New Research Shows Generative AI Is Eating the Career Ladder Junior Developers Climb
A team at Seoul National University has published findings that shift the debate about AI and software jobs. The question is not just whether generative AI replaces junior developers. It is whether the path those developers take to become senior engineers still exists. The paper, accepted to AIES 2026 (the AAAI/ACM Conference on AI, Ethics, and Society), argues that the tasks AI now handles were never just busywork. They were the mechanism through which expertise was built.
The research team, led by Professor Taesup Moon of the Department of Electrical and Computer Engineering, conducted semi-structured interviews with 14 software developers in South Korea: six senior engineers with at least six years of experience, and eight juniors on the verge of entering the workforce. Korea was chosen as a critical case because AI is involved in 51.8% of work-related activities there, nearly double the U.S. rate, while entry-level hiring has collapsed. Entry-level job postings at the 15 largest U.S. tech firms fell 25% from 2023 to 2024. Korean IT postings dropped 43% over the same period, with entry-level positions accounting for just 4.4% of all listings.
What AI Absorbs Is Not Just Work, but Learning
The core finding is a pattern the researchers call Absorption. Generative AI redirects entry-level tasks into senior-AI workflows. Senior developers now use AI to handle basic implementation, debugging, and documentation that used to go to juniors. One startup founder interviewed for the study said the company had substantially reduced its junior workforce without measurable productivity loss, asking, "What makes a junior developer better than a KRW 100,000-per-month AI subscription?"
The consequences go beyond employment numbers. The tasks AI absorbs are the same tasks through which juniors once learned. The researchers frame this through Kapur's theory of Productive Failure and Bjork's concept of Desirable Difficulties. When learners make mistakes and correct them, they develop not only knowledge but also the ability to recognize what they do not know. Junior participants in the study said they could now achieve good results with less effort, but described a persistent feeling of not knowing what they do not know.
One participant took two courses in the same field, received identical grades in both, but said one left real knowledge while the other left only the ability to use AI. The only difference was whether the assignments could be completed using generative AI. The grade remained stable. The learning diverged sharply. Current educational assessment methods do not capture this distinction.
Why the Problem Does Not Correct Itself
The study identifies two structural forces that prevent self-correction. The first is collective pressure in educational settings. When peers use generative AI on coursework, grading on a curve makes AI use effectively compulsory. One participant described having no freedom to work through assignments alone and making mistakes, because everyone else using GPT was earning near-perfect scores. Faculty, the participants said, did not restrict AI use.
The second force is perceptual asymmetry between seniors and juniors. Senior developers, drawing on years of accumulated experience, can evaluate whether AI-generated output is correct. They see the current situation as manageable. Juniors, who lack that experiential foundation, experience the same situation as a loss of the opportunity to build judgment. One senior participant put it plainly: "We have 20 years of accumulated experience, so we can judge whether AI outputs are right or wrong. The next generation will not be able to reach that position. So we are fine."
The researchers connect this to the theory of situated cognition. What people can perceive depends on the position they occupy. Seniors who have already crossed the developmental threshold see a solvable problem. Juniors who are still trying to cross it see a closed door. Neither group can fix the problem alone, because those with the power to change structures do not fully see the issue, and those who experience it directly do not have the power to change hiring or educational systems.
A Pathway That Was Never Protected
The researchers add an important nuance. The erosion did not start with generative AI. Senior developers described their own growth as the product of informal, hands-on work, not structured training programs. One senior recalled starting as a part-time developer doing what he called "development grunt work," writing code as instructed, debugging daily, testing whatever came back. That same developer later said juniors could be replaced by a cheap AI subscription. The pathway that produced him was never institutionally protected. AI has accelerated the erosion of something that was already fragile.
The study draws parallels to other high-stakes fields. Aviation requires pilots to maintain manual flying skills so automation does not erode them (FAA SAFO 13002). Nuclear operators undergo regular simulator retraining (NRC 10 CFR 55.59). These industries deliberately preserve experience necessary for expertise even when automation can perform the underlying tasks. Software engineering has no comparable safeguard.
What the Researchers Propose
The paper recommends institutional intervention across three areas. In university education, courses where AI cannot achieve learning objectives on a student's behalf should be designated as required, and reducing AI dependence should become a criterion for evaluating educational quality. In hiring, evaluations should test not just a candidate's ability to use AI to produce results quickly, but their ability to recognize gaps in their own knowledge and detect errors while working with AI. In companies, learning opportunities should be deliberately created for entry-level developers, such as assigning small modifications to real products and letting juniors experience the full process from commits through code review to deployment.
The paper is titled "Who Will Become the Next Senior? How Generative AI Erodes the Development Pathway in Software Engineering." It was accepted to AIES 2026, held in Malmö, Sweden from October 12 to 14. The lead author is Sumin Yu, a Ph.D. student at the M.IN.D Lab in SNU's Department of Electrical and Computer Engineering, who researches algorithmic fairness, AI governance, and how generative AI affects human learning and social institutions. The work was supported by the National Research Foundation of Korea, IITP, and the BK21 FOUR program at Seoul National University.