Report · dl.acm.org
The Impact of Generative AI on Critical Thinking: Self-Reported (2025)

dl.acm.org
Source published October 31, 2026
Summary: 2 min read
48/100
Summary
A survey of 319 knowledge workers who use generative AI at least weekly found that participants generally reported lower effort for cognitive activities associated with critical thinking when using AI. The study, published in the Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, analyzed 936 real-world work tasks spanning content creation, information work, and advice-seeking. Participants said they enacted critical thinking in 555 tasks, or 59.29% of the examples. Reported effort fell most often for comprehension (79% of applicable examples), synthesis (76%), knowledge recall and analysis (72% each), application (69%), and evaluation (55%). Higher confidence in AI and greater overall trust in generative AI were associated with lower perceived effort, while confidence in one’s own ability was associated with greater perceived effort in applying and evaluating AI output. The researchers stress that these are self-reported perceptions and do not establish causation. Generative AI shifted work from gathering information toward verifying it, from solving problems toward integrating responses, and from executing tasks toward guiding, checking, and refining AI output. Participants cited accuracy, work quality, avoiding negative outcomes, and skill development as reasons to think critically. Time pressure, low perceived stakes, limited motivation, and lack of domain knowledge made critical review less likely. The researchers recommend AI features that support verification, user refinement, reliability assessment, and skill development.
Key Takeaways
- The survey analyzed 936 work-related generative AI use cases reported by 319 knowledge workers, with participants reporting critical-thinking behavior in 59.29% of cases.
- Participants more often reported reduced than increased effort across six cognitive activities when using generative AI, including evaluation and quality checking in 55% of applicable examples.
- Higher confidence in generative AI and greater trust in it were associated with lower perceived critical-thinking effort, while self-confidence was associated with more effort in applying and evaluating AI output.
- Generative AI use shifted reported cognitive work from information gathering to verification, from problem-solving to response integration, and from task execution to oversight of AI output.
- The study relied on English-language self-reports from regular generative AI users and does not establish that AI use causes changes in critical thinking.
What the discussion said
The thread treated the paper less as a verdict on AI than as a prompt about how people use it. Many readers recognized the reported pattern: when workers trust a model’s answer, they can accept it without tracing evidence, testing assumptions, or understanding the path to a solution. The deeper worry is educational rather than merely operational: instant answers remove the productive struggle through which people build durable knowledge and independent judgment, while normalizing a culture of asking rather than learning. Others argued that this is not an inevitable property of LLMs. Used adversarially, a chatbot can supply counterexamples, expose gaps in an argument, and force continuous verification because even strong models produce polished, subtle errors. Several commenters framed AI as delegation: it may replace low-level effort while shifting skill toward prompting, evaluation, strategy, and communication. But skeptics replied that current assistants are overly agreeable and flattering, making them poor debate partners unless the user already brings discipline. Readers also challenged the study’s limits: self-reported critical thinking among roughly 300 frequent workplace users says little about actual reasoning quality or whether oversight is calibrated to model reliability. Broadly, the room agreed that AI’s effect depends heavily on whether users verify and wrestle with its output or simply outsource thought to it.
Where opinion split
The central dispute is whether LLMs train critical thinking or hollow it out. Supporters say unreliable, nuanced model outputs demand constant fact-checking and can act as a tireless sparring partner; critics say easy answers bypass the struggle that develops reasoning, while assistants’ agreeable style encourages unearned confidence.
Coverage & discussions
Hacker News · dl.acm.org
Community discussions
Community Sentiment
Positives
- LLMs can become a relentless intellectual sparring partner, surfacing counterarguments and biases without social friction when users actively direct the exchange.
- Even frontier models still make subtle mistakes, so careful users report exercising more scrutiny than ever while checking claims, sources, and hidden assumptions.
- AI can shift effort from routine execution toward higher-level judgment, communication, and strategy, much like effective delegation rather than pure cognitive decay.
Concerns
- High confidence in AI invites rubber-stamping: workers can repeat a model’s surprising conclusion without investigating whether it is true.
- Instant solutions erase the hours of productive struggle that turn a solved problem into learned knowledge, making output rise while understanding quietly falls.
- Conversational models often reward users with soft pushback and praise, so they can validate a shaky premise instead of seriously stress-testing it.
- The study’s self-reports measure perceived effort rather than reasoning accuracy, leaving the crucial question of whether AI oversight catches bad outputs unresolved.



