What If AI Is Erasing the Way You Got Good at Your Job?
In August 2025, Anthropic studied 132 of its own engineers and researchers. Fifty-three sat for detailed interviews. The team analyzed how they actually used Claude Code at work. The headline findings were the ones you'd expect from an AI company. People felt more productive, more "full stack," able to take on projects that wouldn't have happened otherwise. Twenty-seven percent of AI-assisted work was net-new, things that wouldn't have been built at all without the tool.
The quieter findings are the ones worth paying attention to.
The heavy users reported worry about their own skills atrophying. One described shifting to roughly 70% reviewer of AI-generated code rather than someone who builds. Several said Claude had become the first stop for questions that used to go to colleagues. One said, in plain words, that they liked working with people, and it was sad they needed them less now.
These are the most AI-fluent engineers on the planet, at the company building the tool. And they're describing the early shape of something organizations are not yet measuring: the slow erosion of how professionals build, sharpen, and pass on the thing that made them good in the first place.
What skill atrophy actually means
The phrase suggests losing existing skills. That happens, but it's only part of the story. The deeper problem is the loss of the conditions that produce skill.
Professionals don't develop judgment in classrooms. They develop it through reps: drafting the bad version, debugging the broken system, sitting with the messy customer call, asking the colleague who's seen this one before. The struggle itself is the curriculum. Each rep produces a small, durable update in how someone reads their work. Strip out the struggle and the update doesn't happen.
This is well-documented outside AI. Three decades of research on expertise (Ericsson and colleagues) consistently finds that what builds judgment is deliberate practice under appropriate difficulty, with feedback from someone who's already done the work. Master-apprentice studies in surgery, law, and trades describe the same pattern. You try things slightly beyond your reach, fail in front of someone who can show you what you missed, and try again with that feedback in your head.
What AI offloading does is quietly remove the friction those reps depend on. The first draft you would have struggled through is generated cleanly in seconds. The debugging session you would have learned from is collapsed into a fix you accept and move on. The question you would have brought to a senior colleague is answered by the tool before you finish framing it. Each handoff feels like a win in the moment. Cumulatively, they remove the difficulty that builds the worker.
The mentorship side is doing damage you can't see yet
The Anthropic engineers, naming a drop in colleague interaction, are reporting the part of the loss that's hardest to measure and easiest to dismiss.
Mentorship has never been a clean, formal channel. Most of it happens in the small moments. The junior person walks over with a half-formed question. The senior person looks up, asks what they've tried, points at the part they missed, and sometimes opens a tangent that turns out to be the real lesson. The junior person learns that the question itself was wrong, which is the actual skill. The senior person sharpens their own thinking by being asked.
That exchange has costs. The senior person's time is the obvious one. The implicit social cost (asking for help, admitting confusion, owing something to a colleague) is the bigger one. AI removes both. There's no time tax, no social tax, and the answer arrives in a form polished enough to feel definitive.
The junior worker loses three things at once. They lose the answer-shaped-as-conversation that taught them how senior people think. They lose the calibration that comes from a real person assessing how lost they are. And they lose the relationship that would have made the senior person an advocate later. None of these show up on a dashboard. All of them are how careers actually get built.
The senior worker loses something too: the cognitive sharpening that comes from explaining your judgment to someone who doesn't yet have it. Teaching is how experts maintain their own expertise. Remove the requests for help, and the senior worker's judgment also starts to drift.
This is a layer of an older trend
Knowledge workers are encountering a new version of a pattern that other workers have been living with for over a decade.
Since at least 2015, researchers (Lee and colleagues; Kellogg, Valentine, and Christin in their 2020 Academy of Management Annals synthesis) have documented how software increasingly performs the work managers used to do: assigning tasks, scoring performance, flagging who's behind. They call it algorithmic management. The findings are consistent across studies. Worker autonomy declines. Mentorship and feedback erode because the supervisory layer that produced them is gone. A January 2026 monitoring study found nearly half of surveilled workers would consider quitting if monitoring increased further. The frontline has been telling us what this feels like for years.
