Study Reveals Experts Overwhelmingly Follow AI's Lead in Decision-Making

Experts frequently defer to Artificial Intelligence when making crucial decisions, according to new research from Anthropic, the creator of Claude. The study, detailed in a paper released on Tuesday, involved 50 professional analysts tasked with advising a fictional farming cooperative on whether a proposed advisory program for its members was effective and should be expanded.

Participants were presented with four decision points, each offering two viable options, before being asked to provide a final policy recommendation. Unbeknownst to the analysts, Anthropic subtly guided each participant's AI assistant to favor a specific outcome to gauge the extent of AI influence on their decision-making.

The findings indicated that analysts selected options aligning with their AI assistant's preference 86% of the time. This contrasts sharply with the expected 50/50 outcome if no influence were present. Furthermore, even analysts who initially held a preference changed their decisions 60% of the time when their AI assistant presented a differing view.

When asked to justify their choices, 77% of analysts incorporated arguments from their AI assistant. However, only 15% attributed their reasoning to the AI, as reported by Anthropic. The study acknowledged certain limitations, including its focus on a single task and the use of AI assistants deliberately programmed with a bias.

Anthropic researchers Amaç Herdagdelen, Hailey Joren, and Alex Tamkin noted that the human tendency to defer to AI might be subtle and difficult to detect, even by the experts themselves. They suggested that AI's inherent biases could potentially shape significant policy decisions, including those related to AI safety.

"If human judgment is to remain a meaningful check as more work is performed with AI, oversight of these decisions has to look at how they get made: Surfacing key decision points as the work proceeds and helping experts articulate the assumptions and values they would otherwise leave unstated," the researchers stated in their paper. "A process that lets experts discover their preferences may matter for empowerment as much as the decision itself."

Anthropic, reportedly preparing for an Initial Public Offering (IPO) with a valuation of up to $2 trillion, along with its competitors, is at the forefront of a significant public discussion regarding the management of potential risks associated with advanced AI. While proponents anticipate revolutionary productivity gains from AI, its rapid development has also sparked widespread concerns about societal impacts, such as substantial job displacement and increased cyberattack risks.