Field notes: the mental overload caused by intensive generative AI use

I’ve been working with AI assistants every day for three years. Code, writing, research, note-taking — I have at least one open continuously for half my working hours. In the autumn of 2024 I had, for the first time, the distinct feeling that I was no longer thinking on my own. Not a grand existential revelation: a precise, measurable fatigue that built up at every session.

This article is a structured field report. I describe what I observed, line it up with the literature that is starting to emerge on the topic, and lay out the concrete rules that helped me recover a thinking style of my own — without giving up on AI, which remains powerful for specific uses.

The three symptoms I observed

1. Decision fatigue

Working with an assistant means arbitrating non-stop. Does this suggestion go in the right direction? Does this rewrite preserve my voice? Is this phrase actually mine? Every proposal consumes a decision cycle. After three hours, I finish more drained than after three hours of pure writing — yet I’ve produced fewer original sentences.

This cognitive cost is what the literature calls decision fatigue, formalized by Roy Baumeister and colleagues (Vohs et al., Making Choices Impairs Subsequent Self-Control, JPSP, 2008). The more micro-decisions you make, the lower the quality of subsequent decisions.

2. The atrophy of initial thinking

Without realising it, I had stopped starting. Facing a blank document, my first reflex was no longer to drop an imperfect sentence on the page — it was to open an assistant and ask for a draft. The mental kickoff, the part that demands the most creative energy, I had outsourced.

This behavior has a name: cognitive offloading. The term comes from Sparrow, Liu and Wegner (Google Effects on Memory, Science, 2011), who showed that the mere availability of an external search engine modifies how people encode information. The most thorough synthesis is Risko & Gilbert, Cognitive Offloading, Trends in Cognitive Sciences, 2016: offloading a cognitive task to an external tool reduces subsequent performance on the same task without the tool.

3. Dependency rooted in blank-page anxiety

The most unexpected symptom is that I became less able to write alone. Not from technical incapacity — from anxiety. Facing an empty file, with no assistant in reach, I had an almost physical stress response. As if a safety net had vanished.

This matches the findings of the MIT Media Lab study published in June 2025 by Nataliya Kosmyna and colleagues, “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing” (arXiv preprint, 2025). The researchers compared 54 participants split into three groups — writing unassisted, with a search engine, with ChatGPT — under EEG. Results:

• The ChatGPT group shows the lowest brain connectivity during writing.
• 83 % of ChatGPT-group participants could not quote a sentence they had just “written” minutes earlier.
• When the ChatGPT group was then asked to write without an assistant, their neural engagement remained lower than that of the other two groups.

The authors call this a “cognitive debt”: a mental cost you defer but ultimately pay.

The hidden cost: erosion of critical judgment

Microsoft Research and Carnegie Mellon published a complementary study at CHI 2025: “The Impact of Generative AI on Critical Thinking” by Hao-Ping (Hank) Lee, Lev Tankelevitch and colleagues (Lee et al., CHI 2025, Microsoft Research PDF). 319 knowledge workers reported their weekly AI usage.

The findings are sharp:

• The higher the trust in AI, the less the user reports engaging in critical thinking (significant negative correlation).
• Conversely, the higher the trust in one’s own skills, the more the user critically evaluates AI outputs.
• Self-reported cognitive effort shifts: from production (original work) to verification (checking outputs).

In other words: people don’t think less because they are lazy. They think less because they delegate the generative part and become reviewers — a less demanding, but also less formative, role.

Why writing is particularly vulnerable

Writing is one of the few cognitive acts where formulation is thinking. As Joan Didion put it in Why I Write: “I write entirely to find out what I’m thinking, what I’m looking at, what I see and what it means.”

When you ask an assistant to produce the first draft, you skip exactly the step that structures the thought. The text that comes back may be good, clean, readable — it no longer carries the trace of your clarification. You re-read it; you didn’t think it.

This isn’t an anti-AI stance. It’s an anti-AI-for-everything stance. The uses where an assistant excels (summarizing long documents, transforming formats, initial brainstorming, error catching) are different from the uses where it impoverishes (substantive writing, taking a position, building an original argument).

The rules that helped me recover

Here’s what I put in place after six months of experimenting. Nothing is rigid: it’s what works for me, and it lines up with what the studies suggest.

Rule 1 — Alternate without rationalizing

I impose one hour of unassisted production for every hour of AI-assisted production. The rule is hard because the temptation is constant: a single prompt, just to unblock one sentence, and the session tips over. To hold the line, I use a deliberately silent tool — Draft_, the editor I build on the side, exactly for this reason: it suggests nothing, completes nothing. The friction isn’t a flaw; it’s the feature.

Rule 2 — Put AI downstream, never upstream

AI does not write my first draft. It can:

• Re-read a draft and flag inconsistencies.
• Reformulate a sentence when I’m stuck on a specific phrasing.
• Verify a reference, a date, a calculation.

But never: “write me the introduction”. The test is easy: if the AI output looks cleaner than what you would have written, it has also done the thinking for you.

Rule 3 — Keep one daily session free of any smart tool

Every morning, one hour of pure writing. No assistant, no search engine, no autocorrect. The text produced is less polished, but it’s the only time I know what’s on the screen comes entirely from me. This session serves as a baseline: if it becomes too painful, that’s the signal that my cognitive debt is climbing.

Rule 4 — Evaluate every tool on “am I thinking when I use it?”

Not all digital tools are equal. An assistant that answers a specific question still lets me do the integration work. An assistant that completes my sentences in real time substitutes itself for my formulation process. The second is significantly more cognitively expensive, independent of output quality.

What I did not observe

In fairness, here’s also what I did not experience:

• No loss of technical skill. I code as well without Copilot as with — perhaps a bit slower, but not less correctly.
• No drop in final quality of AI-assisted texts. With serious re-reading and correction, the result holds.
• No difficulty restarting. Three weeks of AI-free writing were enough to recover the initial fluency.

So the issue isn’t AI itself. It’s intensive, unreflective use. Tools are neutral. Routines are not.

In short

• Intensive AI-assistant use produces a specific mental load: decision fatigue, atrophy of starting, anxiety-driven dependency.
• Two recent studies document these effects: MIT Media Lab 2025 (reduced brain connectivity, cognitive debt) and Microsoft/CMU 2025 (negative correlation between AI trust and critical thinking).
• Writing is especially vulnerable because formulation is thought.
• The rule that works: alternate. One AI-free hour for every assisted hour, AI always downstream, and a deliberately silent tool for the daily baseline session.

If you want to try a session without an assistant, Draft_ is built exactly for that: no AI, no suggestions, no autocomplete. Silence returns, thought unfolds.