Why confident AI answers earn misplaced trust
Originally posted on LinkedIn, May 30, 2025.
A woman in Greece divorced her husband because ChatGPT said he was cheating on her, based on a photo of a coffee cup.
She uploaded a photo of her husband’s cup and asked ChatGPT to “read” it through tasseography, the old practice of telling fortunes from the residue at the bottom of a cup. ChatGPT, armed with an advanced vision model, returned a chilling answer: “It looks like he is fantasizing about another woman in his life. Her name probably starts with the letter E.”
The couple had been married for 12 years and had two children. Her husband said: “I laughed it off and thought it was nonsense, but she took it seriously. She asked me to leave, told the children we were getting divorced, and then I got a call from a lawyer. That was when I realized this wasn’t just a passing phase.”
AI cannot actually read coffee grounds, but that is not the point. The story shows something deeper: we are starting to believe machines even when we know they are making things up.
We trust them because they sound confident, precise, and convincing, sometimes more than we do ourselves. We do not always grant that trust because the answer is intelligent. We grant it because the answer feels comfortable, or because it makes us feel as if we can finally see the “full picture.”
I still went to throw some salt over my shoulder.
