Your Brain on ChatGPT: The Productivity Trap Nobody Wants to Talk About

There's a number floating around that gets quoted in every AI earnings call and boardroom deck: 26% more tasks completed when developers get an AI coding assistant, and roughly 25% less time spent on email for knowledge workers. The field experiments behind those figures are real — 4,867 developers here, 6,000 workers there, both measured with actual output data.

But MIT's Media Lab just published something that makes those numbers feel a lot less comfortable. They strapped EEG monitors on people writing essays, half with a chatbot's help and half without. The AI-assisted group produced faster drafts — no surprise there — but their brain engagement dropped noticeably while writing. And here's the kicker: once the tool was removed, they performed worse than the people who never touched it.

The researchers gave that pattern a name: cognitive debt. It accrues quietly, like interest on a card you forgot to check, and it comes due the moment the assistance disappears.

I've been thinking about this while playing with a pair of DGX Sparks sitting on my desk — two mini PCs, 128GB of unified memory each, linked over a ConnectX-7 cable, running DeepSeek V4 Flash (a 284-billion-parameter model) across both. It reads my entire codebase and files GitHub issues for me when it spots a problem. Genuinely impressive setup, the kind of thing that would have sounded like science fiction two years ago.

But I caught myself doing something weird. I stopped reading the diffs. The agent flagged a potential bug, I glanced at the summary, and I moved on. It took a minute to notice what had happened: I outsourced the first question — "wait, is this actually a bug, or just a false positive?" — to the machine.

That's exactly the failure mode Dr. Vishal Kapoor, a researcher who's been digging into this at major banks, warns about. He calls the discipline "AI-fed, human-led": you show up with a hypothesis, a defined gap, a genuine question you care about, and then let the model chew on the evidence and stress-test your reasoning. What you don't do is hand over the first question or the final decision. He measured his own cognitive decline on a brain-training tool after a period of heavy AI-assisted reasoning — roughly 40-50 points on a 1,000-point scale. Not a controlled study, but as a personal data point, it's sobering.

The Script vs. the Model: A Home Lab Reality Check

Meanwhile, over at XDA, a writer ran an experiment I found brutally honest. He replaced his home-lab automation scripts with a local LLM, then watched the whole thing fall apart.

Scripts are rigid, sure — if you want behavior to change based on storage state or service health, you're hand-coding conditionals. But the LLM version failed in ways scripts never do. Not deterministic. Sometimes it just... did the wrong thing. He ended up concluding the scripts won every single time.

There's a real lesson here beyond the entertaining failure. The people pushing "just throw an LLM at it" are selling a story where the model replaces the boring engineering. In practice, for the narrow, well-defined tasks in a home lab, deterministic code wins because it has to win — you can't debug a hallucination the same way you debug an if-statement.

I think the honest takeaway is that local AI is genuinely useful, but for different jobs than the hype suggests. Let it read your codebase and surface candidates. Don't let it replace the act of thinking about whether those candidates matter. Use it to offload the retrieval — the "what did we decide about X six months ago" — but keep the judgment in the loop where it belongs.

The Rest of the Week in AI

• TikTok's transparency center says AI now catches 97% of removed videos before they hit anyone's feed. Impressive stat, but it's worth remembering that moderation is where AI both shines and embarrasses itself — the false-positive rate is the number nobody wants to print.
• Data centers are becoming a political liability. A Vox piece this week dug into why voters in Texas and Australia are increasingly hostile to new construction, and it's not the NIMBY stereotype — it's power prices, water, and the feeling that the gains flow elsewhere. A senator in Australia is openly questioning the long-term financial benefit to the country. That conversation is only going to get louder as training runs get hungrier.

Closing Thoughts

Nobody's arguing we should stop using these tools — that ship has sailed, and honestly, I'd rather have my codebase-crunching agent than not. But the MIT study and Kapoor's framing hit something real: the metric that matters isn't just output, it's what happens to your own reasoning when the tool is removed.

Keep the habit of thinking before you prompt. It's cheap, it's free, and it's the one muscle these tools won't grow for you. Quick add-on note — if you're doing the mental math on anything in this post, Decision Calculator has been a handy sidekick for me lately.

That's it for today's pulse. If you've got a story about your own AI habits — good, bad, or weird — I'd love to hear it.