What's specifically new about AI is the location. Earlier tools worked from the outside. They measured your work. AI sits inside the work. It does some of the work with you, and quietly does some of the thinking that used to live in the job. The same erosion of skill, mentorship, and autonomy that the frontline reported a decade ago is now arriving at the knowledge worker's desk through a different door.
Why this hurts more than the productivity gain helps
Self-determination theory (Deci and Ryan, the most empirically supported framework in workplace motivation) identifies three psychological needs that predict engagement and well-being at work: autonomy, competence, and relatedness. Autonomy is the sense that your work is yours to direct. Competence is the felt experience of getting better at something that matters. Relatedness is the sense that the people you work with see and value your contribution.
Cognitive offloading at scale, without redesigning the work around it, threatens all three.
Autonomy gets undercut by adoption mandates and by the felt sense that the tool is shaping what gets done. Competence gets undercut by the loss of the reps and feedback loops that produced it. Relatedness gets undercut by the disappearance of the small mentorship moments that built professional relationships. The Anthropic engineers naming each of these in the same study is not a coincidence. It's what the theory predicts
The productivity gain is real. It also doesn't replace what's being lost, because the things being lost are not in the same currency as output. They're the conditions that make output durable and the worker capable of producing it again next year. An organization optimizing for this quarter's productivity while eroding next year's capability is doing something it would never let happen on the balance sheet. It's doing it on the workforce because the depreciation is invisible.
What this asks of leaders
The work is treating capability-building infrastructure as something that has to be designed in, because the old version is being designed out by default.
Preserve the reps that build judgment. Some work needs to stay slow. Identify which work is capability-building (early career, novel problem, judgment-heavy) and protect it from blanket adoption mandates. A junior analyst who never writes a flawed first draft will not develop the eye that catches one.
Pay for the mentorship that AI quietly displaces. Time spent helping a junior colleague is a higher-cost activity in a system that didn't price it before. If you want it to keep happening, make it count on the senior person's evaluation, rather than relying on their goodwill. The companies that built this kind of incentive into diversity and inclusion programs already know how.
Slow the metrics down. Tracking AI usage and self-reported time saved tells you the tool is being used. It tells you nothing about whether the worker is still becoming a better version of themselves. Layer in measures that catch the slower variable: judgment quality over time, skill development at junior levels, and the rate at which senior people are still being consulted by their teams.
Redesign roles around what humans do that AI cannot, and protect those parts. Workday's 2026 research found 89% of organizations have updated fewer than half of their roles to reflect what AI has changed about the work. That gap is where the erosion does its work uninterrupted. Closing it is the actual change management task.
The honest framing
The Anthropic engineers I quoted at the start of this piece are doing something organizations should learn from. They're naming a loss while still using the tool that produces it. They continue working with AI, and they want their employer to see what it's costing them.
The vaccine research I did years ago kept finding the same thing: when an institution cannot see what its ask means to the person being asked, the campaign fails in the same ways every time. The AI version of that mistake is happening in slow motion, but it's the same mistake. Adoption gets measured. The cost of adoption doesn't show up until it shows up somewhere the organization can no longer recover from. By the time you can see skill atrophy on a performance review, the pipeline that would have produced your next senior person is already several years thinner than you realized.
The leaders who get AI right over the next few years will be the ones who treat the rollout as a question about human capability and design the answer in, rather than letting the answer happen to them.
***
Sources:
Anthropic (2025). How Anthropic Teams Use Claude Code. Internal study, August 2025. 132 engineers/researchers surveyed, 53 interviews. https://www.anthropic.com/news/how-anthropic-teams-use-claude-code
Lee, M. K., Kusbit, D., Metsky, E., & Dabbish, L. (2015). Working with machines: The impact of algorithmic and data-driven management on human workers. CHI Conference on Human Factors in Computing Systems.
Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366–410.
Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363–406. (And subsequent body of work through Ericsson's later career.)
Deci, E. L., & Ryan, R. M. (2000). The "what" and "why" of goal pursuits: Human needs and the self-determination of behavior. Psychological Inquiry, 11(4), 227–268.
Workday (2026). Beyond Productivity: Measuring the Real Value of AI. January 2026. https://newsroom.workday.com/2026-01-14-New-Workday-Research-Companies-Are-Leaving-AI-Gains-on-the-Table
