#11 — Theory of Mind

Being able to infer what other people are thinking is a crucial skill in human communication. This cognitive ability is what we call Theory of Mind (ToM). You probably use it everyday without even thinking about it. Someone looks at a coffee mug — will they grab it? Someone looks confused at the fridge — are they looking for the milk? Oops I finished that yesterday, but they didn’t know that. Human interactions make deeply use of ToM. This blog post aims to provide a bit of an introduction to the concept of ToM.

Scene from Friends where Phoebe tries to convey a message with vague words. The cool thing is that humans are often capable of communicating like this because they use all other contextual information to fill in the blanks.

What is ToM?

Mental models

Humans have this mental representation in their mind of how everything works. We call this their mental model. When we drop something, we expect it to fall. When we wave at someone, we expect them to react. When we press a power button, we expect the machine to turn on. This mental model allows us to make quick predictions about the world. It allows us to move around in this maze without having to stop at every fork to contemplate our next move.

Mentalizing

Mentalizing describes the action of using one’s Theory of Mind capabilities. When we communicate with other human beings we are constantly mentalizing. So, what is it exactly that we are inferring about others during interaction? There are many-many-many and it is difficult to really classify all, but here are some examples: empathy, emotion, percepts, knowledge, beliefs, desires, and intentions. Especially these final three (beliefs, desires, and intentions) are often looked at together in the so called BDI model.

Higher orders

It is also possible for humans to make a mental representation of someone else’s mental representation. We hall this a higher-order ToM inference. For example: I know that you are probably right now thinking “wow this is such a cool blogpost”. Humans can comfortably do this up to 3rd order recursive, but higher that becomes tricky.

I have not found any literature about this yet (have not really searched for it as well) but this should be higher when you are very familiar with the people and beliefs discussed. I would find it way easier to think about what Tom thinks that Sarah thinks about Mauritz, than what person A thinks that person B thinks is in the box.

Scene from Friends, where Phoebe is performing recursive Theory of Mind.

Testing for ToM

Test 1: Sally-Anne task

So, how do we test for these ToM-capabilities? One famous test from psychology is the Sally-Anne test, which is used to test for false beliefs. The researcher presents a scene where Sally and Anne are in the same space and they put a ball in a box. When Sally leaves the room, Anne moves the ball from the box to the basket. When Sally returns, where will she look for the ball? When the participant realizes that Sally has a mental model with different beliefs than Anne and the reality, that shows ToM-capabilities.

The test participants are mainly young children and people that are neuro-divergent because that appeared to be an interesting group to test. Research has found that younger children until a certain age have a hard time with seeing the difference between people’s mental models and reality. This also comes forward in the video below:

Video showing Sally-Anne false belief tests with young children.

Test 2: Unexpected content task

Another classic task is that of unexpected content of a box. The video below shows a case where there is a box with the label crayons, but the box is instead filled with candles. This test can be performed with any container where a certain content is expected, but the content appears to be different. The participant is asked what they used to think was in the box (before it was opened). Answering candles shows that they don’t differentiate between their own past mental model and the reality or their current mental model. A variable to take into account here is whose (false) beliefs you are asking about. In the previous test it was about other’s and here it is about their own (past) beliefs.

Next posts

Interesting stuff, but how can we make this more concrete? Is this something that can also be used in human-robot interaction instead of only human-human? (Spoiler: yes) What are the benefits of robot being able to understand humans better? What are the dangers? When is it beneficial for a robot to perform behavior that is easily predictable for humans over optimized behavior?

I can go on about this for hours, but that would defeat the purpose of a daily blog post within a limited amount of time. So for that reason, I will keep the rest for later. What a cliffhanger, I know.