Shuichiro Ogawa
日本語

Notes · updated 2026-10-11

What Should Juniors Do When Generative AI Takes Away the Entry-Level Work? (Academic Research)

This note examines, through labor economics, research on expertise and apprenticeship, randomized experiments on generative AI and learning, and studies of creative work, the problem that generative AI takes over simple tasks so that juniors are asked to handle difficult work without the entry-level practice that used …

Contents (7)
  1. The question and the scope of this note
  2. Has junior work shrunk?
  3. What did the groundwork do for expertise?
  4. Does using generative AI replace practice?
  5. What is known about creative work?
  6. What juniors can do
  7. What research has not yet answered

The question and the scope of this note

When generative AI takes over simple tasks, what remains for people is the difficult work that AI cannot handle. Simple tasks were also where newcomers built their skills, so newcomers are now expected to deliver results on difficult work without that groundwork. In sports terms, this is like being sent into a real match without the practice swings or ball drills that normally come first. This note checks this view against research in three parts: whether junior work has actually shrunk, what the groundwork did for expertise, and whether using generative AI replaces practice. It then lists the behaviors for juniors that research supports, together with the strength of the evidence. The focus is creative work (design, illustration, writing), with learning in programming and mathematics, where experiments are more plentiful, used as a supplement. An account based on hiring data and statements from people in the industry is in the industry counterpart, What Should Juniors Do When Generative AI Takes Away the Entry-Level Work? (Industry Sources).

Has junior work shrunk?

The shrinking itself is observed in US payroll records and résumé data. Using US payroll records, Brynjolfsson, Chandar, and Chen showed that employment of 22- to 25-year-olds in AI-exposed occupations was 19% lower than it would have been had it followed the path of the same age group in less exposed occupations, with no comparable gap for experienced workers (Brynjolfsson et al. 2026). The difference came mainly from reduced hiring. The authors present this as a description, not a causal estimate. Hosseini Maasoum and Lichtinger followed firm-level employment by seniority in the résumés of 65 million workers and reported that, after adoption, junior employment fell sharply at firms that posted jobs for integrating generative AI relative to firms that did not, while senior employment kept growing (Hosseini Maasoum and Lichtinger 2025). They call this seniority-biased technological change.

Attributing this shrinking to generative AI is contested. Lambert and Schindler analyzed hiring and postings in the US, UK, Canada, and Australia and reported that when exposure to remote work and exposure to generative AI are estimated jointly, the generative AI coefficient attenuates and often becomes statistically insignificant (Lambert and Schindler 2026). The two exposures overlap in the same occupations, so estimating only one of them can mistake one for the other. Humlum and Vestergaard linked Danish firm surveys with administrative employment records and found precise null effects of AI chatbot adoption on earnings and recorded hours (Humlum and Vestergaard 2025). In a Dallas Fed analysis, the share of employment held by 20- to 24-year-olds in the most AI-exposed occupations slipped from 16.4% in November 2022 to 15.5% in September 2025, but even if all of that decline had turned into unemployment, it would account for only about 0.1 percentage point of aggregate unemployment (Atkinson and Yamco 2026).

In short, the shrinking of junior employment is observed, but whether its cause can be attributed to generative AI is contested among studies. Whatever the cause, juniors now have fewer occasions to gain experience on simple work, and that is reason enough to ask the questions in the sections that follow.

What did the groundwork do for expertise?

Seeing the groundwork as a matter of practice volume has limits. Ericsson and colleagues argued that individual differences, even among elite performers, are closely related to the amount of deliberate practice, that is, practice structured to improve performance (Ericsson et al. 1993). In the meta-analysis by Macnamara, Hambrick, and Oswald, however, deliberate practice explained 26% of the variance in performance for games, 21% for music, 18% for sports, 4% for education, and less than 1% for professions (Macnamara et al. 2014). In professional expertise, a shortage of practice volume is not the whole problem.

