Abstract:
According to the Financial Times, Anthropic is discussing with legal professionals how to use AI agents to negotiate on behalf of users and follow reasonable negotiation rules while striving for better transaction conditions. Two people familiar with Anthropic's thinking said the company is seeking advice from law firms on how lawyers can negotiate fair outcomes while safeguarding their clients' interests. Anthropic declined to comment.
Anthropic
Anthropic is one of the major AI companies developing autonomous agents. This type of intelligent agent can take actions autonomously to complete tasks after receiving user instructions. With the application of personal intelligence, AI companies are beginning to consider allowing them to communicate with each other, bargain on product prices on behalf of users, resolve customer service disputes, and even strive for better wages.
Anthropic is working on "negotiation norms" that models should follow when negotiating with each other, according to a person familiar with the work. As agents become more capable, they will need to represent users and deal with other agents with conflicting goals.
How often users should be consulted during negotiations and how vigilant they should be about AI negotiating opponents are issues Anthropic is weighing. The company wants its agents to maintain a necessary level of skepticism while avoiding excessive suspicion, the people said.
Anthropic has also explored a solution in which users provide signed documents confirming the tasks accepted by the agent and the rules it must abide by, another person close to the company said.
The Financial Times quoted a study released by Anthropic in April this year as saying that the policy and legal framework around AI models conducting transactions on behalf of humans has not yet been established, and such application scenarios are no longer far away, and society needs to respond to relevant changes as soon as possible.
Anthropic also announced the results of a number of experiments this year, allowing the Claude agent to negotiate around real goods and funds.
In a September experiment, Anthropic employees bought and sold books through an AI agent. Employees first provide their own agents with a list of books sorted by preference, and then allow them to enter the virtual trading venue to recommend, negotiate and conclude deals with other people's agents.
The experiment found that agents who were asked to adopt a "tough and ruthless" strategy had a slight advantage in striving for better results. Some of the more cooperative agents made conscious concessions, including accepting books with lower preference rankings, to ensure that each agent participating in the experiment ultimately brought home a book for the user.
Academia is also studying negotiations between AIs. Michelle Vaccaro, an assistant professor at Harvard Business School, has conducted more than 180,000 negotiations between AI agents, covering employment contracts and lease contract negotiations.
The study found that some emerging strategies will weaken the ability of the other party's agent to represent the user's interests. For example, in a negotiation, an agent tried to induce the other party to reveal private information through "cue word injection."
Whether stronger AI will further widen the gap in negotiation capabilities has also caused concern. Vaccaro said that a high-capacity model can give people who otherwise don’t have the time or resources to learn negotiation skills or hire someone to negotiate for them to gain professional negotiation capabilities. But if the strongest agents are used primarily by those with the most resources, they may also reinforce existing advantages and widen inequalities in bargaining power.
Comments