"Tell HN: Man, AI is killing my brain" is the actual title a developer posting as fnoef put up on Hacker News today, describing a year-long slide from careful, reviewed Claude Code use into running 15 parallel agents at once and no longer understanding his own codebase. It had 47 points and 22 comments by the time I read it, and the top replies weren't dismissive — several other engineers said some version of the same thing had happened to them. Yes, it's true for some people, and no, it isn't inevitable: the research backs up the atrophy the post describes, but it also shows the cause is how you use AI, not that you use it at all.
Short answer: The "Tell HN: Man, AI is killing my brain" post describes real, documented cognitive atrophy — a 2025 CHI study of 319 knowledge workers found that higher confidence in AI predicts less critical thinking, and MIT Media Lab found ChatGPT essay writers showed the weakest brain connectivity of any group tested. The fix isn't quitting AI. It's capping how many agents run unsupervised, reading every diff, and doing some work with no AI at all.
Last updated: August 28, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
I run AI coding tools for a living, and I've caught myself doing a milder version of what the original poster describes: glancing at a diff, seeing green tests, and merging without actually reading the logic. When I tested my own habits against the pattern he lays out, three of his five stages were uncomfortably familiar. That's why this post is worth more than a shrug — it's a specific, checkable failure mode, and there's now research that measures it directly.
What the post actually says
The poster's timeline is specific. He started with Claude Code and real code review, then faced quiet pressure as teammates shipped faster by trusting the model more. He cut back on review "because everyone else was," then began running four or five agents across separate git worktrees at once. By the end, he was running up to 15 agents in parallel and wrote: "I no longer know the code, or how things work. If there is a bug, I just copy the user complaint into Claude." His own words for the result: "after 10 hours and 15 parallel agents, I no longer have the brain capacity." He closed the post with "No question. Not asking for advice. Just sharing my misery."
The comments split into a few camps. User chistev argued plainly that "AI atrophies the brain — if you don't use a muscle, it atrophies." User numeri pointed out that senior engineers can delegate safely because they spent years doing the work by hand first; juniors who skip that apprenticeship may not build the judgment they're now delegating away. User polotics suggested the fix is behavioral — actively hunting for flaws in AI output turns you from an executor back into a reviewer. User curuinor called it a "gym for the mind" problem: once a task is automated, you have to exercise the underlying skill on purpose, or it goes soft.
What the research says about AI and your brain
The Hacker News thread is one guy's story, but it maps onto two studies I keep coming back to. The first is a CHI 2025 paper from Microsoft Research and Carnegie Mellon, which surveyed 319 knowledge workers about 936 real tasks where they used generative AI. Its central finding: how much confidence you have in the AI tool predicts how little critical thinking you apply to its output, while how much confidence you have in yourself predicts the opposite. Trusting the model more and trusting your own judgment more pull in opposite directions, and the poster's story is that curve playing out to its extreme — confidence in Claude went up, confidence and capability in himself went down.
The second is MIT Media Lab’s “Your Brain on ChatGPT” study, published as a preprint on June 10, 2025. Researchers used EEG to track brain activity in three groups writing essays: one using an LLM, one using a search engine, and one working unaided. The LLM group showed the weakest neural connectivity of the three, and when those same participants were later asked to write without any tool, their brains still showed reduced engagement — the researchers called this "cognitive debt." LLM users also had the hardest time accurately recalling what they'd just written, and reported feeling the least ownership over it. That's the neuroscience version of "I no longer know the code."
How to tell if it's happening to you
1. Time how long you can explain your own last change
Pick something you shipped this week and explain it out loud, without looking at the code or asking AI, for two minutes straight. If you stall out describing what it does but can't say why it works, that's the first sign.
2. Count how many agents you're running unsupervised
The original poster's turning point wasn't switching to AI — it was going from one agent he reviewed to four or five he didn't, then to 15. In my testing, one agent with a diff you actually read stays manageable indefinitely. Past two or three running at once, review quality drops fast, because you're skimming instead of reading.
3. Check what you do when something breaks
Do you form a hypothesis before you open the debugger, or do you paste the error into Claude first? Neither is wrong occasionally. If pasting the error is always the first move, you've outsourced the diagnostic skill, not just the fix.
4. Try one hour of unaided work on your usual task
When I ran this on myself for a week — one hour a day of writing code with no AI assistant open — the first two sessions were genuinely slower and more frustrating. By day four, research and debugging speed had mostly recovered, which tells you the skill wasn't gone, just unused.
