Learning versus outsourcing

Learning time on AI-susceptible problems declines 2.8% per quarter among college students after ChatGPT's release, cumulating to 26.9% over eleven quarters.

Large-scale observational evidence of exactly the substitution the question is about: proctored retention drops ~25% while non-proctored performance rises — AI raising measured output while eroding the effortful practice that builds knowledge.

unrestricted access significantly outperforms restricted access by 0.21 standard deviations, contradicting concerns about premature AI reliance.

An experimental counterpoint: unrestricted AI access beats restricted access and does not crowd out textbook reading, with gains concentrating among lower-baseline students who self-regulate — the constructive side of the split.

Our primary concerns focus on: (i) study design considerations, including the limited sample size; (ii) the reproducibility of the analyses; (iii) methodological issues related to the EEG analysis

The serious peer rebuttal to the viral MIT 'cognitive debt' EEG study, cataloguing sample-size, reproducibility, and EEG-analysis problems — the reference a reader tracking the cognitive-debt claim needs alongside the original.

Some studies have found that chatbot tutors can backfire because students lean on them too heavily, get spoonfed solutions and fail to absorb the material.

Reports a UPenn RCT with ~800 students showing that how a tutor sequences difficulty determines whether it deepens learning or spoonfeeds — the design question at the heart of learning-vs-outsourcing.