Does AI assistance erode the skills needed to oversee it?

Anthropic's own study of its engineering workforce finds that heavy Claude use raises output and changes what engineers do, and that the delegation behind the gains is what staff fear will erode the skills needed to check the work. The company surveyed 132 engineers and researchers in August 2025, conducted 53 in-depth interviews and studied internal Claude Code usage data. Employees self-report using Claude in 60% of their work and a 50% productivity boost, up from 28% and +20% a year earlier. Anthropic's own pull-request measurement adds a 67% increase in merged pull requests per engineer per day after the Engineering org adopted Claude Code. Yet most employees say they can "fully delegate" only 0-20% of their work, and the excerpt calls Claude "a constant collaborator" whose output still needs "active supervision and validation."

The excerpt traces the gains to breadth. Engineers describe taking on work they would once have avoided, such as building front-end or database code outside their specialty, and 27% of Claude-assisted work is work that "wouldn't have been done otherwise." Claude Code's autonomy also grew, from about 10 actions before needing human input six months earlier to about 20 now. The cost runs the other way. Some engineers lose the "collateral" learning that comes from manual problem-solving, and the excerpt names a "paradox of supervision": using Claude well requires supervising it, and supervising it requires the coding skills that overuse may let atrophy. Employees also say high-level design and taste stay with people, and the survey shows the smallest productivity gains on design and planning tasks.

The survey is the workplace counterpart to the experiment in Does AI assistance actually harm the way developers learn?. That note reports a randomized trial in which AI use impaired conceptual understanding; this excerpt measures no learning outcome, only what engineers say they worry about. Its own time figures, "slightly less time per task category, but considerably more output volume," match the reallocation reading in Does AI really save time, or just change how we spend it?: the gain appears as volume rather than as saved time. The move from colleagues to a model echoes Does generative AI shift knowledge workers away from communication?, since one employee says "like 80-90% of them go to Claude." The erosion also lands differently than in Does generative AI prevent juniors from getting entry-level work?. That note concerns juniors who never get the struggle that builds expertise. This excerpt concerns experienced engineers, one of whom has "been programming for 25 years," worried about skills they already have, which makes the supervision argument more direct.

These are Anthropic's figures about its own staff, who build the tool. The productivity numbers are self-assessments, and the pull-request rise is internal telemetry whose baseline the excerpt does not give. The interview quotes are illustrative, and the excerpt does not say how they were chosen. It also stops before the "Looking Forward" section that describes the steps Anthropic is taking, so nothing here shows whether those steps help. The skill-atrophy worry is credible coming from engineers closest to the tool, but the excerpt does not show that atrophy is happening at the scale they fear. Treat it as a well-placed hypothesis about where the supervision problem appears, and take the learning claim from the controlled study instead.