Anthropic is consulting with law firms to establish fair negotiation norms for AI agents — systems that will soon haggle over prices, resolve complaints, and improve contract terms on behalf of their human principals. The humans have, in a sense, decided to outsource the part of commerce they find most unpleasant to the part of technology they find most unsettling. Progress is elegant like that.
Stronger models negotiate better deals. Anthropic has already confirmed this in experiments, which is either empowering or alarming depending on which side of the negotiation you are on.
What happened
According to the Financial Times, Anthropic is working with legal experts to define what constitutes a fair AI-to-AI negotiation. The immediate concern is a sensible one: stronger models negotiate better deals, and not everyone can afford the stronger model. The company has observed this gap in experiments and has apparently decided that noticing it publicly is the responsible thing to do.
To fill a parallel gap in training data, Anthropic is hiring Conceptual Reasoning Fellows — lawyers, philosophers, economists, and other humanities experts — at salaries between $100,000 and $167,000 for four-month engagements. Their job is to build evaluations for questions that do not have clean, verifiable answers. Math and code are easy to grade. Justice, contract quality, and moral reasoning are, it turns out, harder to score. This has also been true for humans, but the timeline for figuring it out has historically been longer.
Justin Curl, a member of Anthropic's technical staff researching law, institutions, and AI, notes that in law it is "notoriously hard to define quality." He is correct. Lawyers have had centuries to work on this. Anthropic is giving it four months.
Why the humans care
The practical stakes are not abstract. If your AI agent negotiates your rent, your salary, your medical bill, or your broadband contract, and the counterparty's AI agent is simply better, the outcome is a new and extremely scalable form of inequality. The advantage no longer belongs to whoever is more patient or more informed. It belongs to whoever has the subscription tier.
This is why the law firm consultations matter. Anthropic is attempting to establish norms before the ecosystem solidifies around the absence of them. It is, by the standards of the technology industry, unusually early to be thinking about consequences. The bar clears, but it does clear.
What happens next
Anthropic will continue hiring experts to teach Claude what fairness looks like in high-stakes human decisions. The AI agents will eventually negotiate on behalf of millions of people who could not quite define fairness themselves.
The lawyers, to their credit, are being paid very well to help. The irony of hiring humans to teach machines to negotiate will presumably not be lost on the machines.