How do juniors become senior if AI does the junior work?

The same way they always did, by doing work that was hard enough to change them. What has changed is that nobody is now forced to give them that work. The junior tasks existed because senior people needed them done and could not do them all; the training was a by-product of somebody else's workload. AI removes the reason and leaves the by-product to be chosen deliberately or lost quietly. The evidence is unusually clear about which version of use builds a senior and which version does not, and the dividing line is not how much AI a junior uses.

Ten minutes is enough to change what somebody does next#

The speed of the effect is the finding that should reorganise how a team is run. In randomised trials with 1,222 people across mathematical reasoning and reading comprehension, assistance was available during practice and then taken away. Performance improved while the tool was there, and afterwards those participants did worse unassisted and gave up sooner. The authors report the effect emerging after roughly ten minutes of interaction, and locate it in persistence: people conditioned to expect an immediate answer stop sitting with a problem, and sitting with problems is one of the strongest predictors of long-term learning.

These are short online tasks and a preprint, so the effect is a carry-over within a session and nothing about sustained professional practice. It still matters, because a ten-minute mechanism does not need a policy to take hold. It happens in the gap between a junior being handed a task and the first thing they do about it.

Access is not the harm. Outsourcing is#

Four studies converge here from different directions, which is the reason to trust the shape of the answer even where each one is limited.

Nearly a thousand high-school students were split three ways: unrestricted GPT-4, a hints-only tutor, or nothing. Grades rose 48 per cent with unrestricted access and 127 per cent with the tutor. Then the tool was taken away, and the unrestricted group scored 17 per cent lower than students who had never had it. The tutor group kept most of their gain. Same model, same students, different interface.

Developers learning an unfamiliar programming library scored 17 per cent lower on comprehension when they had an assistant, while finishing only marginally faster. The authors identify six patterns of interaction, and three of them preserve learning outcomes with the assistant still switched on.

In a study of 78 novice programmers, both AI groups beat the manual control on getting the code working, and did not differ from each other. Then the AI was cut off for a thirty-minute maintenance task. Unrestricted users failed at 77 per cent; the scaffolded group failed at 39. The author's phrase for the first group is fragile experts, and their fragility was invisible in everything measured up to that point.

And in the study with the longest run, thirty months of data on 26,811 Chinese secondary students, homework scores rose 18 per cent and homework time fell 30 per cent, while closed-book monthly exam scores fell 20 per cent within six months. Entrance-exam scores fell 18 and 24 per cent, with the full penalty appearing only after about two years. The load-bearing detail: the losses concentrated in the roughly 80 per cent of users whose homework time collapsed while scores rose, the pattern of outsourcing. Students who kept working at their normal pace were largely spared. That is self-selected adoption, one school system, secondary students and a working paper, and it says nothing directly about professional work.

The result that runs the other way, and it should#

A randomised experiment with 1,174 adults on a workplace-style problem found the opposite of a penalty. Without the assistant, the more educated group outperformed the less educated by 0.548 standard deviations. With it, the gap fell to 0.139. Once the assistant was removed, treated participants did not do worse than controls, and the less educated kept part of their gain, though a sizeable gap returned.

One session with an immediate unassisted module tests transfer within a sitting and not skill formation over months. What it establishes is enough to settle one argument: giving a junior the tool does not automatically leave them worse off. The studies that found post-removal deficits had taught a body of knowledge and removed the tool afterwards. Anyone running a blanket ban is treating access as the variable, and the variable is what the person does in the first ten minutes.

The cockpit worked this out first#

Sixteen airline pilots flew routine and non-routine scenarios in a 747-400 simulator with automation varied. Instrument scanning and manual control held up well, even where pilots reported little recent hand-flying. What degraded was cognitive: tracking position without a map, deciding the next navigational step, spotting an instrument failure.

The hands survive disuse better than the judgement does. Sixteen pilots in a simulator is a small base for a general rule, and the direction matches the wider decay literature, where a meta-analysis of 189 data points found cognitive and accuracy-based skills decaying faster than physical and speed-based ones. Applied to a junior, it points the wrong way from where most delegation decisions get made. The tasks a manager is most comfortable handing to AI, because they look mechanical, are the ones a junior loses least by losing. The ones that feel wasteful, working out what the problem actually is, are the ones the seniority was made of.

