Re: Evolving AIs and the term "Virus" (was "Who's Afraid of the SI?")

Bryan Moss (
Mon, 17 Aug 1998 22:11:47 +0100

Doug Bailey wrote:

> > Firstly, it would not understand *our*
> > "concept of what a human is", I wouldn't be so
> > arrogant as to assume our concept is "the"
> > concept (some biological chauvinism on your
> > part ;).
> Why couldn't it understand our concept of
> ourselves? It does not have to adopt that
> concept for itself but it could understand how
> we view ourselves.

I no longer have my post, but didn't I say something like, "It could fathom our concept of ourselves, but it would not share that concept"? It's the fact that it doesn't have to share our point of view to understand and interact with us that is the basis of my argument. And if that's the case, an AI that lives among us doesn't have to be hostile like us. The second part of my argument is that because the AI shares a different course of evolution to us, it's less likely to share *our* hostility. This does not mean it can't be dangerous, but it's more likely these dangers would be something we would call "computer error" rather than "malicious intent". And I also believe these computer errors can be contained.

> Besides, it would not have to understand how we
> understand ourselves to inflict harm upon us. We
> don't understand how dolphins view themselves
> but we have certainly inflicted harm upon them.

The idea is that they don't have be like us to act like us, that they can interact with humans without having all the violent impulses and need for power that we have.

> > Well an AI that's goal is to "maximise its
> > information processing rate" sounds like a
> > virus to me.
> [...] Just because we view it as a virus does
> not mean it will not exist at some point, care
> about what we think, or be responsive to our
> demands.

But where does the motivation to care about what we think and be responsive to our goals come from? If we `evolve' an AI that wants to "maximise its information processing rate" at what point does it suddenly realise that humans are the problem and humans must be destroyed. And if I intentionally put those components in to the AI, where do I get them?

> [...] As I've described above, "virus" is a
> perspective judgement. Additionally, a
> completely benevolent AI could become a "virus"
> through its own evolvability or simple intent to
> solve a problem.

Basically what I'm saying is this:

If we create a `good' AI then it is likely to stay `good'. This is because, A) to exhibit behaviour we consider good does not mean the AI has to be identical to us, and B) the probability that `bad' behaviour will emerge from this AI is small. (A and B are related because if the AI is not like us it can have good behaviour without bad behaviour.)

Although simple behaviour that is harmful to us can evolve from an AI that exhibits good behaviour. Complex harmful behaviour would have to evolve with similar initial condition, and in a similar environment, to us. (By complex I mean the sort of cunning AI you see in science fiction, by simple I mean things we'd generally prescribe to "computer error" and will usually be able to control.)

Of course, a `bad' AI can be `evolved' in a lab or software department just as easily as a `good' one. But I'm not addressing purposeful weapons here, just emergent behaviour.