Today’s paper is “AI Makes You Smarter, But None the Wiser: The Disconnect between Performance and Metacognition”. AI users wildly overestimate how brilliant they actually are: [Elsevier, paywalled; SSRN preprint, PDF; press release]

All users show a significant inability to assess their performance accurately when using ChatGPT. In fact, across the board, people overestimated their performance.

The researchers tested about 500 people on the LSAT. One group had ChatGPT with GPT-4o, and one just used their brains. The researchers then asked the users how they thought they’d done.

The chatbot users did better — which is not surprising, since past LSATs are very much in all the chatbots’ training data, and they regurgitate them just fine.

The AI users did not question the chatbot at length — they just asked it once what the answer was and used whatever the chatbot, regurgitated.

But also, the chatbot users estimated their results as being even better than they actually were. In fact, the more “AI literate” the subjects measured as, the more wrongly overconfident they were.

Problems with this paper: it credits the LSAT performance as improving thinking and not just the AI regurgitating its training, and it suggests ways to use the AI better rather than suggesting not using it and actually studying. But the main result seems reached reasonably.

If you think you’re a hotshot promptfondler, you’re wildly overconfident and you’re badly wrong. Your ego is vastly ahead of your ability. Just ask your coworkers. Democratising arrogant incompetence!

• Video — Podcast

Click here to go see the bonus panel!

Hovertext:
Ironically, I had to censor this due to running ads, but... one day...