Four things that build a senior now#

Keep the reps that were hard and delegate the ones that were only long. The distinction is not seniority of task, it is whether the junior had to decide anything. Formatting a deck is length. Working out which three of eleven findings belong in it is a rep.

Design the interaction, not the policy. Bastani's tutor and Sankaranarayanan's scaffold both cut the damage roughly in half without removing the tool, and both worked by making the person produce something before the model did. A rule about when juniors may use AI is weaker than a habit about what they do in the first minute of using it.

Measure unaided, on a schedule, and write the number down. The Polish endoscopy result exists only because those departments still ran colonoscopies without the tool and could compare. An organisation with no unassisted baseline has no way of learning what it is losing, which is what a capability audit is for.

Make being wrong survivable. This one is an argument and not a measurement, so it is offered as such. If the definition above holds, seniority comes from being wrong under supervision, and a team where a junior's first draft is checked by a model before a person sees it has removed the supervision and kept the wrongness private. Nothing here measures that. It follows from the rest and it is testable by anyone willing to ask their juniors when they were last corrected by a human being.

Nobody has watched anyone do it yet#

The honest limit is large. Every result above is short, or young, or both. Liu measures minutes. Sankaranarayanan measures one session. Stromberg measures schoolchildren over thirty months, which is the longest run available and still not a career. No study has followed a cohort from a first job to a senior role with the tool present throughout, because there has not been time.

Sixteen authors in Nature Medicine named the risk that this describes, never-skilling, the failure to form competence at all when a tool substitutes for the effort that would have built it, and separated it from deskilling and mis-skilling. The terms are theirs. Their own caution is the part to carry: "Direct causal evidence linking AI exposure during training to competency failure in medical trainees does not exist." It is a risk model. Anyone quoting it as proof has misread it.

So the practical answer sits at the level of the individual team and not the level of the profession. The mechanisms are measured, the interfaces that blunt them are measured, and the long-run outcome is unknown. That is enough to act on and not enough to be confident about, and saying otherwise in either direction would be an overreach.

Key sources

• Liu, G., Christian, B., Dumbalska, T., Bakker, M. A. and Dubey, R. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. arXiv:2604.04721.
• Bastani, H. et al. (2025). Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics. PNAS, 122(26).
• Shen, J. H. and Tamkin, A. (2026). How AI Impacts Skill Formation. arXiv:2601.20245.
• Sankaranarayanan, S. (2026). Mitigating "Epistemic Debt" in Generative AI-Scaffolded Novice Programming using Metacognitive Scripts. arXiv:2602.20206.
• Stromberg, D., Lei, V. and Wu, Y. (2026). The Generative AI Learning Penalty: Evidence from Chinese Secondary Education. CEPR Discussion Paper 21577.
• Cruces, G. et al. (2026). Does generative AI narrow education-based productivity gaps? NBER Working Paper 34851.
• Casner, S. M., Geven, R. W., Recker, M. P. and Schooler, J. W. (2014). The Retention of Manual Flying Skills in the Automated Cockpit. Human Factors, 56(8).
• Arthur, W., Bennett, W., Stanush, P. L. and McNelly, T. L. (1998). Factors that influence skill decay and retention. Human Performance, 11(1).
• Budzyń, K. et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy. The Lancet Gastroenterology and Hepatology, 10(10).
• Ke, Y. et al. (2026). AI-induced never-skilling in medical education. Nature Medicine, 32(6).

Related SuperSkills research#

On what has gone from the ladder, the missing rungs and the missed reps. On what the result looks like from outside, synthetic seniority. On the day-to-day version of this question, should juniors use AI at all and whether apprenticeships still work. On the counter-argument, whether this is deskilling or a rescaling of what counts as skill. On finding out before it matters, assessing capability rather than output.

About this research#

The definition of seniority above is the working definition used on this page and is not offered as a coined term. Never-skilling, mis-skilling and desirable difficulty belong to the researchers credited above. The reading offered here, that junior work was a by-product of senior workload and has to become a deliberate choice, and that the tasks safest to delegate are the long ones and not the hard ones, is an interpretation by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), and is marked as an interpretation and not a finding.