Artificially Intelligent

Any mimicry distinguishable from the original is insufficiently advanced.

When Learning a Skill Has Costs

| 306 words

Status: shortform

There’s an argument that people sometimes make that goes: “learning how to lie is bad because then people can’t really trust your word anymore.” I think this is at least somewhat true, but I think it’s interesting why it’s true

The key question is going to be “once you can lie really well, what is the thing that you can’t do anymore?” The thing is something like proving that you you think something, which we might call “becoming legible.” For instance, if I’m bad at lying and someone asks me if I stole the money, me saying “no” is proof that I didn’t steal the money. If I’m good at lying, I can no longer become legible in this particular way, because my words need not reflect what I actually think anymore.

We can also apply the same sort of analysis to transparency/honesty about activities. If it is known that some people keep secrets, but there is no easy way for people to become legible about the fact that they don’t keep secrets, then this creates a haze that makes it harder for people to make decisions. This applies less if the people who are keeping things secret are really public about the fact that they have secrets, so it’s easy to identify who they are (of course, this comes with it’s own secrecy tradeoffs).

Another consideration is that if people could prove that they didn’t have secrets, anyone who doesn’t want the absence of such a proof used against them in the counterfactual worlds where they do have secrets ought not ever prove that they don’t have secrets in general and only ever prove that they don’t have specific secrets. A similar consideration applies to lying, although there are probably more situations in which your counterfactual selves wouldn’t have wanted to lie either.