This piece first appeared in German. The English version is a rewrite rather than a line-by-line translation, and the German original stays online in the archive: Outsourcing Minds: Wie KI den Denkprozess leise verändert.

A team at the MIT Media Lab around Nataliya Kosmyna measured the brain activity of 54 people writing essays under three conditions: with ChatGPT, with a search engine, and with no digital help at all. Electroencephalography recorded their brain activity at 32 electrode sites during the writing. The ChatGPT group showed the weakest connectivity between brain regions of the three. In the first session, 15 of its 18 members could not correctly quote a sentence from the essay they had written minutes before.

For most of human history thinking was unavoidable. Writing meant using your memory. Solving a problem meant testing ideas, sitting with contradictions, tolerating uncertainty. There was no shortcut, and your own mind was where the work happened. For millions of people it is now the last place they turn.

Generative AI became an alternative cognitive system

It stopped being a tool. What it delivers is a paragraph, an argument or a strategy, long before your own thinking has started, and the experience is seductive: faster, smoother. Behind the seduction sits a quiet substitution, a relocation of the place where mental effort begins.

The name that emerged for this among users, developers and critics is the Midjourney mentality. It describes the cultural practice of outsourcing the early, high-friction phases of thinking. The subject reaches past technology, to the gradual forgetting of how a productive train of thought even starts, because it can be generated.

The habituation is the mechanism

The same theme recurs in conversations with teachers, consultants and product managers across Europe and the United States. Students submit generated essays within seconds. Consultants draft client strategies by prompt, where they once ran the analysis. Executives rely on AI summaries and leave the reports unread.

What connects those stories is neither cheating nor laziness. It is habituation. Once you know a system can produce a first idea in a fraction of the time, your own impulse to generate one weakens. You stop warming up the engine, because something else is already in the driveway and starts instantly.

That habituation is the core of the pattern. It begins harmlessly, as brainstorming a little. Then it becomes the default, where AI does the first draft. Finally it becomes structural: there is no point thinking this through myself.

What the research shows, and what it does not

A caveat belongs directly beside the MIT result. The study is a preprint and has not been peer reviewed, and its sample consisted of students and young researchers from five universities around Boston, which is not representative of anything broader. A commentary by Milos Stankovic and colleagues also questions the small sample, the EEG analysis and the transparency of the reporting. Treat the results as a first indication and not as proof.

Field evidence comes from a CHI 2025 survey by Hao-Ping Lee and colleagues at Microsoft Research and Carnegie Mellon University. It collected reports from 319 knowledge workers across 936 documented cases of AI use. Its central finding: the more confidence participants had in the AI's ability on a task, the less critical thinking they applied to it. The researchers describe a shift from active task execution to passive task stewardship, where somebody still checks and no longer thinks.

The authors recall an irony that the psychologist Lisanne Bainbridge described for automation in 1983. Automation removes exactly the routine occasions on which judgment gets trained, and leaves the user unprepared when the exception arrives.

Standardized minds in a standardized ecosystem

Thinking is not the only thing getting flatter. The outputs get more uniform too.

Large language models work by predicting statistically plausible continuations of text, which is what makes them capable and what limits them. These are machines of probability and not of originality. When millions of people use the same models to find ideas, cultural output narrows. Writers describe a recognizable model voice, and designers talk about a particular aesthetic. Teachers see essays that look interchangeable worldwide: perfect grammar, no insight.

Emily Bender and colleagues described the underlying problem in 2021 in their paper on stochastic parrots. Language models produce the appearance of understanding by reproducing patterns from training data, without understanding anything. The analysis was contested and has held up as an analytical instrument: what looks like intelligence is in many cases statistical plausibility. Once statistical plausibility becomes the standard tool for producing ideas, the output normalizes.

The Microsoft and Carnegie Mellon study cites a related effect from earlier research, mechanized convergence: people with access to AI produce less diverse work than people without it.

So the Midjourney mentality does more than outsource thinking. It standardizes it.

Where this becomes a political question

The consequences reach past education and creativity into the preconditions for democratic opinion-forming. A society that routinely outsources cognitive work makes itself vulnerable in three places.

The first is epistemic dependency. When people stop forming their own hypotheses before querying a system, they implicitly adopt the framing the system supplies. In a 2025 study of 666 people, Michael Gerlich found that younger participants relied more heavily on AI tools and scored lower on critical thinking. That is precisely the group that will make electoral decisions and shape public debate.

The second is concentrated control over the starting points of thought. When the first idea on a wide range of tasks regularly comes from a system built by a small number of companies, a new form of agenda-setting appears, through model behavior instead of media selection. Which perspectives a model prioritizes, which it omits, which it rates as plausible: all of that shapes the thinking space of its users.

The third is the asymmetry between fast answers and slow thought. AI delivers in seconds and independent checking takes minutes to hours. That asymmetry pushes toward simple narratives and away from differentiated judgment, in the way social media dynamics already do.

