AI: Anthropic's 'fascinating' internal test of AI Agents Shopping. RTZ #1068

Anthropic conducted a week-long experiment called ‘Project Deal’ where AI agents (Claude Opus and Claude Haiku) negotiated and traded goods for employees. Opus agents secured better prices, closing more deals and earning more on average. Despite this, users with Haiku agents rated transactions equally fair, highlighting an ‘invisible inequality’ in AI-assisted decision-making. The experiment revealed that model strength significantly impacted outcomes, with Opus agents earning $3.64 more on average for the same item. Participants preferred Opus runs but showed no statistical difference in satisfaction. Anthropic notes this as an ‘uncomfortable implication’ for AI companies considering mainstream AI agent commerce, emphasizing the need for policy and legal frameworks to address potential risks like economic inequality and security issues.