Another function of the groundwork was that it provided a route into the work alongside experts. Lave and Wenger called the learning by which newcomers take legitimate part in the peripheral work of a community of practice and gradually move toward its central work “legitimate peripheral participation” (Lave and Wenger 1991). In this view, simple peripheral work doubled as a place to watch experts’ judgment up close.

That automation can break this route was pointed out well before generative AI. Bainbridge described the irony that the further automation advances, the more what remains for humans is the difficult handling that cannot be automated, while everyday opportunities to practice shrink (Bainbridge 1983). The view at the start of this note is a generative AI version of these “ironies of automation.” Beane observed that when robotic surgery was introduced, it greatly limited trainees’ role in the work and made the approved ways of learning surgery ineffective, so that “shadow learning,” practices that challenged norms and policies, emerged (Beane 2019).

Economic theory shows that this route can break independently of the number of jobs. In Ide’s model, the transfer of tacit knowledge from experts to novices cannot be written into contracts, and improvements in entry-level automation can reduce growth and welfare even without reducing entry-level employment, by reallocating novices away from the most productive experts (Ide 2025). Garicano and Rayo treat apprenticeship as a bargain in which juniors pay for training by doing menial work, and show that when AI performs that work, apprenticeship may cease to pay (Garicano and Rayo 2025). In the same model, however, if the AI-augmented value of a graduate sufficiently exceeds AI’s standalone output, apprenticeship remains as viable as before. Collapse is not the default; it depends on conditions.

Does using generative AI replace practice?

Randomized experiments on generative AI and learning split by how AI is used. Bastani and colleagues gave nearly a thousand high school students in Turkey access to GPT-4 for mathematics practice; grades during practice rose, but once access was removed, students in the standard chat-style group scored 17% lower than those who never had access (Bastani et al. 2025). In a group whose tutor was designed to give hints instead of answers, this harm was largely mitigated. In an experiment with 1,222 participants, Liu and colleagues likewise found that AI raised short-term performance but that people did significantly worse without it and were more likely to give up (Liu et al. 2026).

Harm tends to appear when the thinking is handed over. Shen and Tamkin, affiliated with Anthropic, randomly assigned developers learning a new programming library and reported that AI use impaired conceptual understanding, code reading, and debugging, without significant efficiency gains on average (Shen and Tamkin 2026). Within the AI group, patterns of use that kept the thinking with the learner preserved learning. In the preregistered experiments of Lehmann, Cornelius, and Sting, using an LLM to substitute for learning activities increased the number of topics covered but decreased understanding of each topic, whereas using it for explanations increased understanding (Lehmann et al. 2025). In an experiment with 117 university students by Fan and colleagues, the ChatGPT group improved its essay scores, but knowledge gain and transfer did not differ between groups (Fan et al. 2025).

Trying first is a condition that turns AI help into learning. In preregistered experiments by Kumar and colleagues, LLM explanations improved learning more than seeing only correct answers, and the benefit was largest for those who attempted the problems on their own first (Kumar et al. 2023). This points in the same direction as research showing that solving problems before instruction leads to better learning than instruction followed by problem solving. In the meta-analysis by Sinha and Kapur of 53 studies, the effect of solving first was a Hedges’ g of 0.36 (Sinha and Kapur 2021). In the experiment by Contractor and Reyes, AI access raised scores on unaided tests, and a smaller gain persisted a week later, but the gains remained for students who used AI as a tutor and faded for students who had AI write for them (Contractor and Reyes 2026).

What is known about creative work?

In creative work, generative AI raises the output of less experienced people substantially. In the experiment by Doshi and Hauser, stories by writers who could draw on LLM ideas were rated as more creative and better written, especially among less creative writers (Doshi and Hauser 2024). Hou and colleagues ran lab and field experiments with students and professional designers; AI raised everyone’s creativity in the ideation stage, while in the implementation stage novices kept benefiting but experts spent more time without becoming more creative (Hou et al. 2025). Stories written with AI ideas in the Doshi and Hauser experiment, however, were more similar to each other.