None of those risks is inevitable. All three get more likely the deeper the habit settles in.

Better design beats abstinence

If the problem is cognitive dependency, the answer is design and not abstinence, and there is empirical ground for that now.

Ian Drosos, Advait Sarkar and colleagues at Microsoft Research tested provocations: short texts attached to AI suggestions that criticize them and propose alternatives. In their experiment with 24 participants, the provocations prompted critical and metacognitive thinking about what the AI had suggested.

The Tools for Thought workshop at CHI 2025 took the question further, with 34 accepted contributions on how generative AI systems could protect and augment human thinking. Reflection questions, source attribution and alternative perspectives are among the forms such friction can take.

The idea now has a theoretical frame as well. Kuangzhe Xu and colleagues propose scaffolded AI friction as a design principle: systems that deliberately expose disagreement, for instance through agents acting as devil's advocates, in place of frictionless answers.

All of it points one way. The next generation of AI tools has to be built differently, and not only made more capable, so that it amplifies human cognition where it currently replaces it. That puts a responsibility on developers which reaches past interface design. What the situation calls for is AI that asks more of the person using it, and is built so that it actually does.

A culture that knows how to begin

The Midjourney mentality is a habit and not a fate, and habits change.

Changing this one takes more than individual discipline. It takes educational institutions that treat independent thinking as the core competence it is, where they currently smile at it as inefficient. Companies have to measure decision quality and not only decision speed. And it takes developers who understand that their product is most valuable when it leaves the user smarter than before, and not only faster.

The human brain keeps its capacity for depth, originality and insight. Whether those capacities keep getting used, or wither in the shadow of an algorithm that finishes the thought before anyone has started it, is the open question.

Sources

Grouped by the section they support.

Opening

  • Kosmyna, Hauptmann, Yuan, Situ et al. (2025), Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task, preprint, MIT Media Lab, arXiv:2506.08872. Supports the 54 participants, the three conditions, the EEG headset with 32 electrode sites, the weakest connectivity in the ChatGPT group, and the 15 of 18 participants who could not quote their own essay in the first session.
  • The paragraph on thinking before AI is the author's own.

Generative AI became an alternative cognitive system

  • No external source. The term Midjourney mentality and its description are the author's own framing.

The habituation is the mechanism

  • No external source. The examples come from the author's conversations with teachers, consultants and product managers.

What the research shows, and what it does not

  • Kosmyna et al. (2025). Supports the preprint status and the recruitment from MIT, Wellesley, Harvard, Tufts and Northeastern.
  • Stankovic, Hirche, Kollatzsch and Doetsch (2025), Comment on: Your Brain on ChatGPT, arXiv:2601.00856. Supports the criticism of the sample size, the EEG analysis and the transparency of the reporting.
  • Lee, Sarkar, Tankelevitch, Drosos, Rintel, Banks and Wilson (2025), The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers, CHI 2025, doi:10.1145/3706598.3713778. Supports the 319 knowledge workers and 936 examples, the link between confidence in AI and less critical thinking, and the shift from task execution to task stewardship.
  • Bainbridge (1983), Ironies of Automation, Automatica 19(6), 775–779, doi:10.1016/0005-1098(83)90046-8, as cited by Lee et al. Supports the irony of automation.

Standardized minds in a standardized ecosystem

  • Bender, Gebru, McMillan-Major and Shmitchell (2021), On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, FAccT 2021, doi:10.1145/3442188.3445922. Supports the stochastic parrots argument.
  • Lee et al. (2025), CHI 2025. Supports mechanized convergence as reported from earlier research.
  • The observations from writers, designers and teachers are the author's own.

Where this becomes a political question

  • Gerlich (2025), AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking, Societies 15(1), 6, doi:10.3390/soc15010006. Supports the 666 participants and the higher dependence and lower critical thinking scores of younger participants.
  • The three risks are the author's own analysis.

Better design beats abstinence

  • Drosos, Sarkar, Xu and Toronto (2025), "It makes you think": Provocations Help Restore Critical Thinking to AI-Assisted Knowledge Work, arXiv:2501.17247. Supports provocations, the 24 participants and the critical and metacognitive thinking they prompted.
  • Microsoft Research (2025), Tools for Thought: Research and Design for Understanding, Protecting, and Augmenting Human Cognition with Generative AI, CHI 2025 workshop, microsoft.com. Supports the workshop, its theme and the 34 accepted contributions.
  • Xu, Shen, Yan and Ren (2026), Cognitive Agency Surrender: Defending Epistemic Sovereignty via Scaffolded AI Friction, arXiv:2603.21735. Supports scaffolded friction as a design principle and the use of agents as devil's advocates.

A culture that knows how to begin

  • No external source. The conclusion is the author's own.
The link has been copied!