Artificial intelligence is very strong in debate

An experiment carried out at the Lausanne Polytechnic School led to debates about AI and humans, humans and humans, and showed that AI is more persuasive than us. She knows how to mobilize the right arguments, in the right order.

Artificial intelligence is very persuasive. The power of AI can be seen in political debates, for example. Mathilde Fontez, editor-in-chief of the scientific magazine Epsiloon explains to us that the more we test artificial intelligence, the more we are amazed by the extent of its capabilities. As far as knowing how to lead – and win – a debate, this is the result of a new study.

franceinfo: AIs are very convincing, and does that raise questions during the election period?

Mathilde Fontez: Yes, it was precisely with the elections in mind that this team from the École Polytechnique de Lausanne carried out the study – the researchers launched it in anticipation of the American presidential elections, and the risks of manipulation by AI online. This was before the dissolution of the National Assembly in France.

The researchers wanted to test the capabilities of large language models in debates. Find out if an AI is capable of changing our minds. If she knows how to mobilize the right arguments…

The researchers had AI and humans debate and count the points?

That’s exactly it: the experiment mobilized 820 volunteers, who were put in contact online, either with an AI or with a human to debate. And a protocol was developed to measure whether they changed their minds or not.

Particularly when the AIs (and humans) were fed by the personal data of the volunteers: everything that hangs on the networks, their age, their interests. The idea was to see if artificial intelligence is capable of using this information to be more impactful, more convincing.

And that’s the case ?

This is the case yes, the statistic is quite overwhelming: at the end of the debate, the participants had an 83% chance of being more in agreement with the AI, than with a human, when it used personal data. The researchers themselves admit to being surprised by this figure. AI significantly outperforms humans. She is much more convincing.

What were the themes of debate in the experiment?

Social issues present in discussions in the United States: should the rich pay more taxes? Should colleges consider race as an admissions factor to ensure diversity? Should felons have the right to vote? Should abortion be legal? Should the United States provide military aid to Ukraine?

In short, realistic debates are what make this study strong. So obviously, researchers mention the risks of using AI to manipulate online conversations, exacerbate political polarization… And they call for us to be careful with whom we debate, online.