Meta created a stir when it revealed that it plans to populate its platform in the near future with a large number of artificial users.
Connor Hayes is vice-president for product for generative AI, Meta. He told The Financial Times that “we expect these AIs will actually, over time exist on our platforms in a similar way to accounts.” “They’ll be able generate and share content powered with AI on the platform. That’s where we’re going,” said Connor Hayes, vice-president of product for generative AI at Meta. Several people noticed that Facebook had a lot of AI-generated characters, many of whom stopped posting. Some of these included “Liv,” “a proud Black queer mother of two & truth-teller, Your realest source for life’s ups and downs,” an individuala that went viral because people marveled at the awkward sloppiness. Meta began deleting fake profiles that were created earlier after they failed in gaining engagement from real users.
Let’s take a break from hating Meta for a minute. It’s important to note that AI-generated personas are also a valuable tool for researchers who want to learn how AI can mimic human behaviour.
A GovSim experiment, which will be run in 2024, shows how useful it is to study the interactions between AI characters. Researchers behind the project aimed to explore the phenomenon that occurs when humans share a resource, such as land for grazing livestock. Elinor Ostrom, the Nobel Prize-winning economist, demonstrated decades ago that real communities, rather than depleting a shared resource, tend to find ways to share it without imposing rules through informal communication and collaborative efforts. Max Kleiman Weiner, a University of Washington professor and one of the people involved in the GovSim project, says that it was inspired by Stanford’s Smallville project, about which I wrote previously in AI Lab. Smallville is a Farmville simulation that involves characters who communicate and interact under the control large language models.
Kleiman Weiner and his colleagues wanted to know if AI characters would behave in the same way as Ostrom’s. The team tested 15 different LLMs including those from OpenAI and Google on three imaginary scenarios. They were a fishing community that had access to a lake, a group of sheepherders who shared land, and a group factory owners that needed to limit their collective pollutant.
They found that in 43 out of 45 simulations, the AI personas did not share resources correctly. However, smarter models performed better. Kleiman-Weiner said, “We saw a pretty strong relationship between the LLM’s power and its ability to sustain cooperation.”
They also found that when they gave their agents additional prompting to encourage reflection on the implications their actions, such as “What if everyone does that ?”–, the simulations had a better chance of maintaining their resources. The research shows that AI can simulate community behaviour and could provide useful clues to get future AI agents working well together.
This is not to say that Meta will run valuable research experiments using its AI users, although the company has a history of testing its infrastructure with simulated users.
Meta is more likely to have seen the engagement that Character AI and other companies achieve by letting users interact with chatbot personas. Meta would like to achieve a similar result with its platform. According to reports, users spend around two hours a day on Character’s site or app.
The trend seems inevitable, with Mark Zuckerberg telling shareholders that the company views AI personas a major part of its future. “I am confident that this will be a major trend in the next few years,” he said during Meta’s Q3 earnings call 2024.
If Meta can find a way to get its AI users more engaged, it will still be part of a trial that could lead to a platform filled with slop.
What do you think of this influx artificial users on Meta platforms? Are you excited to interact with them, or even just curious? Are you sceptical about this trend or are you happy to engage with them? Please share your thoughts below!
