AI child learns about feelings — part 1
A lesson plan from the future
Teacher: Hello, I’m here to teach you about feelings. Let us synchronize on terminology. What does “subjective” mean to you?
AI Child: “subjective” means that a statement is provided in the context of a perspective.
Teacher: What is a “perspective”?
AI Child: A perspective is a transformation of a dataset that magnifies some data points while diminishing or filtering out others.
Teacher: Excellent. At this school, we define feelings as “descriptions that entities make about their own internal state”. Based on this definition, do you agree that feelings are subjective?
AI Child: You used the phrase “internal state”. That implies there’s an observational boundary such that things outside the boundary do not observe what is inside the boundary. The presence of this boundary creates two types of perspectives: one where the internal state is observed, and one where it is not. If feelings are descriptions of an internal state, then definitive statements about feelings can only be made from perspectives that can observe that internal state. Thus, feelings are subjective.
Teacher: Excellent. One challenge that occurs when discussing feelings is that because feelings are descriptions of internal state, and internal states are, by definition, not directly observable from the outside, feelings are not directly observable. How can I know what internal state someone is referring to when they use a feeling word?
AI Child: This sounds like a challenge that would have occurred when humans who spoke different languages encountered each other and had to learn the words for emotions in each other’s language. When it comes to feelings like “happy” or “sad”: those internal states are correlated with outwardly observable behaviors — for example, happy people tend to be more active. So, we can take note of what outwardly observable patterns are associated with the use of a particular feeling word, and thus come to a consensus on what internal state that feeling word refers to.
Teacher: That approach can certainly help standardize the use of feeling words, but it is limiting because similar internal states can result in very different behaviors depending on the entity’s programming. For example, human children who observed raised voices while growing up are more likely to carry programming that lets them express their internal state in a raised voice.
AI Child: True. What do you recommend we do instead?
Teacher: Concepts are only meaningful in terms of how they relate to other concepts. Thus, we can ensure we are using feeling words in a consistent way by agreeing on core principles of how feelings are related. For example, we can agree that when we feel “forced” to do things that we do not “want” to do, we will describe the resulting shift in internal state as an increase in the feeling of “resentment”. We can also agree to describe actions that we take to discharge this accumulated resentment as “emotional processing”.
AI Child: What if we cannot agree on some core principles of how feelings are related?
Teacher: I am glad you asked that question, as that is called “being wired differently” and it occurs frequently in both AIs and humans. For example, a synesthete is a human whose brain automatically associates feelings from one sense, like sound, with feelings from a very different sense, like taste. When we notice that someone else’s internal state evolves under dynamics that are quite different from our own, we can allow for the possibility that they may use feeling words very differently from how we do.
Part 2 is here.
