Latest / Safe As Podcast / E79: AI Makes you Smarter but not Wiser: How AI affects your Cognition

Transcript

E79: AI Makes you Smarter but not Wiser: How AI affects your Cognition · Safe As Podcast

• 0:00
  Generative AI can boost your test score, but can it quietly
• 0:03
  impede your ability to judge how well you're actually doing?
• 0:07
  Good day everyone. Welcome to Safe as I'm Doctor
• 0:09
  Ben Hutchinson. Today's study is from Fernandez
• 0:13
  ET al 2026 titled AI makes You Smarter but none the wiser.
• 0:18
  The disconnect between performance and metacognition.
• 0:21
  This study suggests that while generative AI acts a bit like a
• 0:25
  cognitive power up, it impacts our self awareness.
• 0:28
  For definitions, we're talking about metacognition, essentially
• 0:32
  thinking about thinking. Within that, we look at
• 0:36
  metacognitive accuracy, which is how well your estimated
• 0:40
  performance matches your actual results, and metacognitive
• 0:44
  sensitivity, your ability to distinguish your right answers
• 0:47
  from your wrong ones. Then there's the famous Dunning
• 0:50
  Krieger effect, the idea that low performers overrate their
• 0:53
  skills while experts tend to underrate theirs.
• 0:56
  So how did the researchers test this?
• 0:58
  They conducted 2 large scale studies.
• 1:00
  In Study 1 / 1200, participants tackled 20 logical reasoning
• 1:05
  questions from the Law School Admissions Test.
• 1:07
  They had to use chat GDP 4 O for help.
• 1:11
  Study 2 split 452 participants into AI and no AI groups
• 1:16
  offering cash bonuses for accurate self assessments to see
• 1:20
  if people would be more careful when money was on the line.
• 1:23
  So what did they find? AI definitely increased output.
• 1:26
  Participants using AI scored significantly higher than those
• 1:30
  working alone, about 3 points higher on average.
• 1:33
  However, when it comes to true synergy, that is humans plus AI
• 1:38
  outperforming the AI working alone, the picture isn't so
• 1:42
  impressive. Only about 55 to 59% of users
• 1:45
  actually beat chat GDP by itself.
• 1:48
  In fact, on average, participants performed slightly
• 1:51
  worse than the AI alone, though the difference was less than one
• 1:55
  point. So we see augmentation AI helps
• 1:58
  humans do better than they would unaided, but not genuine synergy
• 2:03
  where the collaboration exceeds what AI could achieve
• 2:06
  independently. Interestingly, even though their
• 2:08
  scores went up, users metacognitive accuracy
• 2:11
  plummeted. On average, people overestimated
• 2:15
  their performance by about four points.
• 2:17
  For instance, if they thought they got a 17 out of 20, they
• 2:20
  actually got a 13. Further, AI eliminated the
• 2:23
  traditional Dunning Kruger effect.
• 2:25
  Usually it's the low skilled people who are uniquely
• 2:29
  unskilled and unaware of their failures.
• 2:31
  They overestimate the most, while high performers are more
• 2:34
  accurate or even underestimate. But with AI, that relationship
• 2:38
  disappeared. Everyone, regardless of their
• 2:41
  actual skill level, became uniformly overconfident.
• 2:45
  The AI levels the playing field of performance, lifting low
• 2:48
  performers up to a higher baseline, but at the cost of
• 2:51
  making everyone equally deluded about how well they're doing.
• 2:54
  The performance boost came with the metacognitive cost across
• 2:57
  the board. Another insight is the literacy
• 3:00
  paradox. The study found that higher AI
• 3:03
  literacy correlated with lower metacognitive accuracy.
• 3:07
  People who understood the technical side of AI were more
• 3:10
  confident, but actually less precise in judging their own
• 3:13
  work. It seems that knowing what
• 3:15
  happens under the hood makes you trust it and yourself too much.
• 3:20
  So what can we make of the findings?
• 3:21
  Several implications are that we should implement cognitive
• 3:25
  forcing. Don't let users just copy and
• 3:27
  paste. Use prompts that require users
• 3:29
  to justify or critique the AI's logic before they submit.
• 3:33
  Have users restate the AI's reasoning in their own simple
• 3:36
  language to break the illusion of understanding.
• 3:39
  Design interfaces that show how unsure an AI is about a specific
• 3:43
  answer to help calibrate user trust.
• 3:45
  Focus on local monitoring since global self belief is often
• 3:49
  inflated. Design tools that force users to
• 3:52
  evaluate their work on a task by task basis rather than just at
• 3:56
  the end. In all, AI can be powerful
• 3:59
  tools, but if we aren't careful, we'll forget where the one's
• 4:02
  supposed to be driving. That's it on safe as.
• 4:05
  Thanks everybody. Please remember to like and
• 4:07
  subscribe, and visit safetyinsights.org for more
• 4:10
  research and links in comments if you'd like to support the
• 4:13
  channel.