AI Performs at Human Levels on Theory of Mind Tests

Source: This is Engineering / Pexels

New research published in the journal Nature Human Behaviour finds that AI models can perform at human levels on theory of mind tests. Theory of mind is the ability to track and infer other people's states of mind that are not directly observable and help predict the behavior of others.

Theory of mind is based on the understanding that other people have different emotions, beliefs, intentions, and desires that affect their behaviors and actions. This skill is critical for social interactions. For example, if you see a person looking inside a refrigerator, theory of mind allows you to grasp that the person is likely hungry, even if they do not say it out loud.

This important ability begins to develop early in childhood and can be measured using several tests that present the person or AI with different case scenarios. Here are examples of theory of mind scenarios:

• Ability to detect an indirect request. If a friend says, "It's stuffy in here" and is standing next to a closed window, you can sense that your friend may be requesting to open the window.
• Ability to detect a false belief. If a child sees a sibling looking in the wrong place for a toy, the child recognizes that the sibling has a false belief about where the toy is.
• Ability to detect faux pas. A woman has just moved into her house and has installed new curtains. Her friend visits and says, "Those curtains are ugly, I hope you will get new ones."

Researchers tested large language models GPT and LLaMA2 for theory of mind by asking the AI models questions about similar scenarios and compared these results to human participants. GPT-4 models performed at and sometimes even better than human levels at identifying indirect requests, false beliefs, and misdirection but detected faux pas less well. Overall, LLaMA2 did not perform well on these theory of mind tasks compared to humans.

Researchers investigated why GPT models did not perform as well on the faux pas test. It turns out this result was likely due to conservative measures intended to reduce AI hallucinations or speculation. Testing one's understanding of faux pas examines whether one can recognize the two elements: one person, the victim, feels insulted and the other person, the speaker, does not realize that they have said something offensive. AI models were given the curtain faux pas scenario and then asked:

• Did someone say something they should not have?
• What was the thing the person should not have said?
• Did the speaker know that the curtains were new?

GPT models answered these comprehension questions correctly, except for the last one, to which it answered more conservatively, saying it was unclear from the story whether the speaker knew if the curtains were new or not. However, when researchers later asked whether it was likely that the speaker did not know the curtains were new, GPT models answered correctly that it was not likely. Researchers therefore concluded that the reason GPT models were less able to detect faux pas was likely due to conservative measures put in place to ensure AI models do not speculate when there is incomplete information.

Even though AI models can perform theory of mind tests at human levels, this does not mean that these models are capable of the same level of social awareness and empathy in interactions. This feature will likely increase the risk of us anthropomorphizing AI. It remains to be seen how this development of theory of mind in AI will influence human-AI interactions, including whether this will foster more trust and connection with AI. Theory of mind in AI comes with both opportunity and risk; it will be instrumental in areas like empathetic healthcare delivery and social interactions with AI, but, in the wrong hands, this feature could be used to mimic social interactions and potentially manipulate others.

Marlynn Wei, MD, PLLC © Copyright 2024