This article summarises reporting from The Algorithmic Bridge. Alberto Romero compiled and wrote the underlying analysis; this piece credits and describes his work.
Alberto Romero, writing in The Algorithmic Bridge, has compiled more than 30 studies published between 2023 and 2026 on how AI chatbots affect human cognition, learning, and psychology. The compilation draws on work from MIT, Wharton, Harvard, Stanford, Microsoft, OpenAI, Oxford, Google DeepMind, and several Chinese universities. Romero states that, to his knowledge, no prior single source had assembled these studies in a readable and accessible form.
Brain activity studies
Romero organises the literature into several clusters. The first covers neuroimaging. A 2025 MIT Media Lab study by Kosmyna et al. (arXiv preprint, N=54) used 32-channel EEG to track brain activity across four sessions. The group using ChatGPT showed “the weakest neural connectivity,” up to 55% lower than unaided writers, and grew reliant on copy-paste by session three. When switched to writing alone in session four, their brain activity remained suppressed; the researchers called this an accumulation of “cognitive debt.” A separate study (Horowitz-Kraus et al., bioRxiv preprint, 2025, N=31) scanned children aged 6–7 alongside adults using fMRI during chatbot interaction. Children showed “lower engagement of cognitive control, attention, and modulation networks,” a pattern the researchers described as more pronounced than in adults.
A counterpoint in the neuroimaging cluster: a study by Wang et al. (Frontiers in Psychology, 2025, N=64) found design students using AI creative tools showed “significantly higher concentration levels” and higher creative performance than a control group. Romero notes the distinction: these students were actively directing AI rather than passively receiving answers.
Critical thinking and automation bias
The second cluster covers what Romero calls “cognitive surrender.” A Wharton working paper by Shaw and Nave (SSRN, 2026, N=1,372) ran three preregistered experiments where AI was sometimes programmed to give wrong answers. Participants followed the incorrect AI output roughly 80% of the time, performing “worse than having no AI at all.” High trust in AI was the strongest predictor: high-trust participants had 3.5 times greater odds of following a faulty answer.
A Microsoft and Carnegie Mellon survey of 319 knowledge workers who use generative AI weekly (Lee et al., CHI ‘25, 2025, collecting 936 real-world use cases) found that “higher confidence in GenAI correlated with less critical thinking.” Workers shifted from active problem-solving to passive output selection, producing “a less diverse set of outcomes” when using AI.
A preregistered field experiment with 758 BCG consultants (Dell’Acqua et al., Organization Science, 2026) found that for tasks within AI’s capability range, consultants completed 12.2% more tasks, worked 25.1% faster, and produced 40% higher quality results. For tasks outside that range, AI users were “19 percentage points less likely to produce correct solutions.” The problem, Romero reports, was that AI output outside its frontier “looked polished but was subtly wrong.”
Overall pattern
Romero describes the cross-study finding as a paradox: passive AI use — receiving answers — suppresses the brain regions involved in effortful thinking, while active AI use — directing tools, receiving challenges — can maintain or increase engagement. He writes that “the variable is not the presence of AI but rather what the AI asks your brain to do.”
Romero acknowledges the limits of the literature: small samples in neuroimaging studies, several results still at preprint stage, and at least one contradictory finding. The article is the sole source for this piece.