Humans are complicated beings. The ways we communicate are multilayered, and psychologists have devised many kinds of tests to measure our ability to infer meaning and understanding from interactions with each other. 

AI models are getting better at these tests. New research published today in Nature Human Behavior found that some large language models (LLMs) perform as well as, and in some cases better than, humans when presented with tasks designed to test the ability to track people’s mental states, known as “theory of mind.” 

This doesn’t mean AI systems are actually able to work out how we’re feeling. But it does demonstrate that these models are performing better and better in experiments designed to assess abilities that psychologists believe are unique to humans. To learn more about the processes behind LLMs’ successes and failures in these tasks, the researchers wanted to apply the same systematic approach they use to test theory of mind in humans.

In theory, the better AI models are at mimicking humans, the more useful and empathetic they can seem in their interactions with us. Both OpenAI and Google announced supercharged AI assistants last week; GPT-4o and Astra are designed to deliver much smoother, more naturalistic responses than their predecessors. But we must avoid falling into the trap of believing that their abilities are humanlike, even if they appear that way. 

“We have a natural tendency to attribute mental states and mind and intentionality to entities that do not have a mind,” says Cristina Becchio, a professor of neuroscience at the University Medical Center Hamburg-Eppendorf, who worked on the research. “The risk of attributing a theory of mind to large language models is there.”

Theory of mind is a hallmark of emotional and social intelligence that allows us to infer people’s intentions and engage and empathize with one another. Most children pick up these kinds of skills between three and five years of age. 

The researchers tested two families of large language models, OpenAI’s GPT-3.5 and GPT-4 and three versions of Meta’s Llama, on tasks designed to test the theory of mind in humans, including identifying false beliefs, recognizing faux pas, and understanding what is being implied rather than said directly. They also tested 1,907 human participants in order to compare the sets of scores.

The team conducted five types of tests. The first, the hinting task, is designed to measure someone’s ability to infer someone else’s real intentions through indirect comments. The second, the false-belief task, assesses whether someone can infer that someone else might reasonably be expected to believe something they happen to know isn’t the case. Another test measured the ability to recognize when someone is making a faux pas, while a fourth test consisted of telling strange stories, in which a protagonist does something unusual, in order to assess whether someone can explain the contrast between what was said and what was meant. They also included a test of whether people can comprehend irony. 

The AI models were given each test 15 times in separate chats, so that they would treat each request independently, and their responses were scored in the same manner used for humans. The researchers then tested the human volunteers, and the two sets of scores were compared. 

Both versions of GPT performed at, or sometimes above, human averages in tasks that involved indirect requests, misdirection, and false beliefs, while GPT-4 outperformed humans in the irony, hinting, and strange stories tests. Llama 2’s three models performed below the human average.

However, Llama 2, the biggest of the three Meta models tested, outperformed humans when it came to recognizing faux pas scenarios, whereas GPT consistently provided incorrect responses. The authors believe this is due to GPT’s general aversion to generating conclusions about opinions, because the models largely responded that there wasn’t enough information for them to answer one way or another.

“These models aren’t demonstrating the theory of mind of a human, for sure,” he says. “But what we do show is that there’s a competence here for arriving at mentalistic inferences and reasoning about characters’ or people’s minds.”

One reason the LLMs may have performed as well as they did was that these psychological tests are so well established, and were therefore likely to have been included in their training data, says Maarten Sap, an assistant professor at Carnegie Mellon University, who did not work on the research. “It’s really important to acknowledge that when you administer a false-belief test to a child, they have probably never seen that exact test before, but language models might,” he says.

Ultimately, we still don’t understand how LLMs work. Research like this can help deepen our understanding of what these kinds of models can and cannot do, says Tomer Ullman, a cognitive scientist at Harvard University, who did not work on the project. But it’s important to bear in mind what we’re really measuring when we set LLMs tests like these. If an AI outperforms a human on a test designed to measure theory of mind, it does not mean that AI has theory of mind.
“I’m not anti-benchmark, but I am part of a group of people who are concerned that we’re currently reaching the end of usefulness in the way that we’ve been using benchmarks,” Ullman says. “However this thing learned to pass the benchmark, it’s not— I don’t think—in a human-like way.”