The paradox of supervision is the name Anthropic’s study How AI is transforming work at Anthropic gives to one reason the atrophy of coding skills is concerning: effectively using Claude requires supervising it, and supervising Claude requires the very coding skills that may atrophy from AI overuse. It is the report’s link between Skill Atrophy — skills getting less practice as more work is delegated — and the active supervision and validation that its respondents describe as part of using Claude.
Usage
The report uses the term in its account of interviews with Anthropic engineers and researchers about skills getting less practice, and ties it back to its survey finding that more than half of respondents could “fully delegate” only 0–20% of their work to Claude, with engineers describing active validation of its output, particularly for complex or high-stakes work. One interviewee put the concern this way: they worried much more about “the oversight and supervision problem” than about their skill set as such, because skills that atrophy or fail to develop would be a problem mainly for their ability to use AI safely on the tasks they cared about, rather than for their ability to do those tasks independently.
The study also records how some engineers respond: deliberately practising without AI from time to time, even when Claude could solve the problem, to keep themselves sharp. A senior engineer said they would be more worried if they were more junior, since they had built the ability to judge what an answer should look like by doing software engineering “the hard way,” and growing that ability while relying on the model would take deliberate effort. The report presents these as the views of its interviewees, and notes that engineers were divided on how much the underlying skill loss matters.