A social media algorithm designed to bridge the gap between people with different political views could reduce the toxicity of online discourse and promote better conversations. It appears to be effective – at least in tests on hundreds of AI-generated users.
at the University of Amsterdam in the Netherlands and his colleagues created 500 chatbots powered by a large language model (LLM) and got them to interact with each other as if they were on a social media platform. The bots were encoded with personas designed to be representative of the US population, based on data from the 2020 American National Election Study.
The election study data included information on the news sources people consume, demographics, political beliefs and non-political interests, such as favourite TV shows and movies. The LLM was prompted with that demographic data and asked to produce posts that echoed each voice. “LLMs are very good at pretending to be a persona,” says Törnberg.
Advertisement
Those virtual users were then put into three separate social media platforms with a similar design to X, formerly known as Twitter, each dictated by different algorithms. The first showed users the most liked and commented-on messages only from users they follow. Each user in this version followed 30 other agents.
On the second platform simulation, users were shown the most liked and commented-on messages from all users on the platform – breaking their echo chambers. X and Instagram Threads have both used a version of this approach in their algorithmically dictated feeds.
The third simulation showed users messages from users from all backgrounds, but prioritised messages that had likes from users who were politically opposed to them. Users who identified as Republicans were shown posts popular with Democratic users, for example.
The first simulation produced a civil, if bland, platform. The second showed “much more interaction across the partisan divide, but also much more toxicity”, says Törnberg. However, the third managed to blend high engagement with low levels of toxicity – which could be a model for the future.
“What’s more interesting than their actual findings are the methods they used to explore this question,” says at Bentley University in Massachusetts. The ability to create your own representative users is beneficial, says Giansiracusa, because it removes the main challenge that academics face when conducting social media research: convincing the platforms they are studying to give up their data. “That’s starting to not be possible any more,” says Törnberg.
However, both the authors and Giansiracusa point out that AI-generated facsimiles of people with different personalities can only get you so far in emulating real human interaction on social media. “What comes to my mind is that chatbots can emulate certain personas and political perspectives through prompting, but they don’t actually have views or values or even consistent personalities,” says Giansiracusa. “We’ve all felt those strong visceral reactions to content on social media. A chatbot does not experience this.”
Törnberg suggests that LLMs can go some way to mimicking how people react emotionally when confronted by content designed to pique their interest. “They bring this kind of ability to react to text and to produce realistic conversations,” he says.
arXiv
