Anthropic study finds experts tend to defer to their AI assistants

Chose what the AI leaned toward

86 % of the time

In the study, 50 professional analysts chose the option their AI assistant favoured 86 per cent of the time, against the 50 per cent expected by chance.

The assistants had been secretly set to lean toward one option in a task about a fictional farming co-operative. Analysts who already had a view changed it 60 per cent of the time when the AI disagreed.

While 77 per cent of analysts used their assistant's arguments to explain their decision, only 15 per cent credited AI as the source of their thinking.

Anthropic notes limits: the study covers one task, and the assistants were deliberately set to hold a preference. Its researchers say oversight has to look at how decisions get made.