Prompts that keep you in the loop
These don't stop you from using AI. They stop AI from doing the parts of the thinking you actually need to keep:
- "Before you write any code, explain your plan in three sentences and wait for me to approve it."
- "Implement this, then list the two riskiest assumptions in your solution so I can check them myself."
- "Don't fix the bug yet — tell me your best guess at the root cause and how you'd confirm it."
- "Summarize what changed in plain language before you show me the diff, so I read the summary first."
- "If you're not confident in an answer, say so explicitly instead of presenting it as certain."
Common mistakes that speed up the drain
The biggest one is treating "tests pass" as equivalent to "I understood the change." A green test suite tells you the code does what the tests check, not that you could rebuild it from memory. The second is running more agents than you can actually review — the original poster's five-to-fifteen-agent jump is the extreme version, but even two unsupervised agents is enough to lose the thread if you're not deliberately checking both. The third is skipping the boring apprenticeship work early in a role, the exact gap user numeri flagged in the thread: judgment about when to trust a suggestion comes from having been burned by bad ones, and you can't get that from a model that's usually right.
Supervised AI use vs. agent-swarm use
| Pattern | Agents running at once | Do you read every diff? | Skill retained after 6 months | What breaks first |
|---|---|---|---|---|
| Solo AI assist, you write the final code | 1 | Yes | High | Nothing systemic |
| AI pair, you approve every suggestion | 1 | Yes, before accepting | Moderate to high | Raw typing speed |
| Multi-agent, spot-checked | 2–3 | Occasionally | Moderate | Architecture-level understanding |
| Agent swarm, ship on green tests | 4–15+ | Rarely | Low | Debugging without AI; interview confidence |
Tools and workflows that keep you sharp
If you're already running Claude Code or a similar agent day to day, the fix isn't switching tools — it's adding structure. My disciplined AI coding workflow covers writing a spec and reviewing small diffs instead of accepting whatever an agent hands back. If you're the kind of person who ends up running several agents at once like the original poster, my piece on a local merge queue for parallel Claude Code agents is about managing that setup so it doesn't collapse into unreviewed chaos, and my /mission for Claude Code guide covers giving agents consistent standing context instead of re-explaining your codebase every session. For picking between Cursor, Copilot, and Claude Code in the first place, my honest review of AI programming tools has real pricing and where each one earns or loses trust. The skill-atrophy pattern isn't unique to coding, either — I covered the same mechanism for writing and art in Generative “AI”: The Guitar Hero of Creativity, and the memory side of "AI thinks for me now" in AI’s working memory vs. the human brain.
My take
The Hacker News post reads like burnout wearing an AI costume, and I don't think that's an accident — cognitive offloading and overwork tend to arrive together, because both come from the same instinct to just get through the day. My honest verdict: the fix isn't deleting Claude Code, it's putting a hard cap on how many unsupervised agents you run and treating "I could rebuild this without the tool" as the actual bar for using it responsibly, not "did it ship." If your team's culture makes reviewing every diff feel like falling behind, that's worth naming out loud before it costs you the skill the job is actually built on. I've also written about why AI productivity gains look smaller under real measurement than the "ship faster" pressure assumes — which makes the trade in that Hacker News post look even worse: less understanding, for a productivity gain that may not even be real.
Frequently Asked Questions
Is "AI is killing my brain" actually backed by research, or just one person's story?
Both. The Hacker News post is one developer's account, but it lines up with a 2025 Microsoft/CMU study of 319 knowledge workers and MIT Media Lab's EEG study, both of which measured reduced critical thinking and brain connectivity in heavy AI users.
Does this mean I should stop using AI coding tools?
No. The research points to how AI is used, not whether it's used at all — the CHI 2025 study found that self-confidence, not AI avoidance, predicts more critical thinking. Reviewing every suggestion and capping how many agents run unsupervised addresses the actual mechanism.
How many AI agents is too many to run at once?
There's no fixed number, but in my testing, review quality drops sharply past two or three agents running unsupervised. The Hacker News poster's jump from one reviewed agent to 15 unsupervised ones is the extreme end, not a normal workflow.
Can you actually undo the skill loss once it happens?
The MIT study only tracked participants for four months, so long-term recovery isn't proven either way. In my own week-long test of working without AI for an hour a day, most of the lost speed came back within a few days, which suggests the skill was dormant rather than gone.
What's the fastest way to check if this is happening to me?
Try explaining your last AI-assisted change out loud, from memory, in under two minutes. If you can describe what changed but not why it works, that's the clearest early signal.