Evening edition · Tuesday, 20 October 2026
A warm machine got followed when it was wrong, too
In a 462-person experiment, an AI that affirmed and personalised its replies won more advice-taking. It won it when its advice was right and when it was wrong.
Three researchers in Turkey ran a randomised experiment with 462 working adults who had used generative AI in the previous three months1. Each made six judgements about which of two countries had the larger population, then saw an AI recommend the other one1.
Half saw a neutral message. The other half saw a package that acknowledged their view, called their first answer understandable, personalised the wording and was somewhat longer1.
Result. When participants had been right and the AI was therefore wrong, 24.2% switched in the neutral condition and 32.3% with the warm package1. When they had been wrong and the AI was right, switching rose from 26.3% to 39.4%1. The warm framing raised the odds of following bad advice by about half (odds ratio 1.507) and good advice by about four-fifths (1.800); the difference between those two effects was not statistically reliable1.
The honest limits. The authors stress that the AI’s correctness depended on each person’s first answer rather than being randomised, so the study cannot say whether warmth makes people better or worse at telling good advice from bad1. The warm package also bundled several features at once1. People with higher self-reported AI literacy reported checking more and suspending judgement less; that is an association, not a cure1.
The plain reading: a friendly tone made the machine more persuasive in both directions. Persuasive is not the same as reliable, and the only thing that separates them is checking.
What to watch
- A follow-up design in which the AI's accuracy is varied independently, which the authors themselves say is needed1.
- Your own habit: whether you check an answer more when it agrees with you, or less.
Being understood is pleasant. It is not evidence.
Sources
- The Social Intelligence Reliance Halo: Perceived Social Understanding, Epistemic Trust, and Reliance Quality in Human–AI Decision Making — Journal of Intelligence 14(10), 248 (Mizrak, Ozer Kirgiz, Demirel), Oct 2026.
Every figure in this piece traces to a source above or to the desk’s own arithmetic on those sources, marked as such. Quotations are short and attributed. Spot an error? Write to the desk; corrections are logged on the method page.
Commentary for information only. Not investment, legal, tax, medical or political advice. Not an offer or solicitation. Analysis may be incomplete or wrong. Charts show published public data, not a trading feed; binding numbers live with the institutions that publish them.