The Confidence Trap: Why Trusting AI Makes You Think Less

A team at Microsoft Research just published something that should make every
knowledge worker pause:
a study of 319 professionals using
ChatGPT, Copilot, and other AI tools revealed a troubling pattern. The more
confident you are in AI's ability to do a task, the less you engage your own
critical thinking—even when that thinking is exactly what separates good work
from mediocre output.

I've been watching this play out in real-time with the professionals I train at
EverydayAI. But seeing it quantified across 936 real-world work examples? That's
a wake-up call.

The Confidence Paradox

Here's the finding that stopped me cold: higher confidence in AI is
statistically associated with less critical thinking. Not because the AI is
doing all the thinking for you—but because you stop questioning whether it
should.

Meanwhile, workers who had high confidence in their own abilities engaged in
more critical thinking, even though they reported it requiring more effort. They
checked sources. They refined outputs. They caught errors. They maintained
agency over their work.

Think about what this means. The path to excellence isn't trusting the AI
more—it's trusting yourself enough to challenge what it produces.

What "Less Critical Thinking" Actually Looks Like

The researchers didn't measure vague feelings. They tracked specific cognitive
activities across Bloom's taxonomy—the gold standard framework for understanding
thinking skills. When workers had high AI confidence, they reported reduced
effort in:

• Analysis: Breaking down problems into components (-15% effort)
• Evaluation: Judging quality against criteria (-23% effort)
• Synthesis: Combining ideas into new meanings (-12% effort)
• Knowledge recall: Even basic fact-checking dropped (-11% effort)

This isn't efficiency. It's cognitive offloading—outsourcing the mental work
that makes you good at your job.

One participant, a market researcher, said it plainly: "I trust ChatGPT for
straightforward factual information. It usually gives good answers, so I don't
think about it."

Except when ChatGPT hallucinates a competitor that doesn't exist. Or cites a
study that was never published. Or confidently states outdated regulations as
current law.

The Self-Confidence Shield

Here's the hopeful finding: workers who were confident in their own skills did
the opposite. They:

• Cross-referenced AI outputs against external sources (114 out of 319 did this
  consistently)
• Verified information using their domain expertise
• Evaluated outputs against both objective criteria and subjective quality
  standards
• Integrated only relevant parts of AI responses rather than wholesale copying

A nurse in the study used ChatGPT to create an educational pamphlet for diabetic
patients. But she didn't just export and print it. She cross-checked every
medical claim against her hospital's diabetes management guidelines. She caught
three subtle errors that could have confused patients about insulin timing.

Her self-confidence made her skeptical—in the best possible way.

Why This Matters for Your Career

I talk a lot about the "Second Renaissance"—this compressed period where AI
transforms knowledge work faster than any previous technology. But if the
printing press era teaches us anything, it's that tools amplify existing
inequalities.

The workers who develop deep expertise and maintain critical thinking habits?
They'll use AI to 10x their output quality and speed.

The workers who over-rely on AI as a replacement for thinking? They'll plateau.
Or worse, they'll produce increasingly generic, error-prone work and not even
realize it.

The researchers found that 59% of knowledge workers reported engaging in
critical thinking when using AI tools. That means 41% just... didn't. They
accepted outputs at face value, especially for "routine" or "low-stakes" tasks.

But here's the cognitive trap: if you only practice critical thinking on
high-stakes work, you're not building the muscle. You're like an athlete who
only sprints during the championship game. When it matters most, you won't have
the reflexes.

The Bainbridge Problem, AI Edition

There's a concept from automation research called "Bainbridge's Ironies of
Automation." The idea: when you automate routine tasks and leave only exceptions
to humans, you deprive people of the routine practice needed to handle those
exceptions skillfully.

The Microsoft researchers found this exact pattern with GenAI. Workers said
things like:

• "AI handles the easy stuff, but when it generates something wrong in an area
  I don't know well, I can't tell" (Participant 290)
• "I use it because I must hit quotas. I don't have time to verify"
  (Participant 295, sales role)
• "The AI writes my emails. I make sure they sound like me, but I don't really
  evaluate the content" (Participant 254)

This is how skills atrophy. Slowly, task by task, until you need them and
they're just... gone.

What Actually Works: Building the Right Confidence Balance

The study identified three types of confidence that matter:

1. Confidence in yourself doing the task (positive correlation with critical
   thinking)
2. Confidence in AI doing the task (negative correlation with critical
   thinking)
3. Confidence in evaluating AI's output (positive correlation with critical
   thinking)

The sweet spot? High on #1 and #3, appropriately calibrated on #2.

Here's how to build that:

For self-confidence:

• Deliberately practice tasks without AI first
• Use AI as a draft generator, not a final product creator
• Track instances where you caught AI errors—this builds your evaluation
  instincts

For evaluation confidence:

• Learn the failure modes of your specific AI tools (hallucination patterns,
  recency limits, bias tendencies)
• Develop domain expertise that lets you spot implausible claims instantly
• Create personal quality checklists for different task types

For AI confidence calibration:

• Assume AI is wrong until proven right, not the other way around
• Test AI outputs on tasks where you know the correct answer
• Document when AI fails so you build realistic mental models

The Long Game

Here's what I tell the educators and professionals I work with: AI should make
you better at thinking, not better at avoiding it.

The researchers found that workers who maintained critical thinking habits cited
three motivations:

1. Improving work quality (74 people mentioned this)
2. Avoiding negative outcomes (116 people mentioned this)
3. Skill development (13 people mentioned this)

That last one is criminally underrepresented. Only 13 out of 319 knowledge
workers were thinking about AI as a tool for learning and growth, not just
productivity.

One participant (P154) stood out to me. When ChatGPT solved a coding problem,
they said: "I make sure that I understood how it works and can do it by myself
next time."

That's the mindset. Use AI to expand your capabilities, not replace them.

Your Move

The next time you use ChatGPT, Copilot, Claude, or any GenAI tool, ask yourself:

• Am I using this because I'm confident in my ability to evaluate the output?
• Or am I using this because I'm confident the AI will handle it and I don't
  need to think about it?

The first leads to growth. The second leads to dependence.

The researchers put it bluntly in their conclusion: "GenAI tools reduce the
perceived effort of critical thinking while also encouraging over-reliance on
AI, with confidence in the tool often diminishing independent problem-solving."

Don't let the efficiency gains blind you to the capability losses.

Your competitive advantage in the AI era isn't trusting the tools more. It's
trusting yourself enough to question them.

Related Reading

This post is part of a series on maintaining critical thinking in the AI era:

• From Doer to Steward: How AI Is Rewiring the Way You Think -
  Understand the three cognitive shifts happening to your brain
• Don't Let AI Make You Lazy: A Practical Guide to Staying Sharp -
  Concrete tactics for overcoming awareness, motivation, and ability barriers
• The Second Renaissance: Why AI Isn't Like the Printing Press -
  Why this transformation is compressed and what it means for you

Explore the full analysis:
The Second Renaissance project page offers a
comprehensive look at AI's transformation of knowledge work.

This post draws on: Lee, H.P., Sarkar, A., et al. (2025). "The Impact of
Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort
and Confidence Effects From a Survey of Knowledge Workers." CHI Conference on
Human Factors in Computing Systems.