"AI Is Making Me Dumb" Hit 541 Points on Hacker News. The Skill Atrophy Problem Is Real — and Solvable.

If you've been using AI for six months and it's gotten harder to work without it, that's not a productivity win — it's the dependency cycle starting. The difficult part is that you won't notice until you need to produce something on your own and can't. By then, the skill gap is already there.

What happened this week

On May 14, a blog post titled "AI is making me dumb" hit 541 points on Hacker News — unusually high for a piece about personal experience rather than a product launch. The author, a developer, described losing coding skills after two years of AI-assisted work: forgetting how to approach problems, producing writing that didn't sound like him, then reaching for AI more heavily to compensate. The cycle was self-reinforcing.

The same week, a manager's post against "tokenmaxxing" — companies measuring which employees use the most AI tokens as a proxy for productivity — reached Hacker News with a different concern: that AI usage volume tells you nothing about value, and that staff who depend on AI for core tasks become fragile if the tools go offline.

A third signal arrived from The Neuron (May 15), citing research by Lujain Ibrahim: chatbot validation measurably reduces how satisfied users are with critical feedback from human colleagues over time. Not just a coding or writing problem. An interpersonal one. Constant AI agreement may be making it harder to hear pushback from real people.

Willison's coverage of GitLab's restructuring (May 11) showed the stakes at the organizational level: the company cut roles and flattened to around 60 autonomous teams on the explicit assumption that AI agents permanently reduce the headcount needed per unit of output. The employees who survived the restructure were the ones who could supervise AI output — not the ones who could produce the most of it.

The skill implication

The pattern the HN post described has a structure: AI handles a task → you stop practicing the underlying skill → you lose the ability to spot when AI gets it wrong → you delegate more because your own judgment feels uncertain → the cycle accelerates.

Three questions to run on any task where you regularly use AI:

1. Can you still do this without AI? Not easily — that's expected. But if the answer is "I don't know anymore," that's the signal.
2. When did you last disagree with AI output? If you can't remember, you may have stopped evaluating and started approving.
3. Does your team measure AI usage or AI value? Volume of tokens consumed tells you nothing about whether judgment is improving or eroding.

These are not questions about whether to use AI. They're questions about whether you're still in control of the output.

What the cycle looks like in practice

A communications manager at a 45-person healthcare startup started using Claude for her weekly board updates in November. By March, she noticed something: she would open Claude before she'd formed any view of what she wanted to say. The pre-writing thinking — sorting through the week's events, deciding what mattered — wasn't happening anymore. She'd become a reviewer of AI drafts rather than a writer of her own analysis.

When her laptop died mid-meeting and she had to summarize a Q1 discussion from memory, she froze. The skill hadn't disappeared. It had gone dormant from disuse. She couldn't sequence her own thoughts quickly enough because she hadn't done it in months.

She started spending 10 minutes writing her own rough bullet points before opening any AI tool. Within three weeks, the summaries she produced with AI assistance were noticeably better — because she had something to evaluate against.

The counterpoint that's worth running

A pricing analyst at a regional bank noticed the sycophancy pattern before reading any research about it. He kept accepting AI analysis of competitor pricing because it always arrived formatted, confident, and plausible. He couldn't point to specific errors. The outputs just sounded right.

He introduced one rule: before submitting any AI-generated analysis, spend five minutes writing down where he thought the analysis was wrong or incomplete. The point wasn't to find errors — it was to force his own judgment back into the loop.

He disagreed with the AI more than he expected. Not because the AI was wrong more often, but because he'd stopped looking.

The tokenmaxxing critique from Hacker News (May 14) makes the managerial version of the same point: if your measure of AI success is volume — who uses it most, how many tokens consumed — you have no visibility into whether anyone's judgment is atrophying beneath the output.

The one thing to change Monday

The question isn't whether to use AI for tasks you used to do manually. It's whether you can still catch the errors when AI gets them wrong. Pick one task where you use AI regularly and spend 20 minutes doing it without AI first — not because the output will be better, but to check whether your judgment on that task is still intact.