The short answer
Current research does not show that using AI inevitably or permanently reduces critical-thinking ability. It does show that some workflows lower active engagement, increase acceptance of incorrect advice, or create performance that does not transfer when assistance is removed; other workflows support learning and better performance. The task, user, interface, and division of mental work all matter.
The evidence
The honest answer is conditional, not catastrophic.
AI can reduce critical-thinking effort within a task when it supplies the frame, analysis, and conclusion before the person engages. That is not the same as demonstrating a permanent reduction in critical-thinking ability or intelligence.
A 2025 CHI study surveyed 319 knowledge workers about 936 real examples of generative-AI use. Higher confidence in AI was associated with less self-reported critical-thinking activity, while higher task-specific self-confidence was associated with more. Participants also described thinking shifting toward verification, response integration, and task stewardship.
The study is informative but should not be overstated. It relied on self-reports and observed associations rather than directly testing long-term skill change. It supports a concern about how effort is allocated; it does not establish that AI inevitably makes people less capable.
Ask which thinking moved, which thinking vanished, and which thinking improved—not simply whether AI was present.
A mixed record
AI can replace engagement, but it can also distribute expertise.
Experimental studies establish a narrower risk: people sometimes accept incorrect AI recommendations, and deliberate cognitive friction can reduce that behavior. Merely adding an explanation does not reliably improve judgment because the explanation may make the recommendation more persuasive without making it more accurate.
Workplace evidence is not uniformly negative. A published study following 5,172 customer-support agents found an average productivity increase of roughly 15 percent and evidence consistent with learning, especially among less experienced workers, in that specific setting. AI can transmit useful patterns and support skill development when the workflow exposes people to applicable knowledge.
An OECD review of experimental evidence concludes that outcomes depend on the task, user experience, and type of human-AI collaboration. Long-term business effects and workers' understanding of model limitations remain important evidence gaps.
| Supported with qualification | Not established |
|---|---|
| Some AI workflows are associated with lower self-reported critical-thinking effort. | AI use universally or permanently reduces intelligence. |
| People can over-rely on incorrect recommendations in controlled tasks. | Every person will become dependent on AI. |
| Interaction design and deliberate friction can change reliance behavior. | One checklist eliminates overreliance in every domain. |
| AI has produced productivity and learning benefits in some workplaces. | Productivity gains always represent durable human learning. |
Use the distinction
Offload production—not the judgment that makes production useful.
People have always used tools to offload memory, calculation, and routine execution. A checklist, calculator, map, or spreadsheet can free attention for higher-value thought. Cognitive offloading is not automatically harmful.
The risk is offloading the mental work that creates transferable judgment. If AI decides what the real problem is, which sources deserve trust, which assumptions matter, and what downside is acceptable, the person may produce an answer without exercising the capabilities needed to evaluate it.
- Keep the larger goal and problem frame human-owned.
- Separate observations, reports, interpretations, predictions, values, and assumptions.
- Judge source independence and identify missing information.
- Choose what a good outcome means and which downside is acceptable.
- Make the commitment and return to learn from the result.
Use AI deliberately
Give AI four productive roles—and retire the role of oracle.
The same model can weaken or strengthen a decision depending on the assignment. Prompts that request a verdict encourage delegation. Prompts that request alternatives, tests, and competing explanations create material a person can evaluate.
| Role | Useful request | Human responsibility |
|---|---|---|
| Generator | Produce materially different options, not five phrasings of one idea. | Test feasibility and keep only genuine alternatives. |
| Critic | Find counterarguments, weak assumptions, disconfirming evidence, and failure mechanisms. | Verify the critique and judge which risks matter. |
| Simulator | Show how the case could look through another incentive, stakeholder, time horizon, or system layer. | Treat simulated perspectives as hypotheses, never as facts about a person. |
| Organizer | Structure notes and identify unresolved questions. | Preserve provenance and distinctions between evidence, inference, and assumption. |
| Oracle | Tell me what to do. | Retire this role; a generated verdict cannot own the objective or consequences. |
A repeatable practice
Think before, evaluate during, and learn after.
Before consulting AI, describe the decision in your own words. Record the larger goal, current evidence, initial options, important unknowns, forecast, and confidence. This baseline is not a claim of human superiority; it is a reference point that makes influence observable.
During AI use, request alternatives and challenges rather than a final recommendation. Trace material factual claims to suitable sources, ask what may be missing from the prompt, and identify the assumption that carries the conclusion.
After AI use, state the final reasoning in your own words. Record what changed, why it changed, what would change your mind again, and what result you expect. When the outcome arrives, review the quality of the process separately from luck.
If you cannot explain the final rationale without reopening the chat, the thinking is not yet ready to become a commitment.
The product connection
Build a thinking practice that remains available without AI.
Give AI a challenge to generate—and yourself a question to investigate. Reality Skill is a decision-making workspace in which evidence, interpretations, assumptions, alternatives, forecasts, and AI contributions remain distinguishable. Decision lenses help you choose a relevant test instead of treating an articulate response as proof.
Set a Reality Date to compare expectations with outcomes. That is real-world decision practice: keeping the reasoning available and returning to what happened. The aim is to develop judgment through use and reflection; the platform itself has not been shown to guarantee improved critical-thinking ability.
Research and sources
Evidence behind this guide.
- The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers Microsoft Research / CHI 2025
Based on self-reported examples and associations; it does not directly measure longitudinal loss of skill.
- To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI Proceedings of the ACM on Human-Computer Interaction
A controlled decision-task experiment; observed effects and usability tradeoffs are context-specific.
- Generative AI at Work The Quarterly Journal of Economics
A staggered workplace deployment among customer-support agents; results should not be generalized to every occupation.
- The Effects of Generative AI on Productivity, Innovation and Entrepreneurship OECD
A 2025 review of experimental evidence that emphasizes variation by task, experience, and implementation as well as long-term evidence gaps.
- Appropriate Reliance on Generative AI: Research Synthesis Microsoft Research
Distinguishes appropriate reliance from both overreliance and under-reliance.