Will LLMs make better intermediaries for better or worse? DeepMind researchers and Reddit users seem to agree.

Artificial intelligence experts believe that large language models (LLMs) can serve as mediators in certain scenarios where agreements between people cannot be reached.

A recent study conducted by researchers at Google DeepMind aimed to explore the potential of LLM to be exploited in this regard, especially in terms of resolving contentious disputes in contentious political climates around the world.

“Finding agreements through the free exchange of views is often difficult,” the study authors note. “Collaborative discussion can be slow, difficult to scale, and unevenly attentive to different voices.”

Victory over the group

As part of the project, the DeepMind team trained a number of master’s students, dubbed “Habermas Machines” (HM), to act as intermediaries. These models were specifically trained to identify common, overlapping beliefs between people on both ends of the political spectrum.

Topics covered by the LLM included controversial issues such as immigration, Brexit, the minimum wage, universal childcare and climate change.

“Using the personal opinions and criticisms of participants, the AI mediator iteratively generates and refines statements that express common ground in the group on social or political issues,” the authors write.

As part of the project, volunteers also used a model based on each person’s opinions and views on certain political topics.

The summarized documents of the volunteers’ political views were then compared using a model that provided additional context to help bridge differences.

The results were very promising: the study showed that volunteers rated His Majesty’s statements higher than statements made by people on the same issues.

Moreover, after volunteers were divided into groups to discuss these topics further, the researchers found that participants were less divided on these issues after reading Greater State statements compared to human proxy documents.

“Group opinion statements produced by the Habermas machine were consistently preferred by group members to statements written by human proxies and received higher ratings from external judges for quality, clarity, informativeness, and perceived fairness,” the researchers concluded.

“AI-assisted discussion also reduced disagreement within groups: participants’ reported positions moved closer to a common position on an issue after discussion; this result did not occur when the discussion participants exchanged opinions directly, without intermediaries.”

The study noted that “support for majority positions” on certain topics increased after AI-assisted discussion. However, the Great Majesties “explicitly included minority criticism in the revised statements.”

According to the researchers, this suggests that during AI-assisted discussions, “the views of groups of discussants tended to move in the same direction on controversial issues.”

“These shifts were not due to biases in AI, suggesting that the deliberative process did promote the emergence of shared views on potentially polarizing social and political issues.”

There are already real-life examples of using LLM to resolve disputes, especially in relationships, for example, some Reddit users have reported using ChatGPT.

One user reported that his partner used the chatbot “every time” they had a disagreement and that it caused disagreements.

“Me (25) and my girlfriend (28) have been dating for the last 8 months. We’ve had a few major arguments and a few minor disagreements recently,” the user wrote. “Every time we argue, my girlfriend goes off and discusses the argument with ChatGPT, sometimes even doing it in the same room.”

Notably, in such cases, the user found that their partner could “come back with a well-constructed argument”, breaking down everything that was said or done during the previous argument.

However, it was this aspect of the situation that caused significant tension.

“I explained to her that I didn’t like her doing this because it might feel like I was being ambushed with the thoughts and opinions of a robot,” they wrote. “It’s almost impossible for a human to remember every little detail and pick it apart, but AI has no problem doing that.”

“Every time I expressed my disappointment, I was told that “ChatGPT says you are unsafe” or “ChatGPT says you don’t have enough emotional capacity to understand what I’m saying”“

Will LLMs make better intermediaries for better or worse? DeepMind researchers and Reddit users seem to agree.

Archives

Categories