The Impact of Generative AI on Critical Thinking [pdf]

What the study actually measures

• Several commenters stress the paper studies self‑reported critical thinking within AI-assisted tasks, not whether AI makes people generally less intelligent.
• Core finding as interpreted: higher trust in GenAI correlates with less perceived critical thinking effort; higher self-confidence correlates with more.
• Some argue this mainly reveals a general human tendency: when a tool gives plausible answers, people stop digging unless they have a clear reason.

Methodological criticisms

• Strong pushback on survey design:
  • “Critical thinking” is self-assessed, not objectively tested.
  • Items like “I always trust AI” vs “I question AI’s intentions” blur constructs (frequency, trust, and critical thinking intertwined).
  • Concerns about small, self-selected sample and lack of control group.
• Some still find the qualitative examples helpful for articulating how work shifts from creation to verification.

AI as cognitive tool: shifting vs losing skills

• One camp: technology historically reallocates cognition (printing press, calculators, GPS, high-level languages). We offload rote skills and move “up the stack” to abstraction and design. Losing obsolete skills can be fine if outcomes improve.
• Opposing camp: some “obsolete” skills (navigation, writing, arithmetic) are still practically and psychologically important (resilience, autonomy, motivation). Tech-driven offloading may contribute to alienation and dependence.

Experiences using LLMs at work

• Mixed utility reports:
  • Helpful for boilerplate code, style rewrites, intros, search-like Q&A, and tutoring/quiz generation.
  • Others find them slow things down: more time spent “babysitting” or rewriting “AI slop” than coding directly.
  • For complex or novel systems, AI often fails; by the time humans take over, the learning curve is steeper.
• Concern that replacing junior work with AI undermines experiential learning.

Verification, trust, and critical thinking

• Strong emphasis that LLM outputs must be verified; hallucinations are frequent, especially dangerous for learners who don’t know what to check.
• Some compare AI to GPS: great for most cases, but failure modes can be severe and people tend to follow it uncritically.
• Others argue AI, like a pry bar, gives “cognitive leverage,” freeing time for higher-level reasoning—if users understand limitations and switch to other tools (books, experts) when precision matters.
• Multiple commenters note that people already over-trusted Google; LLMs accelerate an existing problem of treating convenient outputs as ground truth.

Education and cognitive development

• Worries about students using AI/Grammarly instead of learning spelling, writing, or problem solving, and about a generation habituated to an “easy button.”
• Counterargument that strict spelling and similar norms are largely social conventions whose importance might be overstated.
• Desire for longitudinal studies on how AI changes what skills are learned and which atrophy, rather than just how people feel about thinking effort.

Knowledge ecosystem and “truth decay”

• Some fear that as more sources become AI-generated or polluted by hallucinations, independent verification of facts may become impossible, worsening existing misinformation dynamics.
• This is framed as a potential threat not just to individual critical thinking but to shared knowledge and social trust.

Societal and governance concerns

• Skepticism that even strong evidence of cognitive harm would slow AI deployment, given money and inertia; climate change is cited as an analogy with mixed precedent.
• Concerns about future AR+AI scenarios where people are continuously guided by models, becoming “meat robots”; others note society is already heavily shaped by education, media, and platforms.
• Some see a paradox: AI vendors warn users to verify outputs while simultaneously marketing systems as highly capable, making widespread over-trust predictable.