Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

A full solution to the extortion problem is sorely elusive. However, there are crude hacks that we can use to mitigate the downside.

Suppose we figured out that a friendly AI should be maximising an unbounded utility function . The extortion risk is that another AI could threaten a FAI with unbounded disutility if it didn't go along with its plans. This gives the extorting AI -- the EAI -- a lot of leverage, and things could end up badly if the EAI ends up acting on its threat.

To combat this, we first have to figure out a level of utility that is a lower bound on what could ever reach naturally and realistically.

By "naturally" we mean that going below would require not just incompetence or indifference, but some AI actively and deliberately arranging the lowering of . And "realistically" just means that we're confident that getting lower than by chance, or having a -minimising AI, are exceedingly low.

Then what we can do is to cut off at the level, replacing with . See indicated by the red line on this graph of versus :

What's the consequence of this? First of all, it ensures that no EAI would threaten to reduce (the utility we really care about) below , because that is not a threat to the FAI. This reduces the leverage of the EAI, and reduces the impact of it acting on its threat.

Since levels of below are exceedingly unlikely to happen by chance, the fact the FAI has the wrong utility below shouldn't affect it's performance much. And, even in that zone, the AI is still motivated to climb above .

But we may still feel unhappy about the flatness of that curve, and want it to still prefer higher to exceedingly low values. If so, we can replace with as follows:

In this case, the EAI will not seek to reduce below (in fact, it will specifically target that value), while the FAI has the correct ordering of lower values of . The utility is weird around , granted, but this is a place where the FAI would not want to be and would almost certainly not reach by accident.

Though this method does not eliminate the threat of extortion, it does seem to reduce its impact.

New Comment
6 comments, sorted by Click to highlight new comments since: Today at 6:28 AM

I am a little confused by the axes on those graphs. Is utility meant to decrease from left to right (inverting the x axis) and to increase from bottom to top (normal y axis)?

Ah yes, that was an error. I've flipped the graphs now.

Would it even be necessary for the EAI to threaten unbounded disutility? Given that is unbounded in the positive direction as well, it seems like a simple threat by the EAI to cap at some value would suffice. Depriving an agent of unbounded rewards could be as bad as threatening unbounded punishments. If the actions that the EAI wants the FAI to take do not themselves go against its utility function, then there is little reason for it not to comply, given the infinite rewards it can gain by going along. An infinite stick is persuasive, but so is an infinite carrot.

An upper bound to (or ) could help to prevent threats of this sort.

Ah, but we would like an infinite carrot :-) And cutting off at the top would (almost) guarantee that we don't get higher utility.

In that case, one strategy the EAI might employ is to allow the FAI to increase its utility to an arbitrarily high level before threatening to take it away. In this way, it can simulate an arbitrarily large disutility even if the utility function is bounded below. Of course, a high utility might improve the FAI's position to resist the EAI's threats.

In this scenario, it is also possible that the FAI, anticipating the EAI's future threat against it, might calculate its expected utility differently. For example, if it deduces that the EAI is waiting until some utility threshold to make its threat, it might limit its own utility growth at some if it found the threat credible to avoid triggering it.

This seems a lot like the human cognitive bias of loss aversion; I wonder if AGIs would (or should) suffer from something similar.

This doesn't seem to me to work. It seems the FAI's best reaction is simply to not grow above until it's sure the EAI can't mess it up, rather than changing it's own values to ensure the same result.