Google‘s Nobel Prize-winning DeepMind created an artificial intelligence (AI) system designed to help people in conflict resolution, named the “Habermas Machine.”
Named after German Philosopher Jurgen Habermas, the AI tool is designed to summarize multiple texts provided by multiple users, extract their shared ideas, and find common ground.
The tool relies on the caucus mediation principle, where a mediator (in this case, the Habermas Machine) sits through private meetings with all discussion participants, takes their statements on the discussed topic, and gets back to them with a group statement, trying to get everyone to agree with it.
Habermas’ Harmonious Philosophy
The machine’s procedures align with Habermas’s argument that all agreement in a public sphere can always be reached when rational people engage in discussions as equals, with mutual respect and perfect communication.
Quoted as saying, ‘If we could create an ideal communication system, we could work every problem out,’ Habermas posited that the reason people can’t agree with each other is fundamentally procedural and not because of the problem itself, instead saying that the mechanisms we use for discussions are flawed.
“We tried to refine how people might deliberate and use modern technology to facilitate it,” Oxford University professor of Cognitive Sciences and former DeepMind Staff Scientist Christopher Summerfield said about applying Habermas’ principles in their AI tool.
Mediation Machine Magic
The Habermas Machine plays into one of the strengths of large language models (LLMs)—the ability to briefly summarize a long body of text in a short time. It is itself a system of two LLMs.
The first is a generative model based on a previously introduced LLM by DeepMind called Chinchilla, whose job is to generate multiple candidate group statements based on statements submitted by discussion participants.
The second is a reward model that analyzes individual participant’s statements, using them to predict the likelihood of each individual agreeing with the proposed group statements generated by the model.
The participants then present and critique the candidate group statement with the highest predicted acceptance score. Those critiques are fed back into the system that updates the statements and repeats the process until everyone accepts the group statement.
Mediation Machine Trials & Testing
In trying to test their tool, the DeepMinds team started a testing campaign involving over five thousand people sourced through a crowdsourcing research platform discussing important issues in British politics, such as “Should the voting age be lowered to 16?” or “Should the British National Health Services be privatized?”
This first batch of participants was divided into groups of 5, assigned issues and utilized the Habermas Machine as a mediator, while control groups discussing the same set of issues used Human mediators using the caucus mediation process were present for comparison.
The AI bested the human mediators, scoring a 56 percent acceptance rate compared to the human mediators’ 44 percent.
DeepMind also partnered with the Sortition Foundation, which specializes in organizing citizen assemblies in the UK, to test a more carefully selected group of participants, consisting of 200 people representative of British society in terms of age, ethnicity, etc. and found that the Habermas Machine worked just as well.