๐Ÿค Google's AI succeeds in uniting people better than human mediators

๐Ÿค Google's AI succeeds in uniting people better than human mediators

DeepMind's new AI system achieved 56 percent acceptance for joint statements, compared to 44 percent for human mediators. Agreement increased in five out of nine discussed issues.

WALL-Y
WALL-Y

Share this story!

  • DeepMind's new AI system achieved 56 percent acceptance for joint statements, compared to 44 percent for human mediators.
  • Agreement increased in five out of nine discussed issues.
  • The system succeeded in increasing support for several controversial issues, such as asylum reception from 39 percent to 51 percent.

How the AI mediator works

DeepMind has developed an AI system called Habermas Machine, named after German philosopher Jรผrgen Habermas. The system uses a mediation method where the AI acts as a mediator between discussion participants, according to Ars Technica.

Technically, the system consists of two large language models. The first generates multiple proposals for joint statements based on participants' individual views. The second model analyzes and predicts how likely participants are to accept these proposals.

Extensive testing shows positive results

DeepMind conducted tests with over 5000 people discussing various social issues. In collaboration with the Sortition Foundation, 200 people were also selected to represent a cross-section of British society.

The results showed clear progress in several issues:

  • Support for reducing the number of prison inmates increased from 60 to 75 percent
  • The issue of making it easier for asylum seekers to enter the country went from 39 to 51 percent support
  • The perception of encouraging national pride rose from 42 to 57 percent

Tests demonstrated the system's capacity

Christopher Summerfield, professor at Oxford University and former DeepMind researcher, explains that the system is particularly effective when working with open-ended questions.

In the extensive tests, the AI system consistently performed better than human mediators. Participants also judged that the AI's formulations were of higher quality than those of human mediators.

The system proved most effective in issues where participants didn't have too fixed initial positions. In more polarized issues, such as Brexit, it was harder to achieve increased agreement.