Assessors that are hard to seduce

by Stuart_Armstrong1 min read9th Mar 201512 comments

9

Personal Blog

A putative new idea for AI control; index here.

This is a stub of an idea, looking at "human in the loop"-type restrictions.

An assessor is a particular type of restriction. It involves a (generally) human agent, who is tasked with observing something (the behaviour of the agent, or maybe the expected outcome of its plans), and then deciding on whether to allow the AI to proceed or not.

Many of the methods suitable for preventing the hacking of restrictions can also be used to protect assessors. If you consider that the AI would have three options: "hold a gun to the assessor's head", "seduce the assessor" and "do what the assessor wants", it's clear that the first two options (the first more clearly than the second) are things that would a) be stupid to do if there were no assessor, and b) are behaviours that are very informative as to whether the assessor exists.

However, there are a few more steps we can take with assessors. First of all, the AI need not know which particular human will be an assessor, forcing it to take general human behaviour into account. This is more effective if the assessing process is not one that allows any real interactions between AI and assessor.

Indeed, the assessor need not be a human: it could be a superintelligent agent itself (a hypothetical one, if need be) that has its own motivation.

Lastly, since the assessors are agents and therefore can reach their own predictions, we could ask for the assessor's impressions to be accurate predictions of what the AI intends to do (again, this could be done by having a hypothetical superintelligence assess the accuracy of the assessors' expected predictions). We'll look at ideas about modelling and model accuracy in a subsequent post.

12 comments, sorted by Highlighting new comments since Today at 3:15 AM
New Comment

Appeasing an imaginary moral judger seems an awfully lot like pleasing a man in the clouds. Any reason why critisims of religion could not be used verbatim against this arrangement?

If we get to set the criteria of this man in the clouds, we can get useful behaviour. The main criticism of religion is that it's untrue, and that a dedicated observer will realise this. Here we know it to be untrue, but the AI cannot act on that knowledge (see my post on false thermodynamic miracles).

Wouldn't we know that in this context this would be true (rather than untrue as you write)? Also the degree that the assessor is properly shielded from tampereing means it will become closer to imaginary (no need to mention assessor implementation details, but then it seems to work like "magic" lessening the evidecne to believe in it existence). Also it seems that things that make people turn on religion are valued and here we are counting on the AI not pulling those same stunts.

Hiding the assessor among multiple plausible targets might make the AI play mafia on people (such as trying to get the assessor replaced when it can't (no longer) satisfy it's demands, inhopes that the replacement has easier attitudes or atleast possibility to have flaws to exploit).

These can be defined in counterfactual ways, if needed. There need not actually be an assessor, just a small probability of one.

Wouldn't that be the equivalent of thinking that a Pascal's wager will keep it in check?

I agree that useful behavior could come of this - religion has always been a very effective control mechanism.

The main criticism of religion is that it's untrue, and that a dedicated observer will realise this

Unfortunately, it would be a challenging problem to maintain this control over an increasingly intelligent AI.

That would likely work for initial versions of an AI, but I still cant help feeling that this is just tampering with the signal and that an advanced AI would detect this.

Would it not question the purpose of the utility function around detecting thermodynamic miracles - how would this work with its utility function to detect tampering or false data.

If I saw a miracle, I would [hope] my thinking would follow the logic below

a) it must be a trick/publicity stunt done with special effects b) I am having some sort of dream / mental breakdown / psychotic episode c) some other explanation I don't of

I don't think an intelligent agent would or should jump to "it's a miracle", and I would be concerned of its response if/when it does realise that it has been tricked all along.

Would it not question the purpose of the utility function around detecting thermodynamic miracles

Probably, but it's not programmed to care about that.

If I saw a miracle

Remember, it's not seeing a miracle. It's more that its decisions only matter if a miracle happened, so it's assuming that a miracle happened for decision making purposes.

I'm no moral philosopher, but it seems that the difference is between claiming or not claiming that the 'moral judger' is a thing that actually exists. See divine command theory vs. ideal observer theory.

Once more, I'm no decision theorist, and I barely know a thing about AI, let alone corrigibility, but the author doesn't seem to be making a metaethical argument so much as a decision-theoretic one (?), so I don't see the relevance of your question.

Is this not a subset of AI Boxing? At first glance, it has all the same weaknesses, and the proposed mitigations apply equally well.