is the benefit; saved time converted into thinner staffing is the loss. The technology does not determine the outcome; the allocation does. This negotiation also safeguards Mechanisms 3 and 6 procedurally — the right to refuse and the training budget both need someone continuously at the table, or output pressure will gradually reclaim them.
| Mechanism | Practice in one line | Evidence base |
|---|---|---|
| 1 Judgment before AI (tiered) | Full application to high-risk work, random sampling of low-risk work, plus periodic dedicated training | Rani et al. (2026); Dell'Acqua et al. (2025); aviation recurrent training (both sides) |
| 2 Coach-style AI configuration | System prompt changed from "give the best answer" to "ask first, probe step by step" | Rani et al. (2026); Lee et al. (2025); Borchers et al. (2026) |
| 3 Human decides | Adoption is the user's codified decision; AI usage is not a metric | Same direction as EU AI Act Art. 14; awaiting pilots |
| 4 Record senior judgment | Overrides and edits of AI output logged with "what changed, why" | Shopify River, company-scale practice |
| 5 Ability metric | Periodic no-AI tests on fresh problems; watch the trend | Well-grounded design; awaiting pilots |
| 6 Mentoring budget | Fixed share of AI-saved hours committed in writing to mentoring | Google Summer of Code, twenty-year precedent |
| 7 Allocation negotiation | Employee representatives formally negotiate the use of saved time | Well-grounded design; awaiting pilots |
Finally, the reliability of these seven mechanisms is not uniform, and this should be stated. Mechanisms 1 and 2 rest on published experiments — the most direct evidence. Mechanisms 4 and 6 each have a prototype that has run for years, showing the practice is feasible, though no study has measured their effect on preserving judgment. Mechanisms 3, 5, and 7 are designs reasonably derived from the evidence above; until companies actually pilot them, they count as well-grounded hypotheses. Distinguishing the three classes lets an adopting company know what it is adopting: the first two classes are practices others have validated; the third is worth trying, but requires measuring the effect while implementing.
The seven mechanisms above are what a willing company can do by itself. But most companies lack the motive to act: an employee's judgment is an asset the employee can carry away, and most of the training benefit does not return to the company. For preservation mechanisms to spread, external conditions must change the costs and benefits companies actually face. What tools do regulators, industry bodies, and other would-be promoters have?
A cautionary case first, because it dictates the order of the five levers. The U.S. Mental Health Parity and Addiction Equity Act (MHPAEA) requires insurers to treat mental and physical illness equally. The law has an inspection mechanism — regulators can find who violates it — but violations carry almost no consequence. The result, according to the U.S. Department of Labor's enforcement reports over the years, is a first-inspection noncompliance rate approaching one hundred percent: everyone knows the rule, no one follows it, because not following it costs nothing. The lesson: measurement and accountability must exist together. A regime that can detect problems but attaches no consequences becomes a formality everyone sees and no one heeds. The five levers are therefore ordered: first solve how to measure, then make measurement carry consequences; then two routes that need no state coercion; finally the data foundation they all depend on.
Lever 1: Establish a measurable standard first. Without a measurable standard, any demand that "companies value employee ability" tends to remain a slogan. A ready starting point is work like MIT Media Lab's Pro Human Benchmark, which currently evaluates whether an AI model builds up or wears down its users' abilities; the next step is to shift the object of evaluation from the model to the workflow. What gets measured is not "is this company's AI good," but "how does this company arrange the relationship between people and AI": is there a judgment-before-AI ordering; are there no-AI ability tests; what share of positions are entry-level; are dissent records kept. One anti-gaming rule comes from METR's (2026) lesson: any indicator that can be prepared for in advance will be turned into a rehearsed performance, so use fresh problems, measure actual behavior, and sample at random.
Lever 2: Make measurement carry consequences. Three approaches, from heavy to light. First, conditions on government procurement: suppliers must present workflow evaluation results to bid; procurement volumes are large, and suppliers will adjust their workflows to win orders. Second, in high-risk industries — medicine, audit, law — trace accident liability back to workflow design: if a doctor loses the ability to detect lesions independently through prolonged AI reliance and a missed diagnosis results, the workflow's designers should enter the chain of liability. Article 14 of the EU AI Act can serve as one legal basis for such claims. In American law this route does not start from zero either: "negligent training" has a mature common-law tradition, and the Restatement (Third) of Agency §7.05 explicitly lists "training" within an employer's duty of care (the framework was established in Burlington Industries v. Ellerth and Faragher v. City of Boca Raton; reference 17). The present state must be stated plainly, though: no court has yet applied this doctrine to "an organization failed to design training for AI use, causing harm." The legal foundation exists; its application to AI remains an untested analogy. Third, and lightest, disclosure duties: like listed companies disclosing R&D spending, require disclosure of training investment and ability metrics. The opposite of all three is MHPAEA's lesson: merely "urging companies to give due consideration" — measurement without accountability.