Raising output and developing skill are separate questions. In posting data covering more than four million artworks, Zhou and Lee reported that artists who adopted image generation AI became 25% more productive and 50% more likely to receive a favorite per view, while average content novelty declined, and argued that ideation and filtering become key skills (Zhou and Lee 2024). Fu and colleagues had 36 designers create advertisements with and without generative AI and reported gains in divergent thinking among those new to generative AI (Fu et al. 2026). Both measure artworks or short tasks, and neither measures how juniors’ skills develop over months or years.

At the entry point to the work, there are signs that creators who do not use AI are being pushed out. Kim, Jin, and Lee reported that after the launch of image generation AI, illustrators on Pixiv who did not adopt AI as a primary tool posted less (Kim et al. 2026). Research on image generation AI and commissioning is collected in Do Decision-Makers Without Taste Stop Commissioning Creators? Checking the Problem of Image-Generating AI Through Three Links (Commissioning, Habituation, and the Creator Base) and Relating It to Research on AI Slop.

What juniors can do

From the research above, six behaviors for juniors have some support. The strength of the evidence differs by item and is noted for each. Most experiments involve students and programmers, and no study was found that followed the effects of these behaviors in creative juniors over time.

The first is to try first, and then ask AI for explanations. Learning was best when people first worked on problems themselves and then received explanations (Kumar et al. 2023), and the advantage of solving first is supported by a meta-analysis (Sinha and Kapur 2021). The evidence consists of randomized experiments with students and a meta-analysis.

The second is to use AI for understanding rather than handing it the generation. Substituting AI for one’s own work lowered understanding, whereas asking for explanations raised it (Lehmann et al. 2025; Shen and Tamkin 2026; Contractor and Reyes 2026). These are also experiments with students and developers, measuring from right after the task to about a week later.

The third is to keep time to check oneself without AI. Gains from practicing with AI can disappear once AI is removed (Bastani et al. 2025; Liu et al. 2026). Retrieving one’s understanding without help and spacing practice over time are rated as high-utility techniques in a review of learning techniques (Dunlosky et al. 2013). In creative terms, this corresponds to regularly keeping time to draw, write, or lay out work without AI. No study was found, however, that measured the effect of this kind of practice on creative skills.

The fourth is to seek feedback actively. As more situations can be settled by asking AI, there are fewer occasions to ask experts. A meta-analysis of feedback seeking at work found that organizational tenure and age are negatively related to seeking feedback, so newcomers tend to seek it more (Anseel et al. 2015). The same meta-analysis, however, found that the relationship between feedback seeking and performance was small. What one asks and how may matter more than asking itself, and research has not yet answered this.

The fifth is to practice the judgment of ideation and selection. The more generative AI increases the number of works, the more weight falls on deciding what to make and choosing among what comes out (Zhou and Lee 2024). Because works made with AI ideas become more similar (Doshi and Hauser 2024), the judgment to pick out differences among similar works becomes what distinguishes creators. This is an inference drawn from observational studies and short experiments; whether practicing selection develops skill has not been measured.

The sixth is to choose places where one can work alongside experts. What may be lost is less the volume of practice than the route of watching experts’ judgment up close (Lave and Wenger 1991; Macnamara et al. 2014; Ide 2025). When the official route narrows, people end up learning outside the norms (Beane 2019). Whether apprenticeship remains viable depends on how far the AI-augmented value of a graduate exceeds AI’s standalone output (Garicano and Rayo 2025). From the junior’s side, choosing work and workplaces where one can come into contact with experts’ judgment is a way to keep this route open. This is an inference from theory and observational research, and no study has compared the effects of such choices.

What research has not yet answered

No longitudinal study was found that followed how creative juniors develop their skills over several years while using generative AI. Peer-reviewed research on early careers in video and advertising was not found either. Experiments on generative AI and learning almost all use tasks lasting from tens of minutes to a few weeks, with students and programmers as participants. No research focused on creative workplaces in Japan was found.

Unverified items

No unverified statements remain in the main claims.

References

Author: Shuichiro Ogawa (Design Researcher / Consultant) About me →