Social Intelligence
Theory of Mind (ToM): A fundamental element of social intelligence and consciousness, and a key topic for AGI
Theory of mind refers to the capacity to understand other individuals by ascribing mental states to them. A theory of mind includes the understanding that others' beliefs, desires, intentions, emotions, and thoughts may be different from one's own.
The canonical representation of this, which is pretty much the base case of ToM, is the false belief test (though I like to call it the false reasonable belief test), also called the Sally-Anne Test. It is very easy to understand: “Sally has a basket. Anne has a box. Sally has a marble. She puts the marble into her basket. Sally goes out for a walk. Anne takes the marble out of the basket and puts it into the box. Now Sally comes back. She wants to play with her marble. Where will Sally look for the marble?”
Simon Baron-Cohen (Borat’s brother), et. al. (1985).
Human children cannot pass this test until the age of four.
The following neuroeconomics paper (the authors include neuroscientists such as Sebastian Seung) describes an intriguing way of increasing the accuracy of crowdsourcing by a method that is a sort of ToM for social networks. In essence, it states that the more accurate opinions come from people who can cogently explain why others have a reasonable false belief.
Better wisdom from crowds | MIT News
Prelec, D., Seung, H. & McCoy, J. A solution to the single-question crowd wisdom problem. Nature 541, 532–535 (2017). https://doi.org/10.1038/nature21054
ToM is a fundamental component of social intelligence and consciousness, which may not be absolutely necessary for AGI, but cannot be ignored as possible components.
And since the essence of ToM is captured by the Sally-Anne Test, I believe that using it as a framework in neuroscience and finding a way to inject it into AI systems.
ToM <> Social Intelligence <> <> Consciousness <> Empathy / Values Alignment
Which one of the above is necessary for any of the others is up for debate, but ToM
Emotional Intelligence & Agentic Drive
Artificial general intelligence is a goal, or at least a guiding north star, of artificial intelligence (AI) researchers. I propose that general intelligence is not a binary property (all or none) but exists on a spectrum from highly-specialized, narrow intelligence(s), where the general intelligence is zero or close to zero, to a high-level of general intelligence. The more flexible and adaptive to many situations an intelligence is, the higher its general intelligence rating would be. I also propose that general intelligence is an emergent property of one or more narrow intelligences, which is how human general intelligence evolved in nature, and points to how an artificial general (AGI) intelligence may also eventually evolve out of computer science research efforts.
While emotion imparts meaning and value to our experience and our lives, and likely is an essential part of what makes us human, that is a topic for another essay, or many essays. What I am referring to here is the role that emotion plays in synthesizing information, allowing us to make decisions and to take action. While science does not yet have a robust understanding of how this works, we do have striking examples of people who have emotional deficits that result in intellectual impairments, impairments that I would characterize as deficits in general intelligence.
In the 1980s, the neuroscientist Antonio Damasio (author of Descartes’ Error) worked with a patient named Elliot. Elliot was by all accounts a well-adjusted, productive member of society. Elliot was respected by his colleagues, friends, and was a dutiful father and husband. In his 30s, Elliot developed a benign brain tumor that required surgery, which involved removal of part of his orbitofrontal cortex. The surgery was a success and Elliot seemed to recover well. In the short-term, it was not apparent that he had any neurological deficit. His intellectual faculties, which always had been strong, remained intact. He continued to score well on tests of memory, spatial relations, sorting of information, and understanding words and numbers. However, Elliott’s life quickly fell apart, and he lost both his job and his marriage.
While it might not have been immediately apparent upon meeting him, the deficit that Elliot suffered from was that his life had become completely devoid of emotional content. While discussing events and situations, whether sad or disturbing or laden with some other meaning that one would normally easily see in them, Elliot remained flat and emotionless. And while this deficit of course caused problems with his personal relationships due to lack of empathy and other prosocial behaviors, it caused even more basic problems. For instance, while he was still able to execute individual cognitive tasks, such as the sorting of documents, he was unable to contextualize each task and assign it an appropriate weight and importance. He might spend all day or all week sorting unimportant documents while ignoring much more important things. And this deficit, this lack of general intelligence, profoundly impaired his ability to navigate life.
