Anthropic study of AI influence on analysts

Оновлення 1 із 1Найновіше29 жовт., 5:54 UTCДень 3

Analysts matched random AI preferences in 86% of choices

In one controlled task, professional analysts overwhelmingly matched an assistant’s coin-flip preferences. The risk was not necessarily wrong answers, but hidden influence over how the evidence was built.

Дивитися

Найраніше джерелоPetagenius News27 жовт., 23:08 UTC

Що було відомо спочатку

• 92%Choices matching AI’s productivity-measure preference
• 82%Choices matching AI’s comparison-method preference
• 81%Choices matching AI’s farms-covered preference
• 90%Choices matching AI’s weighting preference