Who Thinks When AI Thinks? Two Ways of Working With It

You ask a chatbot to draft a reply to a client. The text looks tidy, so you send it. An hour later a colleague asks what exactly you promised — and you can’t quite remember. It isn’t carelessness. You didn’t really write that message.

Handing parts of our mental work to outside helpers is nothing new. We write shopping lists, set reminders, use calculators and follow the sat-nav. Psychologists call this cognitive offloading: using an action or a tool to reduce the demands a task places on our own memory and attention (Risko & Gilbert, 2016). Most of the time it’s sensible. The freed-up capacity goes to something that matters more.

Generative AI changes the scale. We can now offload not just memory or arithmetic, but reasoning itself: building an argument, choosing a structure, reaching a conclusion. Hence the question we hear more and more: are we forgetting how to think? Research from 2025–2026 gives a careful answer. What seems to matter is less whether you use AI and more how.

What the research shows

The most widely cited study is MIT Media Lab’s “Your Brain on ChatGPT” (Kosmyna et al., 2025). Fifty-four participants wrote essays over three sessions in one of three conditions: with ChatGPT, with a search engine, or with no help at all. The researchers recorded EEG throughout.

Those who wrote unaided showed the strongest and most widely connected brain networks. The ChatGPT group showed the weakest, with search users in between. ChatGPT users also felt the least ownership of their essays and struggled more to quote what they had just written.

It’s a vivid result, but one to hold loosely. The paper is a preprint that has not been peer reviewed, and other researchers have published methodological criticism of it. The sample is small, and only 18 people took part in the fourth session, where groups switched conditions. Weaker EEG connectivity is not a direct measure of intelligence either — it may simply reflect less effort on that particular task.

A more revealing study appeared in 2026 in Frontiers in Psychology (Zhu et al.). The authors surveyed 589 people, most of them university and graduate students, the rest young professionals, in three waves two weeks apart. Rather than asking how much people used AI, they asked how, and separated two modes:

• Dependent offloading — handing the thinking itself to AI and accepting its output with little scrutiny.

• Autonomous offloading — using AI as scaffolding to suggest, structure and check, while decisions and responsibility stay with the person.

In the moment, both modes felt equally helpful: participants rated the immediate benefits — speed, less effort — about the same. The difference showed up later. Dependent offloading was linked to what the authors call a transfer of cognitive agency — a habit of letting the machine make the call — and to lower intrinsic motivation, the interest in the work for its own sake. Autonomous offloading showed no link to handing over decisions and was associated with higher motivation.

One important caveat: every measure in this study was self-reported. People rated their own creativity, depth of processing and independent judgment; nobody tested them. So the findings describe how people see their thinking, not proven changes in ability — as the authors themselves stress.

Even so, the implication is uncomfortable: an unhealthy way of working with AI is hard to spot from your own experience. It feels just as convenient as a healthy one.

A systematic review of 67 studies on ChatGPT in higher education (2022–2025), published in Computers and Education: Artificial Intelligence, paints a similar picture. ChatGPT supported critical thinking when it was built into structured tasks with room for reflection. Unstructured use led to over-reliance and shallow engagement. Creativity followed the same pattern: helpful for generating early ideas, but without clear guidance it risked drowning out the writer’s own voice.

A possible mechanism

Why would the mode matter? Three explanations fit what we know about learning, though none has yet been proven directly for AI.

Effort. What we think through, phrase and check ourselves sticks better than what we merely read. When AI does all the heavy lifting, memory has little to hold on to — consistent with MIT participants struggling to quote their own essays.

Ownership. If the result doesn’t feel like yours, you have less reason to understand it. Losing the sense of authorship and losing intrinsic motivation may well be two sides of the same process.

Metacognition — the ability to monitor your own thinking. In the Frontiers study it helped only partly. Among people who kept closer track of how they used AI, dependent offloading was less strongly linked to handing over decisions — but that self-monitoring did not protect their motivation. Paying attention helps, but it may not be enough on its own; what seems to matter is who actually does the thinking.

It’s just as important to note what the data do not show. They don’t show that AI “rots the brain” or “makes people stupid.” Headlines like that overstate the evidence.

Where the explanation falls short

The evidence here is young, and its limits are real.

Correlation isn’t causation. Most studies are observational or short-term. If dependent users have lower motivation, the reverse may also be true: less motivated people may be quicker to hand everything to AI. Four weeks of follow-up beats a one-off survey, but it isn’t years.

Self-report isn’t performance. The strongest data on the two modes come from questionnaires. Whether people who offload dependently actually think worse — not just feel they do — still has to be tested.

Who was studied. Participants are mostly students. In the systematic review, more than half of the studies come from Asian universities, and from STEM, teacher education and language courses. How this translates to an accountant, a doctor or a retiree is an open question.

The lab isn’t life. An experimental essay isn’t real work. For routine tasks, delegating to AI can be entirely rational — nobody worries that we no longer do long division by hand.

History urges caution. In 2011 a paper in Science on the “Google effect” made headlines with the idea that the internet changes how we remember. When the Social Sciences Replication Project retested one of its key findings in 2018 — that hard trivia questions make us think of computers — the result did not replicate. That doesn’t disprove the whole idea. It does show how easily a striking claim about “technology rewiring our minds” can run ahead of the evidence.

Why it matters in practice

A better question than “should I use AI?” is “who is making the decision here?” A few habits in the spirit of autonomous offloading — not a guarantee, just a way to keep the thinking yours:

• Your version first. Sketch a draft, a hypothesis or a plan — even three lines — before you open the chatbot.

• Ask for critique, not answers. “Find the weak points in my argument” gives your brain more to do than “write me an argument.”

• The explain-it-back test. When you’re done, try to restate the result in your own words. If you can’t, the decision isn’t yours yet.

Chess has lived with this for years. Engines play far better than any world champion, and there’s one in every phone. Yet a common piece of coaching advice is to analyse your own game first — write down where you think it went wrong — and only then switch on the engine. The engine will show you the better move. It won’t tell you why you missed it. That part you have to work out yourself, and that is where a player improves.

Conclusion

AI doesn’t automatically make us smarter or duller. The evidence so far is consistent with a simpler idea: it amplifies whatever way of working we bring to it. Used as scaffolding, it leaves the thought ours. Used as a substitute, it may gradually take over not just the work, but the decisions and the interest in them.

The main risk isn’t the tool. It’s that the unhealthy mode feels just as comfortable as the healthy one. So it’s worth asking yourself, now and then: who was thinking when I got this answer?

Questions for discussion

1. What task do you happily hand to AI — and what would you never give it? Where exactly do you draw the line, and why there?

2. Think of something you recently did with AI’s help. Could you explain it without the chat open? What does that tell you?

3. If you play chess or Go: do you open the engine before or after your own analysis — and has that choice changed how you play?