Is there any research work on the usefulness of the concept of "personality" for an AGI?

Question

Humans have been endowed with personalities by nature, and it is not clear (to me at least) if this is a feature or a bug. This has been explored in science fiction by various notions of Borg-like entities. It is my belief that, for narrative reasons, such stories usually end with the humans with their flawed personalities winning in the end.

Are there experts who have analyzed, perhaps mathematically, design criteria for an AI agent with weakly enforced goals (e.g. to maximize reproduction in the human case) in an uncertain environment, and ended up with the answer that a notion of personality is useful?

If there are AI researchers, philosophers (or maybe even science fiction) writers who have examined this question in their work, I would be happy to know about those too.

Answer 1

'Personality' is something of a 'suitcase word' (Minsky) for quite a large collection of (presumably reasonably consistent) observable traits.

It seems clear that there is a certain collective advantage in having a consistent personality - specifically that it affords observers some learning gradient in an otherwise uncertain environment. This is of particular importance because those consistencies might have been arrived at using different learning mechanisms than the ones a given observer has.

Hence, in any non-trivial coevolutionary system, other organisms will inevitably make use of any such consistencies. Consider a simple robot, called Alice, say, that has the trait of 'quickly flashing red when it sees a blue robot'. It makes sense for all observers to exploit everything that they perceive as correlating with Alice's behavior, in particular, the prediction that a blue robot is likely to be present.

The best reference I can recommend on this (which shows that we tend to ascribe 'personality' to even very simple mechanisms) is 'Vehicles' by Valentino Braitenberg.

Answer 2

First, a note on the question itself.

Humans have been endowed with personalities by nature, and it is not clear (to me at least) if this is a feature or a bug.

In my opinion, this is a statement that constrains the question, since it assumes that the personality is given. To me, it feels a bit like playing god: Artificial (given) Intelligence would hence imply Artificial (given) Personality. This approach to the problem seems to be supported by the next fragment:

a notion of personality is useful

I point to the above because I don't think that artificial intelligence... Intelligence itself, actually, need to be given or assigned, or even have a use in the sense of a purpose.

---

The previous note was about emergence, which is a topic that user217281728 briefly addressed in their answer. In this second approach, the particular traits just happen, or develop. The interaction between the (so-called) agents and their environment, as well as fellow agents can give place to new behaviour patterns, not designed beforehand.

In an evolutionary approach, if the personality would happen to have an advantage (or at least not represent a disadvantage), then it could just appear. Of course, I am making a number of assumptions and demarcations here as well:

• I am thinking about embodied intelligence
• I speak of evolutionary robotics
• I think on social issues being of importance
• I assume that personality could emerge

Now, an example that I find extremely interesting is that of the little mobile robots which could move around and end-up in a pool of food or a pool of poison. And they, somehow, by some odd chance, recognised or made a relation between signals sent by other robots, and the presence of food. Or not. That was more or less the thing: Some robots (kind of) learned to conceal information and thus had more time to eat themselves. Well, I would have a couple of personality adjectives for such guys.

Here you find the article and here you find some videos and related stuff.

And with that, we land at my last point: We humans put the adjectives, according to our social conditioning. We call Marvin depressive and R2D2 lovely and charming.

If they perceive their personalities as constructive or damaging, will always depend on our own judgment. In the end, it is quite common under humans to disagree on personality issues, too.

---

Bonus

Remember when HAL got emotional, on the face of death?

It gets human when it loses its cool, before the flawed-personality human astronaut :)

Answer 3

"Usefulness" can only be measured against some purpose. Once you pass AGI - which really means "generally animal-like AI, because it seems general to us" - then you've passed into a world of potentially undefined behavior.

Part of what makes a human free and sets us apart from the other animals is the fact that our purposes, capabilities and possibilities aren't fully defined. We're open ended.

To clarify terms, I interpret "Strong AGI" as "potentially super intelligence, but at least human level."

When we say "Strong AGI" vs just "AGI," we're not saying that one is more open ended than the other. We are saying that the stronger one is simply smarter on some axis.

So to ask whether a particular trait would be "more useful" to a Strong AGI - that would depend on the purpose of the AGI. But here's the catch: if a thing had just one purpose, then the most efficient solution to fulfilling that one purpose will always be a narrow solution, not a general one. When the purpose of the object is known before hand, giving that object more general capability than is necessary for that purpose is counterproductive.

That's why it's impossible to make declarative prescriptions about what a free, open-ended AGI should or shouldn't need. Such prescriptions would nullify the open-ended freedom of utility that its generality implies. We can speak declaratively about lesser robots and animals.

But for any given problem, the solution we will want to find is the most well-defined, narrow, efficient one available - not the most general one.

In other words, sure, personalities could be useful for a Strong AGI, assuming the problems in question involved personalities.