At the brief's latest capture, a warning that AI coding will hollow out the next generation of software expertise had reached 509 points and 397 comments on Hacker News. That response is a useful measure of developer concern, rather than proof that the prediction is right. The essay making the case connects a genuine problem to a much larger forecast: assistants can let people finish unfamiliar work without understanding it, but the available studies do not establish that the profession's expertise pipeline will collapse.
The distinction matters for engineering teams deciding how junior developers should use coding agents now. The strongest evidence is about short learning tasks, immediate comprehension and the way a person interacts with an assistant. It says far less about what happens over years of work, how mentoring changes the result, or whether newer agents alter the trade-off between speed and learning.
The warning has a measurable core
Lars Faye's argument starts with an uncomfortable dependency. Reviewing generated code requires enough knowledge to spot a bad abstraction, a missing edge case or an invented API. A beginner who delegates the work used to acquire that knowledge may still produce a passing result. The essay calls this a threat to skill formation, especially when companies reward output speed and treat AI use as mandatory.
A 2024 study of novice programmers observed that split at close range. Researchers ran 21 lab sessions using interviews, observation and eye tracking. Twenty participants completed the assigned problem, yet the routes they took differed. Students who already knew what code they wanted could use generative AI to move faster and dismiss poor suggestions. Students who struggled were more likely to be pulled further off course and to finish with an inflated view of their understanding, according to the paper accepted at ICER 2024.
That finding is easy to flatten into a verdict against AI assistance. The study itself describes a divide among users. Its productive group had enough of a mental model to judge suggestions; the struggling group lacked the metacognitive awareness to recognize when the tool had made matters worse. AI amplified a difference that was already visible in how the students approached programming.
A newer randomized experiment gives the concern firmer quantitative support. Anthropic researchers recruited 52 software developers, most of them junior, who had used Python weekly for at least a year but did not know the Trio asynchronous programming library. Participants completed two Trio tasks with or without an AI assistant and then took a quiz. The AI group averaged 50%, compared with 67% for the hand-coding group. Their roughly two-minute speed advantage was not statistically significant, according to Anthropic's report.
Debugging produced the largest gap. That detail is more consequential than a missed syntax question because debugging is also how developers check generated code when it fails. The experiment's paper found weaker conceptual understanding, code reading and debugging in the AI-assisted group on average. Full delegation was faster for some participants, but they learned less about the library they had just used.
Usage within the AI group varied enough to complicate the average. Anthropic identified six interaction patterns. Four people who delegated the whole task, four who gradually handed it over and four who relied on AI for debugging all averaged below 40% on the quiz. Seven participants who asked conceptual questions and wrote the code themselves were the second-fastest interaction group overall while retaining stronger understanding. The researchers treat those clusters as associations, not causal findings, because each contains few people.
A short quiz cannot measure a career
Both studies support a focused claim: handing an unfamiliar programming task to an answer-generating system can reduce what a novice learns during that task. They do not demonstrate that developers who use AI will fail to become experts. The 2024 work observed 21 lab sessions. Anthropic measured comprehension minutes after one Python-library exercise and explicitly says it does not know whether those quiz results predict long-term skill development.
The gap between those statements is where the viral argument stretches. A workforce pipeline is shaped by repeated assignments, code review, production incidents, formal instruction and changing responsibilities. Neither experiment followed promotions, maintenance work or performance under a real outage. Neither compared teams with strong mentoring against teams that simply accept generated patches. Those missing conditions do not disprove the warning. They limit its reach.
The evidence also separates people more clearly by behavior than by access. In the ICER study, capable novices used AI to express code they already intended to write. In Anthropic's experiment, participants who asked for concepts or explanations retained more than those who asked the assistant to take over. An IDE toggle marked the experimental condition, but the cognitive choice happened at the prompt: seek an answer, or use the response to build a model of the system.
That makes the headline claim, that AI coding will prevent expertise, too deterministic. A tool can remove productive struggle from a task, and a workplace can encourage that pattern at scale. It can also expose a learner to an explanation on demand and let an experienced developer skip work they have already mastered. The studies show these modes producing different outcomes. Treating all assistant use as the same intervention discards their most useful result.
Productivity is moving under the argument
The practical pressure to delegate is unlikely to stay fixed. METR's early 2025 randomized trial had 16 experienced open-source developers complete 246 tasks in repositories they knew well. Allowing AI made them 19% slower even though they believed it had made them faster, as recapped in METR's follow-up.
By February 2026, METR said its follow-up data pointed in another direction and was harder to interpret. Among ten returning developers, the estimate shifted to an 18% speedup, with a confidence interval that still included a 9% slowdown. Newly recruited developers had an estimated 4% speedup, again with an interval spanning no effect. The researchers changed their experiment design because developers who valued AI most often refused AI-free tasks, while some participants ran agents in parallel and could not cleanly report task time.
That update does not settle whether agents make developers faster. It shows why a prediction about expertise must account for changing tools and work habits. A slow assistant may invite a developer to inspect every suggestion. An agent that completes a routine migration in the background creates a different learning environment. Measuring access to AI without recording delegation, explanation requests, review effort and task familiarity will blur those cases together.
Teams do not need to wait for a decade-long study to measure the risk locally. For work intended to teach an unfamiliar system, a review can require the developer to explain the generated change, alter it without the assistant and diagnose a seeded failure. Routine transformations can be evaluated on delivery time and defect rate instead. This separates learning work from production work using observable outcomes, rather than assuming that every generated line has the same educational cost.
Tool makers have a more direct design question. Anthropic's highest-scoring patterns included conceptual inquiry and requests that paired code with explanation. An assistant can make those actions easier to choose, while still allowing direct generation when the user knows the domain. The experiment did not test which interface intervention works best, so claims that a learning mode solves the problem would be premature.
The missing test lasts longer than one afternoon
The next useful result needs to follow developers across months, record how they use agents and test whether they can maintain or debug the same systems without assistance. It should also separate unfamiliar learning tasks from repetitive work, since Anthropic's study tested the former and METR found that task selection changes when agents are available. Until then, the 509-point debate has identified a credible engineering-management risk. It has not shown that expertise is destined to disappear.