Critical Thinking in the Age of AI: Separating Performance Gains from Learning

Generative AI can boost learners' immediate performance, but such performance gains are not the same as learning. The Nature Reviews Psychology comment explicitly distinguishes performance gains from learning, stating that AI uses "do not promote the deep cognitive and metacognitive processing that are required for high-quality learning", shifting the educational-technology discussion from score gains toward the quality of cognitive processing. The comment is a commentary/review piece; its full text is subscription-only, and the evidence bundle contains only summary-level statements and a reference list, with no experimental data.

Existing studies suggest AI assistance may bring cognitive ease at the cost of depth. The comment cites work including Stadler et al. on "LLMs reduce mental effort but compromise depth in student scientific inquiry" and Fan et al. on "metacognitive laziness", making the impact of AI on learning processes concrete in terms of reduced mental effort and metacognitive engagement. The evidence bundle lists only bibliographic records for these works, without sample sizes, effect sizes, or experimental design details.

Research on whether AI erodes critical thinking is not yet consistent. The feature notes that "research does not yet show clearly what impact generative AI is having on learning and critical thinking", because the tools are new and varied, studies conflict, and few rigorous or long-term studies exist; it also cites a 2024 study in which secondary-school students solved mathematics problems better with a ChatGPT-like tool but performed worse than those who never had access once the tool was removed. The feature relays the direction of that 2024 study's findings, but the evidence bundle provides no sample size, effect size, or statistical detail.

Critical thinking is domain-specific, so teaching it must happen within specific subjects. Educational psychologist Paul Kirschner states that critical thinking is "not generic and free-floating — you have to teach it within specific subjects", illustrating this with corroborating records and artefacts in history versus understanding the relationships between theory, hypothesis, method and evidence in science; the feature also cites a study showing neurologists are not particularly good at diagnosing cardiology cases. This is a relay of expert opinion and prior research; the evidence bundle does not provide the sample or design of that study.

This work speaks to education researchers, teachers and curriculum designers: it suggests that when generative AI is introduced, tasks should be structured to require learners to reason themselves, interpret evidence and grapple with uncertainty rather than receive ready-made answers; domain specificity means such practices must be redesigned within each subject rather than applied as a generic recipe. The feature also mentions OECD work with Macat to develop a standardized critical-thinking assessment for secondary-school students and undergraduates, and an EEF commitment of up to £2.5 million for rigorous studies, both of which are directions to watch.

Readers should still watch: the long-term impact of generative AI on learning and critical thinking has no consistent conclusion, and the feature itself notes studies conflict and rigorous long-term research is scarce; the comment's full text is subscription-only, so the bundle shows only the abstract-level statement and references, leaving its detailed argument and data unverifiable; the survey and experimental figures cited in the feature are relayed rather than reported with sample sizes and effect sizes; and how to measure critical thinking in a standardized way remains an open question.