Have AI chatbots developed theory of mind? What we know and don’t know.

Mind reading is common among us humans. Not the way psychics claim to do, by accessing the warm streams of consciousness that fill each individual’s experience, or the way mentalists claim to do, by blasting a thought from your head at will. Everyday mind reading is more subtle: we observe people’s faces and movements, listen to their words, and then decide or guess what might be going through their heads.

Among psychologists, such intuitive psychology – the ability to attribute mental states to other people different from our own – is called theory of mind, and its absence or deficiency has been linked to autism, schizophrenia and other developmental disorders. Theory of mind helps us communicate and understand ourselves; it allows us to appreciate literature and movies, play games, and make sense of our social environment. In many ways, ability is an essential part of being human.

What if a machine could also read minds?

Recently, Michal Kosinski, a psychologist at the Stanford Graduate School of Business, made this argument: that large language models like ChatGPT and OpenAI’s GPT-4 – next-word prediction machines trained on large amounts of text from of the Internet – have developed a theory. of mind. His studies have not been peer-reviewed, but they have sparked scrutiny and conversations among cognitive scientists, who have tried to answer the often asked question these days – Can ChatGPT do This? — and move it into the realm of more robust scientific investigation. What are the capabilities of these patterns and how might they change our understanding of our own minds?

“Psychologists wouldn’t accept any claims about young children’s abilities based solely on anecdotes about your interactions with them, which seems to happen with ChatGPT,” said Alison Gopnik, a psychologist at the University of California, Berkeley. and one of the first researchers to look at the theory of mind in the 1980s. “You have to do pretty careful and rigorous testing.”

Dr. Kosinski’s previous research has shown that neural networks trained to analyze facial features like nose shape, head angle, and emotional expression can predict people’s political views and sexual orientation with a surprising degree of accuracy (about 72% in the first case and about 80% in the first case). the second case). His recent work on large language patterns uses the classic theory of tests of the mind that measure children’s ability to attribute false beliefs to other people.

A famous example is the Sally-Anne test, in which a girl, Anne, moves a marble from a basket to a box while another girl, Sally, is not watching. To know where Sally will look for the marble, the researchers say, a viewer would have to exercise theory of mind, reason about perceptual evidence, and Sally’s belief formation: Sally didn’t see Anne move the marble in the box, so she still believes in it. That’s where she left it for the last time, in the basket.

Dr. Kosinski presented 10 major language patterns with 40 unique variations of these theory of mind tests – descriptions of situations like the Sally-Anne test, in which a person (Sally) forms a false belief. Then he asked the models questions about these situations, prompting them to see if they would assign false beliefs to the characters involved and accurately predict their behavior. He found that GPT-3.5, released in November 2022, did so 90% of the time, and GPT-4, released in March 2023, did so 95% of the time.

The conclusion? Machines have a theory of mind.

But soon after those results were published, Tomer Ullman, a psychologist at Harvard University, responded with a series of his own experiments, showing that small tweaks in prompts can completely change the responses generated by even large ones. more sophisticated language models. If a container was described as transparent, the machines would fail to deduce that someone could see inside. The machines had difficulty taking into account the testimonies of people in these situations, and sometimes could not distinguish between an object being inside a container and being on top of it.

Maarten Sap, a computer scientist at Carnegie Mellon University, introduced more than 1,000 theory-of-mind tests into large language models and found that the most advanced processors, like ChatGPT and GPT-4, only passed about 70% of the time. (In other words, they were 70% successful in assigning false beliefs to the people described in the test situations.) The discrepancy between his data and Dr. Kosinski’s might come down to differences in the tests, but Dr. Sap said that even passing 95% of the time would not be proof of a true theory of mind. Machines typically fail systematically, unable to engage in abstract reasoning and often making “spurious correlations”, he said.

Dr. Ullman noted that machine learning researchers have struggled over the past two decades to capture the flexibility of human knowledge in computer models. This difficulty has been a “shadow discovery,” he said, hanging behind every exciting innovation. Researchers have shown that language models often give wrong or irrelevant answers when primed with unnecessary information before a question is asked; some chatbots have been so baffled by hypothetical discussions of talking birds that they eventually claim that birds can talk. Because their reasoning is sensitive to small changes in their inputs, scientists have called knowledge of these machines “fragile.”

Dr. Gopnik compared the theory of mind of large language patterns to his own understanding of general relativity. “I’ve read enough to know what the words are,” she said. “But if you asked me to make a new prediction or say what Einstein’s theory tells us about a new phenomenon, I would be puzzled because I don’t really have the theory in mind.” In contrast, she says, the human theory of mind is tied to other commonsense reasoning mechanisms; it stands strong in the face of scrutiny.

In general, Dr. Kosinski’s work and the answers to it are part of the debate over whether the capabilities of these machines can be compared to the capabilities of humans – a debate that divides researchers working on the natural language processing. Are these machines stochastic parrots, or extraterrestrial intelligences, or fraudulent crooks? A 2022 field survey found that, of the 480 researchers who responded, 51% believed that large language models could possibly “understand natural language in a non-trivial sense”, and 49% believed that they could not. couldn’t.

Dr. Ullman does not discount the possibility of machine understanding or machine mind theory, but he is wary of attributing human abilities to non-human things. He noted a famous 1944 study by Fritz Heider and Marianne Simmel, in which participants saw an animated film of two triangles and a circle interacting. When the subjects were asked to write down what happened in the film, almost all described the shapes as people.

“Lovers in the two-dimensional world, no doubt; little triangle number two and sweet circle,” one participant wrote. “Triangle-one (hereafter known as villain) spies on young love. Ah!”

It is natural and often socially required to explain human behavior by talking about beliefs, desires, intentions and thoughts. This tendency is at the core of who we are – so central that we sometimes try to read minds of things that have no minds, at least not minds like ours.