Now I question everything David Epstein says

Even if you happen to intuitively agree with the conclusions, this the kind of dangerous and uninformed statement from a bestselling author that spreads like wildfire and becomes impossible to contain. The "new" study from MIT appeared as a preprint in June 2025 and has not yet been published, so using it as "settled law" is at the very least flippant. The preprint did get a lot of press because, well, MIT Media Lab is very good at it and because it reinforces the general feeling people have about AI and how it affects our brains. The authors themselves acknowledged they released it early to gather feedback, and asked journalists not to use words like "stupid," "brain damage," "harm," or "brain rot", language that nevertheless dominated coverage because the authors did come up with catchy expressions like "cognitive debt" or "brain first, AI second". When you ask journalists to cover your half-baked but bombshell-level work with moderation, what do you expect? Even though it is still a preprint, some scientists felt compelled to chime in to counter the deluge of press (link in the comments). Here are some of the issues (many more have been raised): 1️⃣ The study had 4 "sessions." Sessions 1–3 had 54 participants split across three groups (~18 per group). Session 4, which produced the most-cited findings about "cognitive debt" lingering after LLM use, was optional, and only 18 participants returned (about 9 per group). Michael Demidenko, PhD points out that the headline result rests on a small, self-selected subsample, with no analysis of who came back or why. Low statistical power makes a "significant" result much less likely to reflect a real effect. 2️⃣ The EEG data interpretation is a joke, reverse-inference-style. The paper assumes higher neural connectivity = better/deeper cognition, and lower connectivity = under-engagement or "cognitive debt." But connectivity differences could equally mean greater efficiency, different strategies, or simply different task demands not better or worse cognition. 3️⃣ The phrase "accumulation of cognitive debt" implies a causal trajectory, but the design can only show group differences in a constrained task. There was no measurement of tool use outside the lab, and participants in the LLM group reportedly varied widely in how much they actually used ChatGPT (one used it only for proofreading, others wholesale). The authors were tackling a timely topic and collecting a rich multimodal dataset, but their conclusions are beyond weak. We need real evidence, not casual correlational statements packaged as truth, even (actually, even more so) if they fit your narrative.