Damasio discusses at length how human beings rely on emotional experience and literal feelings in the body to make decisions— he calls this somatic marking. For example, when deciding whether to buy one brand of headphones over another, we might consider a few different features or each, before arriving at a decision. That decision is ultimately an emotional decision— a gut feeling. Those of us who are more obsessive about headphones and gadgets, or who are more obsessive in general, might list several features, or even dozens of features, before deciding. If one were to extend this process too long, one would recognize this obsessiveness to be a deficit of general intelligence, that is, one would recognize that one is trying to over-optimize this one decision at the expense of others (and our own overall utility and welfare). Were we to completely lack emotional grounding, we might wind executing a task such as playing the game of Go forever, albeit possibly very well.
Damasio tells the story of another patient that illustrates advantages (in terms of narrow intelligence) and disadvantages (in terms of general intelligence). That patient talked to Damasio about driving to see him on an icy day with dangerous road conditions. The patient stated that other people had trouble with the conditions because they were not following “proper, rational procedures." By contrast, this patient handled the roads “calmly and surely” and had no problems. He had not panicked because he could not feel fear, and this lack of fear allowed him to act calmly and safely. In this context, the patient’s emotional impairment was an advantage.
However, when the same patient contemplated scheduling another appointment, he “enumerated reasons for and against each of the two dates: previous engagements, proximity to other engagements, possible meteorological conditions, virtually anything that one could reasonably think about concerning a simple date." Regarding these two tasks, this patient seems to be proficient in one or more narrow intelligence skills but deficient in general intelligence. By analogy, a computer emotion layer, based to some degree on our understanding of human emotions, could be a useful component in building an AGI out of one or several narrow AIs.
In animals, emotion and instinct can be viewed as heuristic devices that synthesize information and lead to decisions and action. The same is true in humans. However, in humans, there is an additional complexity. We have a cognitive faculty, and while this cognitive faculty is not entirely disconnected from emotion, it is capable of processing information independently of emotion to some degree. We humans can do the same mathematical calculation in two very different emotional states such as happy and sad, and in both situations arrive at the same correct answer. This cognitive faculty can be viewed as a narrow intelligence, analogous to ANI. Thus, human beings can be described largely as feeling machines coupled to one or several cognitive processors of narrow intelligence. As Damasio says, human beings are not thinking machines, we are “feeling machines that happen to think”.
While human cognitive processing is sometimes merely epiphenomenal, a wholly epiphenomenal view is too reductive. Our thinking is sometimes epiphenomenal in that it is just an expression of our emotional state, as in purely post-hoc rationalization (making up a rationale for why you did something that is not the actual emotional driver). However, human cognition cannot be totally epiphenomenal— it must have power on its own, otherwise, it would be completely useless, and clearly it is not.
There is no doubt that emotion and instinct drive most behavior in humans, as we are still primarily feeling machines, not thinking machines. However, if we place human cognitive processing (and the more limited cognitive processing that some animals have) at the center of a general intelligence model, I believe that such emotion can be viewed as a layer that integrates the narrow intelligences of our various cognitive faculties into a general intelligence.
Once again, I am not referring to the subset of emotional intelligence that endows us with the ability to understand and to relate to others, e.g., theory of mind and social intelligence. Nor am I referring to self-awareness and consciousness, which also may be necessary for general intelligence, depending on which philosopher, neuroscientist, or computer scientist you ask.
Instead, the emotional faculty that I refer to is the one that helps us navigate our way through the world in terms of cognitive resource allocation and decision making. I propose that a modeling of this emotional faculty can help computer scientists achieve something closer to AGI. It also may be valuable for AI alignment.