AINEWS
Junior developers now ship features with AI assistants they cannot fully read, and the friction that built the underlying skill is being absorbed into the tools themselves.
A junior developer ships a feature with the help of an AI coding assistant. The code passes review, gets merged, and runs in production. Then it breaks, and the same developer cannot read the output well enough to fix it. That is the failure mode Lars Faye, a Swedish practitioner, named the Expert-Novice problem in a recent essay on his personal blog.
Faye's argument, laid out in the essay, is structural rather than moral. AI coding assistants are not making developers worse at their jobs in some vague "kids these days" sense. They are reshaping the training pipeline itself, in a way the industry's own messaging does not acknowledge. The tools demand expert judgment to use well. The same tools absorb the manual reading and writing that built that judgment in the first place. The contradiction is not a paradox to manage around. It is the actual shape of the problem.
The cost shows up first inside enterprise teams. On a Hacker News thread that picked up Faye's essay, a senior engineer described a workplace where management had embraced AI for everything from sprint planning to code review. The engineers' job was now "mostly filtering AI output rather than building." A prompt like "hey Claude, read this Jira ticket and implement the feature in this code base" produces a 1,500-word Jira ticket, mostly LLM boilerplate, and a code change that looks plausible on first read. Neither is a $200,000-a-year engineer's deliverable. The human in the loop is now a filter, and the filter is in short supply.
What the industry is saying out loud, in the same window, makes the contradiction sharper. On one side: "If you're writing code manually, you're doing it wrong." On the other: "Vibe coding is a dead end." Vibe coding, the practice of shipping AI-written code without reading it, leaves the developer without the judgment that takes years to build. Both messages are landing on the same audience: the cohort of developers now entering the workforce, the people expected to produce the "experts" these tools will require three to five years from now.
The deeper problem is that previous tool shifts were absorbed by the pipeline without destroying expertise, because none of them removed the manual reading-and-writing loop. Compilers did not stop developers from writing code. Stack Overflow did not stop them from reading it. AI coding assistants are the first tool layer that does, at least at the entry level, by producing complete outputs the user can ship without parsing. The friction that builds expertise is being automated away at the exact moment the expertise required to use the tool well is going up.
The commercial logic is straightforward. Sam Altman of OpenAI has framed the market plainly. "We see a future where intelligence is a utility like electricity or water and people buy it from us on a meter." That is a per-token subscription business, and it depends on a steady supply of customers who can use the product. It does not depend on those customers being able to do the underlying work without the product. The company building the meter has limited incentive to maintain the apprenticeship pipeline that the meter would, in the long run, hollow out.
What the argument needs, before it can carry more weight than one practitioner's blog, is a second voice. A CS educator, a hiring manager at a company that has rolled out AI coding across its engineering org, or a published study on skill atrophy under AI assistance would each let the lens escape the single-author frame. The HN thread is corroboration of a felt problem, not independent measurement of one.
What the piece can already say is this: the contradiction is not in the developers. It is in the pipeline. The tools are being mandated before the training loop has been rebuilt, and the people being told to use them today are the same people who will be responsible for the systems these tools are being used to build. That is the question the industry has not yet answered, and it is the one a working developer, a tech-adjacent manager, a CS student, or a parent paying for a CS degree should be asking before the next onboarding class starts.