If we see precursors to deception (e.g. non-myopia, self-awareness, etc.) but suspiciously don’t see deception itself, that’s evidence for deception.

Stated like this, it seems to run afoul of the law of conservation of expected evidence [? · GW].  If you see precursors to deception, is it then the case that both (a) seeing deception is evidence for deception and (b) not seeing deception is also evidence for deception?  I don't think so.

The direct patch is "If you see precursors to deception, then you should expect that there is deception, and further evidence should not change your belief on this"—which does seem to be a first approximation to your position.

Above some threshold level of deceptive capabilities we should stop trusting the results of behavioral experiments no matter what they show

I agree, and if we don't know how to verify that we're not being deceived, then we can't trust almost any black-box-measurable behavioral property of extremely intelligent systems, because any such black-box measurement rests on the assumption that the object being measured isn't deliberately deceiving us.

It seem that we ought to be able to do non-black-box stuff, we just don't know how to do that kind of stuff very well yet. In my opinion this is the hard problem of working with highly capable intelligent systems.

Suppose that when all is said and done they see the following:

And my interpretation of that graph is "the rightmost datapoint has been gathered, and the world hasn't ended yet. What the ????"

Deception is only useful to the extent that it achieves the AI's goals. And I think the main way deception helps an AI is in helping it break out and take over the world. Now maybe the AI's on the right have broken out, and are just perfecting their nanobots. But in that case, I would expect the graph, the paper the graph is in, and most of the rest of the internet, to be compromised into whatever tricked humans in an AI convenient direction. 

And "if we see X, that means the datasource has been optimized to make us do Y" Well if we decide not to do Y in response to seeing X, then the optimizer won't show us X. We can say "don't trust the data source, and never ever do Y whatever it tells us." This isn't practical with future superintelligences and the whole internet. 

Impair bigger models (e.g. with dropout or random or adversarial perturbations) and see if they get better at some deceptive capabilities. Impairing models shouldn’t improve  capabilities, so this would be good evidence that the bigger models are capable of deception but choosing to hide that capability.

This experiment is a little tricky to implement, but the core idea is to hope that interventions impair something critical in the model's ability to execute on deception but don't impair all of the deceptive capabilities.

Even if the largest models can continue hiding their capabilities despite the impairment, models that are just barely able to hide their capabilities may fail to do so once impaired, which should increase the model size where deceptive capabilities peak.

I feel like this assumes that deception is a property immediately downstream of pure capabilities, but I can think of a couple cases where deception may arise at a particular point for other reasons - such as in the case of suboptimality deceptive alignment [LW · GW]. 

There are also other considerations that come to mind, such as that a model once deceptive can make it so that it isn't that hard to sustain deception past impairment (especially if it manages to exert some control over its gradient), or whether this might happen naturally once deception is "amplified" as a strategy by gradient descent to the point where we notice it and run the test.

It's definitely possible I'm missing something about the experiment setup that covers these cases, or that you aren't very worried about them in this context, so I'd love to hear your thoughts.

The main idea seems good: if you're in a situation where you think you might be in the process of being deceived by an AI, do not relax when the AI provides great evidence that it is not deceiving you. The primary expected outputs of something really good at deception should be things which don't look like deception.

Some of the things in the post don't seem general enough to me, so I want to try to restate them.

Test 1 I like. If you understand all of the gears, you should understand the machine.

Test 2 I like. Tweak the model in a way that should make it worse at process A, and then check the outputs of process A to see whether they change in ways that reveal biases you can triangulate with past outputs in order to catch deception.

Test 3 scares me. This seems to me to be a test for the very specific form of deception "hide capabilities while in testing but not when deployed" which seems like a small domain in deception space.

Test 4 I kind of like. One expected behavior from a deceptive AI is to hide capabilities that scare people studying them. If you see scary capabilities getting less scary, you should take that as a sign of deception unless you have good reason to believe those capabilities should get worse with scale. Thus it is a good idea to find out which things should get worse with scale ahead of time. I do worry that this paradigm relies too much on AI which improves via "more dakka" (eg more GPUs, larger datasets, better processors, etc) rather than via algorithm improvements or something, in which case I don't know that people will have a good handle on what capabilities will get worse. The "scaling helps" section also worries me for this reason.

In the section "deceptive models know this" you suggest "deciding on a level of deceptive capabilities that’s low enough that we trust models not to be deceptively aligned". Won't that just optimize on things which start deceiving well earlier? I think I may be misinterpreting what you mean by "deceptive capabilities" here. Maybe your "deceptive capabilities" are "smoke" and actual deception is "fire", but I'm not sure what deceptive capabilities that aren't deception are.

How would you measure deceptive alignment capability?

The only current models I can think of with the capability to be deceptive in the usual sense (GPT-N and other Transformer models) tend not to have any incentive to deceive (Gato may be an exception, I'm not sure). LLMs trained with next-token predictive loss shouldn't care about whether they tell the truth or lie, only about what they think a human would write in a given context. In other words, they would try to tell the truth if they thought a human would do so, and they would try to lie if they thought a human would do that.

I think that to measure deceptive capability, you would first need a model that can be incentivised to deceive. In other words, a model with both linguistic understanding and the ability to optimize for certain goals through planning-based RL. Then you would have to devise an experiment where lying to the experimenters (getting them to accept or at least act on a false narrative) is part of an optimal strategy for achieving the AI's goals.

Only then could I see the AI attempting to deceive in a way that matters for alignment research. Of course, you would probably want to airgap the experiment, especially as you start testing larger and larger models.