Last night, I read a white paper from Google DeepMind about an AI system they are developing designed to mediate discussions, especially on divisive topics, to help groups reach consensus. By the numbers, the system was more successful than human mediators. I was so intrigued that I read the paper twice. It kept me up all night.
DeepMind’s system uses LLMs to analyze and synthesize viewpoints, producing summaries that reflect shared positions while including dissenting opinions. It operates by soliciting feedback from participants, revising its summaries based on their input in a continuous loop until an acceptable version is reached. While the AI is effective at producing agreeable summaries, DeepMind acknowledges that it lacks certain human mediation skills, such as moderating off-topic remarks and ensuring factual accuracy. This sounds awesome, except…
When people ask me what I fear most about AI, my answer is “nothing.” I’m not worried about Skynet or other apocalyptic scenarios. I am scared to death about artificial control, which I explored in The Next Great Decoupling: AI Takes Control.
As I read this new paper, all I could think about was how easily model bias could impact any outcome. There are no “neutral” mediators. All humans have biases. The danger is believing there could ever be a neutral AI mediator. My challenge to DeepMind: prove me wrong.
Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.