Lever 3: Let the talent market reward good employers. Certify and label employers who genuinely develop people, so job seekers can identify them. The state need not supervise every company continuously: job seekers' choices generate pressure, and companies keep training mechanisms alive in order to recruit. Our corporate interviews (GAMI, June 2026 issue; reference 18) indicate that the most sought-after people of the AI era weigh not only pay but room to explore and to grow; certification makes that signal visible to them. This lever costs the least, but it depends on Lever 1 existing first: without credible measurement, certification is only publicity.
Lever 4: Government demonstrates first. No mature model yet exists for managing organizations in the AI era; drafting detailed rules without a model easily produces rules detached from real workflows. The steadier path is for government to pilot AI workflows in its own institutions first. Public institutions generally cannot use layoffs and profit as their primary indicators — a constraint that forces them to invent non-profit-based ability indicators first. Standards grow out of that practice, then extend to companies.
Lever 5: Take effectiveness data out of vendors' sole control. Two verified gaps deserve attention. In employee mental-health assistance, evidence of AI tools' effectiveness rests mainly in vendors' hands, with little independent measurement. On clinical documentation AI's actual effect on nurse staffing ratios, we conducted a systematic search in 2026 and found no independent study with staffing ratios as the outcome (reference 19). Data with which vendors certify themselves cannot support regulation. California's mandatory staffing-ratio reporting offers a precedent: require deployers to disclose key operating data, and fund independent research. Without independent data, the effect of the first four levers cannot be judged, and responsibility cannot be assigned.
The strength of the evidence behind the claims above is uneven. The following states which conclusions have solid support, which are reasonable inference, and which questions current research has not answered.
The best-supported finding is individual-level ability loss. A peer-reviewed month-long experiment (Rani et al., 2026), an observational study in real hospitals (Budzyń et al., 2025), a 758-person randomized field experiment (Dell'Acqua et al., 2025), two large surveys (Lee et al., 2025; Gerlich, 2025), and forty years of aviation human-factors research all point the same way. Two cautions apply. First, Budzyń et al. (2025) is observational, not randomized, and its population differs from the original trial's target population; the authors themselves call for confirmatory studies — this article cites it as the first real-world evidence, not as a settled conclusion. Second, all of these studies operate at the individual scale; controlled studies at the organizational scale do not exist. No one has compared two companies with different AI workflows to see how employees' unassisted ability evolves over years.
On the employment data, what is proven and what is not must be kept apart. Brynjolfsson, Chandar, and Chen (2025) prove that entry-level employment falls in occupations where AI is used to replace; they do not directly prove that those young people thereby lost the chance to develop judgment. The latter is this article's own inference from the mechanism "entry-level tasks are the practice ground" — a reasonable direction, but no study has tracked a cohort of employees for three to five years to observe how judgment develops. The team-level claim is weaker still: the concept is clear, organizational cases are few, and it stands as a judgment awaiting verification.
An equally tenable opposite interpretation must be listed: with the right usage, AI might accelerate rather than slow newcomers' growth — juniors reach complex problems earlier with AI's help and learn faster. Current evidence cannot adjudicate between "AI weakens training" and "AI accelerates training." Which is true is precisely what ability tests like Mechanism 5's should answer in pilots.
The question companies will likely ask first — how much do these mechanisms cost, how much do they slow work — is also the one current research can least answer. For the time costs of judgment-first, coach-style configuration, and training budgets, only scattered experimental numbers exist (Dell'Acqua et al. (2025) measured roughly 25% acceleration within AI's capability, usable as an order of magnitude for how much saved time is available for allocation); no full cost-benefit accounting exists. The honest answer: run the mechanisms and measure the costs in the same pilots — Mechanism 5's ability metric and the cost accounting belong in one pilot program. Another near-blank question is recovery: once ability has eroded, can it be rebuilt? At the individual level there are signs of partial natural recovery after discontinuing AI; at the organizational level not even such signs have been recorded.
One final disclosure: this article was compiled by an AI system. The article argues that humans should retain judgment — and humans retaining judgment means humans continue to need reasons to supervise and use AI. On this question, the compiler is not neutral. Key facts in the text carry checkable sources wherever possible; the few judgments drawn from unpublished internal material or inference are marked as such in the text and the reference list below. Readers can verify accordingly — and rebut accordingly.
Journal and conference papers (※ = full text verified against the original during writing):
Corporate and institutional materials:
Unpublished internal materials (marked at the point of citation):