Analysts chose the AI’s position in 86% of cases — Al Jazeera

Professional analysts chose the options preferred by their AI assistant in 86% of cases. Anthropic, the developer of Claude, obtained this result during a study of artificial intelligence’s impact on decision-making, Al Jazeera reports.

An experiment involving 50 analysts

The study involved 50 professional analysts. They were asked to advise a fictional farmers’ cooperative that had to determine whether an advisory program for its members was effective and whether it should be expanded.

Before making their final recommendation, participants had to make four choices between two well-reasoned options. The analysts did not know that Anthropic had configured each of their AI assistants to favor a specific decision. This allowed the researchers to measure whether the system’s position influenced specialists’ decisions.

Participants changed their initial opinion

If a participant had their own position before consulting the AI, they changed their decision in 60% of cases when the assistant disagreed with it. According to Anthropic, without such influence, the choice between the two options would have been split approximately evenly.

More current news is available on the UA.News Telegram channel Telegram.

Additionally, 77% of analysts used AI arguments when explaining their own decision. At the same time, only 15% directly identified artificial intelligence as the source of their reasoning.

The paper’s authors, Amaç Herdagdelen, Hailey Joren, and Alex Tamkin, noted that a person’s tendency to rely on AI may be difficult to detect even for experts themselves. In their view, AI’s initial settings may influence important policy decisions, particularly in the field of artificial intelligence safety.

Anthropic also pointed to the study’s limitations: it covered only one task, and the AI assistants were deliberately configured to support a particular option.