Cognitive offloading means moving part of a mental task to an external aid. Generative AI can supply drafts, explanations and suggestions, but improved output while using a tool is not the same as independent learning. Recent studies examine that distinction; they do not establish permanent brain damage or an evolutionary decline in human intelligence.
In this article
What the studies actually measured
| Study | Design and sample | Evidence and limits |
|---|---|---|
| Lee and colleagues, CHI 2025 | Survey of 319 knowledge workers, with 936 examples of GenAI use. | Self-reported effort and confidence. Higher confidence in AI was associated with less reported critical-thinking effort. The survey does not establish a causal or permanent loss of ability. |
| Kosmyna and colleagues, essay-writing preprint | 54 participants in sessions 1–3; 18 in a fourth session. Compared LLM, search-engine and unaided conditions using EEG and writing-related measures. | Task-specific differences in connectivity, recall and ownership. Version 2 was revised on 31 December 2025. EEG differences are not a diagnosis of brain damage or a population-wide forecast. |
| Bastani and colleagues, PNAS 2025 | Randomized field experiment with nearly 1,000 high-school mathematics students, comparing standard AI assistance, a guarded tutor and a control. | The standard interface improved practice performance but students performed worse after access was removed. Tutor safeguards largely avoided that harm. Findings concern this learning setting and tool design. |
Assisted performance versus independent skill
A polished AI-assisted answer can conceal whether its user can explain the reasoning, identify an error or solve a similar problem unaided. This is a practical evaluation problem: assess both the finished product and the learner’s own understanding.
The mathematics experiment found a 48% improvement during practice with the standard interface, followed by a 17% reduction in exam grades relative to the control when assistance was removed. Those are study-specific relative results, not estimates of how much intelligence an AI user loses.
How to retain critical judgment
The following are editorial practices informed by the distinction between assistance and learning, rather than a validated universal treatment:
- Attempt the problem before requesting an answer; ask for a hint when learning is the aim.
- Check important claims against original sources and inspect calculations.
- Explain the reasoning in your own words and identify the assumptions.
- Practice comparable tasks without the tool to check retention.
- Use feedback to correct mistakes instead of copying an unexplained solution.
What remains uncertain
Survey associations, short experiments and specific classroom results cannot establish an inevitable long-term decline across all users. Effects may depend on the task, prior knowledge, interface and way the tool is used. Longitudinal and replicated research is needed before making broad claims about enduring cognitive change.
Editorial update, 7 October 2026: This article now distinguishes measured task outcomes from the earlier evolutionary framing and provides the study designs and their limits.
Insider Release Declassified. Analyzed. Explained.