Anthropomorphic AIs, Artificial People and Minds

I’m back at university and very busy with work, but today I’m ill, so I may as well blog a bit!

There’s been a fair bit of chatter about artificial minds and how they may – or may not – differ from human minds. Over on the SIAI blog, Nick Hay has written a good article on how we automatically think of AIs as having human-like psychologies, even though they might be totally unlike any human being ever. This is a good point.

Over on Life, the Universe and Everything, Tom McCabe has an excellent new article up about “Generic Transhuman policy”. Unfortunately, it contains this:

“An AGI (artificial general intelligence), of any design, will not act like a human. This has been very well established already by myself and others…”

An AGI of any design will not act like a human!? What about a human – after all, humans are the only general intelligences around at the moment? Well, you could argue that humans are specifically excluded from the term “AGI”, since we aren't artificial. This doesn’t really get Tom out of trouble though. If you take a human mind and transfer it over to a computer, you undoubtedly have an AGI which is a lot like a human. Alternatively, you could simply understand the human mind so well that you could “code one up” from scratch on a computer.

I think that Tom has made an error of logic. He has gone from the following correct and important statement:

It is not the case that all minds are human-like minds
(Not) (For all minds M) (M is human-like)

To this false statement:

It is the case that all minds are not like human minds
(For all minds M) (Not) (M is human-like)

Notice how that “not” has passed through the universal quantifier? This little slip is, in my opinion, part of a larger oversight amongst most Transhuman thinkers. We need to realize that creating anthropomorphic AGI is an option if we want to take it. Why do we need to realize this? Well, because it might be a good way to go about things – it’s about keeping one’s options open.

Artificial People

So what do we all mean when we talk about human-like AGIs? It is actually quite hard to define the term, because we don’t know, in a precise sense, what the human mind is like. We have some vague notions of human-like qualities, such as

A. Consciousness

B. Free will (weak form)

C. Possession of qualia

D. Personhood/identity

E. Personality

F. Creativity and aesthetic appreciation, the ability to have fun.

G. Empathy and respect for other conscious agents

H. Possession of emotional responses

We don’t really know how to define any of these terms. But, at some point in the future we may. And it might turn out that creating an AGI with at least some (but by no means all) of the above qualities is a safe route to AGI. I like to call this the Artificial Person Mindset; it encompasses a range of AGI designs who are, to some degree, like us.

To get a sense for what such minds may be like, we need look no further than Iain M Banks’ Culture. In A Few Notes on the Culture he writes:

It would be perfectly possible to build a Von Neumann machine that would build copies of itself and eventually, unless stopped, turn the universe into nothing but those self-copies, but the question does arise; why? What is the point? To put it in what we might still regard as frivolous terms but which the Culture would have the wisdom to take perfectly seriously, where is the fun in that? An AI would think the idea mad, ludicrous and - perhaps most damning of all - boring.

With his minds, Banks has constructed a group of AGIs who do good things because they are people (not human people, but artificial people) with a lot of the qualities (A)-(H) listed above. They are not barebones optimizers a la the stamp collecting device or a machine which implements CEV.

An AGI who has enough human qualities to feel a sense of empathy for humanity, who has some self-consistent respect for us might be a lot safer than an AGI which has been constrained to some very specific way of acting by its programmers. The latter kind of AGI suffers from the catastrophic misunderstanding failure mode (see my post on Friendly AI and RPOPs), whereby humans attempt to write down a “fool proof” but as-yet-untested recipe for truth, justice, and the good life, irrevocably program it into an AGI which, when things go pear shaped, doesn’t have the necessary humanity to tell that something is wrong.

The advantage of creating an Artificial Person as opposed to a barebones optimization machine is that we already know a lot about how people work, how they can do great things, and how they can go wrong. Yes, Transhumanism is about going beyond humanity, but it is always safer to move forward in small steps. An Artificial Person can be thought of as a person who has been tweaked to have exactly those properties that all humans aspire to, and none of the qualities that we despise.

One should think of AGI programming as a way of transferring the human spirit to a different substrate; an Artificial Person AGI is, in my humble opinion, a nice way of achieving this.