# Could More AI Make Us Less Skilled? > A Trends in Cognitive Sciences article warns that cognitive offloading can weaken skill acquisition and accelerate decay, while stopping short of claiming that AI lowers basic intelligence. _Source: Trent N. Cash et al., Trends in Cognitive Sciences · 2026-07-29 · 8 min read · Verified against primary sources_ Canonical: https://iyu.app/e/ai-cognitive-offloading-risks ## The 60-second version A Trends in Cognitive Sciences synthesis argues that fully offloading thinking to AI can impede skill acquisition and accelerate skill decay, while foundational cognitive abilities may be more resilient. **Key points** - This is a review and perspective article, not a new dose-response experiment on AI use and intelligence. - AI-assisted task scores can rise even when later unaided performance falls. - Seven LLM-versus-search experiments involving 10,462 participants found shallower learning from automated syntheses. - Specific skills and knowledge appear more vulnerable than basic abilities such as working memory. - Hints, critique, retrieval practice and human verification can turn AI from a substitute into a scaffold. **Verdict.** AI is most likely to weaken what it repeatedly replaces. Keep the learner inside the reasoning loop when the goal is durable capability, and reserve full automation for tasks where learning is not the objective. ## Full explainer > **i** Evidence boundary: this is a Science & Society synthesis, not a new experiment showing that heavier AI use causes lower intelligence. It reviews evidence about cognitive offloading, skill acquisition and skill decay. The authors explicitly distinguish learned skills from foundational cognitive abilities. ### The question — What exactly could AI make worse? The paper by Trent N. Cash, Megan O. Kelly, Brooke N. Macnamara and Evan F. Risko asks a deliberately provocative question: **Is AI making us stupid?** Its more precise answer is narrower. Completely handing a task to AI can reduce the practice needed to acquire and maintain a specific skill. That is not the same as proving a broad decline in intelligence. Psychologists call the delegation of mental work to external tools **cognitive offloading**. A calculator stores arithmetic steps, navigation software carries route memory, and a generative model can now draft explanations, plans or solutions. Offloading is often useful. The risk appears when the tool removes the very mental operations a person is trying to learn. - **15** — References synthesized by the focal article - **10,462** — Participants across seven LLM-versus-search experiments - **~1,000** — Students in a field experiment on AI math tutors ### Performance vs learning — A better answer today can hide weaker performance tomorrow One cited field experiment gave nearly a thousand high-school students access to two GPT-4-based math tutors. During practice, the unrestricted system raised scores by 48%, while a guarded tutor raised them by 127%. But after AI access was removed, the unrestricted group scored 17% below students who never had AI. The guarded tutor largely mitigated that loss. > **⚑ Caveat:** These percentages belong to one high-school mathematics experiment. They do not establish that every AI user learns less, nor do they measure intelligence. They show why assisted task performance and unaided learning must be evaluated separately. The mechanism is familiar: learning requires retrieval, error correction, synthesis and repeated practice. A system that supplies the finished route can improve immediate output while bypassing those operations. The user may also develop an illusion of competence because the joint human-AI product looks better than the user's unaided skill. ### Depth of knowledge — Summaries can remove productive discovery A second cited program of seven online and laboratory experiments compared learning from LLM syntheses with traditional web search while holding core facts constant. Across 10,462 participants, LLM users developed shallower knowledge, felt less invested in the resulting advice, and produced advice that was sparser and less original. The relevant variable was not simply whether the facts were correct. Search required people to inspect sources and build a synthesis; the automated summary delivered that synthesis. Convenience reduced the amount of constructive work performed by the learner. > AI can improve the product without improving the producer. ### What may decay — Specific skills are more vulnerable than basic capacity - **Specific learned skills:** Algebra procedures, diagnostic search, writing structure and source synthesis depend on practice and can weaken through disuse. - **Knowledge:** Facts and conceptual links may be encoded less deeply when the tool performs retrieval and synthesis. - **Metacognition:** Users may overestimate what they can do unaided or fail to notice skill loss. - **Basic cognitive abilities:** Working memory, attention and general reasoning capacity are more resistant to task-specific training or disuse; evidence does not show broad collapse. This distinction matters. Losing fluency in long division after years of calculator use is plausible; losing the foundational ability to reason because one used a calculator is a much stronger claim. The authors argue that basic capacities may be comparatively resilient, even while knowledge and practiced skills erode. ### Design matters — A coach and a crutch are not the same tool The review is not an argument to avoid AI. The math experiment shows that guardrails can preserve much of the benefit: a tutor can offer hints, ask for intermediate steps and withhold final answers. Other work suggests collaborative use can support learning when the human remains responsible for generating, checking and revising ideas. - **Attempt first.** Produce an outline, calculation or diagnosis before asking AI to intervene. - **Request critique, not replacement.** Ask for errors, counterarguments and missing evidence instead of a finished answer. - **Retrieve without the tool.** After an AI-assisted session, explain the result from memory or solve a nearby problem alone. - **Keep provenance.** Inspect primary sources and mark which claims came from the model. - **Use tutor constraints.** Prefer hints and questions when the goal is learning; use full automation when the goal is only throughput. ### What remains unknown — Long-term exposure has not been settled The evidence base is still young and heterogeneous. Many studies examine short tasks, students or particular interfaces. Longitudinal effects, childhood exposure, domain expertise, source-monitoring errors and the possibility that users adapt their behavior over time remain open questions. The defensible conclusion is therefore conditional: intensive offloading can weaken the specific skills and knowledge that are no longer practiced, especially when AI gives direct answers. Current evidence does not justify saying that using more AI simply lowers intelligence. The practical question is not only how often AI is used, but **which cognitive steps it replaces**. ## Primary sources - [Trends in Cognitive Sciences paper](https://doi.org/10.1016/j.tics.2026.06.004) - [PubMed record](https://pubmed.ncbi.nlm.nih.gov/42425850/) - [LLMs versus web search experiments](https://doi.org/10.1093/pnasnexus/pgaf316) - [High-school mathematics field experiment](https://doi.org/10.1073/pnas.2422633122) - [AI assistance and skill decay perspective](https://doi.org/10.1186/s41235-024-00572-8) --- _Published by iyu (https://iyu.app) — the day's AI news, checked against primary sources and rewritten in plain language. Free to quote with attribution and a link to the canonical URL._