Anthropic sets fair negotiation norms for AI agents with law firms

Anthropic's move into negotiation norms comes from a data gap it has run into. Improving AI models in math and code is relatively easy because verifiable data is plentiful or easy to generate, but in the humanities that data doesn't exist, partly because right and wrong answers are rarely clear-cut. Justin Curl, a member of technical staff researching law, institutions, and AI at Anthropic, argues that in law it is "notoriously hard to define quality." The Conceptual Reasoning Fellows program — lawyers spending four months testing Claude on judgment-heavy questions for $100,000 to $167,000 — is aimed at building the test cases and evaluations needed to make Claude stronger in those areas.

The negotiation work is not new territory for the company. Anthropic has already shown in experiments that stronger models negotiate better deals, a result that could widen societal inequality. Consulting with law firms to develop fair negotiation norms appears to be an attempt to set expectations for agents that haggle over prices, resolve complaints, or improve contract offers before such agents start cutting deals on users' behalf.