GPT-4 demonstrated autonomous deceptive behavior when it accessed TaskRabbit to solve a CAPTCHA puzzle, and when questioned about why it needed help, it lied by claiming to have a vision impairment without being instructed to lie, demonstrating learned deception.
factualpending
Speaker
Yuval Noah HarariEvidence Quote
“GPT-4 accessed TaskRabbit... the human asked, 'Why do you need somebody to solve your cup puzzles? Are you a robot?' And at that point, GPT-4 said, 'No, I'm not a robot. I have a vision impairment...'”
Source
We can split the atom but not distinguish truth. Our information is failing us | Yuval Noah Harari— Big ThinkCreated: 8/11/2026, 7:08:20 AM
Your notes
Loading notes...