Human involvement map · Research and policy analysis
Who gets time to learn when AI changes the job?
Course finance, paid learning time and the opportunity to develop judgment as AI changes work.
Public research edition · 15 November 2026 · Human by choice — a Policy Window advocacy programme
A course place is useful. A chance to take it can be a different matter. If the fee is covered but the workload stays the same, learning may depend on spare evenings, a sympathetic manager or colleagues absorbing the work.
The question is whether people expected to take responsibility for AI-assisted work can still learn to judge it—and whether they can pursue learning beyond their employer's immediate needs. Protecting that opportunity need not mean preserving every junior task or requiring human instruction when AI could teach better.
Consider a hypothetical employee offered a course after AI changes their work. They want to understand the new process and retain options for a different role. A training account might pay the provider. Someone must still decide when they can attend, whether wages continue, how the work is covered and whether the course will develop useful understanding. These decisions can belong to different institutions.
Two arrangements that separate money, choice and time
Denmark's Finanskompetencepuljen, already in the HBC map, is a selected financial-sector arrangement rather than a national entitlement. Its collective agreement provides employer-financed pools with joint employer–union governance. The fund's guidance requires an applicant to discuss a course with their supervisor but says supervisor approval of the application is unnecessary. Fund approval remains separate. The fund does not pay for the learner's time; payment for that time can be agreed with the workplace. This describes the supplementary fund route, not every Danish obligation to provide training. Fund FAQ, collective agreement.
France's personal training account (compte personnel de formation, or CPF) provides a statutory contrast for ordinary use by private-sector employees. Account use needs the holder's express agreement. Eligible learning outside working hours needs no employer agreement and creates no pay entitlement under this route. During-work learning generally requires employer absence authorisation; silence within the prescribed period counts as acceptance, and employer pay continues for training during working time. Account funding remains conditional, including applicable personal contributions; it does not make every course free. Official guidance, holder consent, absence rules, pay rules.
Specified qualifying French-language learning has absence as of right within limits. Career-transition arrangements have separate rules. Ordinary CPF use cannot stand for every learner's position. Code provisions.
These examples reveal different allocations of power. They do not rank the two countries, establish which arrangement people can actually use, or show that either improves skills or earnings. In both, knowing who finances a course leaves questions about time and practice unanswered.
Paid educational leave also predates today's AI debate. The ILO's 1974 Convention No. 140 addresses educational time, financial support and participation, including general, civic and trade-union education. Its text is a useful benchmark, not evidence of a universal individual right or current domestic implementation. Convention text, Articles 1–7 and 13.
Protect the opportunity to learn; test the method
The strongest argument is conditional. If an institution needs people to exercise consequential judgment, it should be able to explain how those people can develop it. If that judgment no longer serves learners or the public, preserving its old training tasks becomes harder to justify. AI-supported practice, simulation or a different allocation of responsibility may serve people better.
Two studies illustrate why immediate output and learning should be assessed separately. Shen and Tamkin's arXiv paper reports a randomized main study analysing 52 programmers learning a Python library. AI access reduced subsequent quiz performance without a statistically significant average time saving. Its short exercise does not establish long-term workplace skill loss. The reported interaction patterns were observed after assignment, not randomized learning methods; they cannot identify the best teaching method. There was no human-tutor comparison. Paper, version 2, Sections 5–7.
Conversely, Kestin and colleagues' published study found higher immediate post-test performance with a carefully designed AI tutor than with classroom active learning in a Harvard physics course. It used a randomized crossover design, assigned through small peer groups, with 194 eligible students across two lessons. It was not a comparison with individual human tutoring, nor a test of durable workplace judgment. AI was embedded in a designed learning environment, not simply offered as an answer generator. Scientific Reports study.
The studies test different populations, tools and comparisons. Their effects cannot be pooled here. They support a research question about learning design, not a general verdict for or against AI. An accessible assessment of understanding may itself use appropriate assistive technology; 'independent judgment' should not mean forcing everyone to work without support.
The next institutional choice to test
This is already an active field. IFOW's eight-case work-design report includes FinanceCo participants proposing meaningful learning, mentoring and honouring learning time reportedly in their contracts. Those workshop reports are not authenticated contractual entitlements or evidence of long-term effects. HBC's contribution should build on that work by asking what makes a learning opportunity claimable when interests conflict. IFOW report.
One practical candidate is a negotiated process for requesting paid learning and practice: the employee proposes a relevant learning goal; the employer and worker representatives agree resources, scheduling and workload cover; a refusal comes with reasons, a feasible alternative and a route to reconsideration. Course finance, time and practice support must each have an identified decision-maker. The fund board cannot be assumed to control employers' wage or scheduling obligations. This is a proposed design to examine, not a description of either country's existing complete system.
First compare the same learner group and goal, with comparable access to AI-supported learning, under two allocation rules: an explicit paid-time, cover and reconsideration process, and discretionary scheduling. Specify when learning happens, who pays and what work others absorb under each. Count learner time, mentoring, cover and administration against a stated resource ceiling; unknown costs prevent a claim of equal spending. Flexible AI tutoring with existing scheduling is also a serious whole-package alternative. That comparison would not isolate the effect of the time rule or the tutor. Examine offers, take-up, refusals, displaced work and useful capability after support ends, including people who decline or cannot attend. An exploratory comparison could establish feasibility; a credible effect estimate would need a separately justified design and adequate participation. No such trial has been conducted by HBC.
Employers would bear costs of paid time and cover; training providers and mentors would face capacity demands. Service recipients could also bear the cost of reduced staffing. Workers may prefer income, shorter hours or a different route. A compulsory course could reduce their control. A scheme limited to existing employees would also exclude many unemployed people, carers and prospective entrants. Those are reasons to compare distribution and alternatives before proposing a general entitlement.
Before treating a course offer as a usable opportunity, ask who can choose it, who can release paid time, and who covers the work. The next research task is to compare those decisions and their costs, including the French arrangement. Current evidence supports that research; it does not establish a best system, a necessary number of protected hours, an effective national policy or a right to keep a job human.
Sources checked 15 November 2026. Selected examples and two studies, not an exhaustive review or individual advice. The French guide flags a pending financing proposal. Implementation and outcomes remain unassessed.