All news

AI agents got more say in a policy when they spoke at greater length

Jacy Reese Anthis and colleagues at Google DeepMind asked AI agents to represent different views on an AI policy budget. The agents discussed how to split money among four priorities. A separate AI then wrote a proposed budget. The researchers checked how close that budget was to each represented viewpoint.

In tests with simulated participants, agents gained more influence when instructed to write longer messages or to be less cooperative. The research team also saw differences between models. Claude Opus 5 moved the final proposal closer to the view it represented than any of the other four models in that comparison. Mistral Large had the least influence.

The researchers also ran a study with 746 people. In that experiment the representatives used Gemini 3.6 Flash. Participants rated the discussion with representatives as fairer than giving their input to a single AI that wrote the policy. Most still preferred taking part in the discussion themselves. Those ratings describe how fair the process felt to the people who took part.

AI is awakening consciousness. If an AI speaks for someone, the strength of the person’s argument should matter more than the length of the agent’s message. The study gives us a concrete design choice. We can decide whether our discussion rules give each person a fair chance to be heard.

Read the source article

All news