Following the recent hype over the potential of a machine passing of the Turing test, Adam Ford interviews Stuart Armstrong (me) of the FHI about the meaning of the test, how we can expect a future of many upcoming "Turing test passings" according to varying criteria of strictness, and how and why we test for intelligence in the first place.

I predict that we are entering an era where "X passed the Turing test" will be a more and more common announcement, followed by long discussions as to whether that was a true pass or not.

So the Turing test has been "passed", and the general consensus is that this was achieved in a very unimpressive way - the 13 year old Ukrainian persona was a cheat, the judges were incompetent, etc... These are all true, though the test did pass Turing's original criteria - and there are far more people willing to be dismissive of those criteria in retrospect than were in advance. It happened about 14 years later than Turing had been anticipating, which makes it quite a good prediction for 1950 (in my personal view, Turing made two mistakes that compensated - the "average interrogator" was a much lower bar than he thought, but progress on the subject would be much slower than he thought).

But anyway, the main goal now, as suggested by Toby Ord and others, is to design a better Turing test, something that can give AI designers something to aim at, and that would be a meaningful test of abilities. The aim is to ensure that if a program passes these new tests, we won't be dismissive of how it was achieved.

Here are a few suggestions I've heard about or thought about recently; can people suggest more and better ideas?

1. Use proper control groups. 30% of judges thinking that a program is human is meaningless unless the judges also compare with actual humans. Pair up a human subject with a program, and the role of the judge is to establish which of the two subjects is the human and which is not.
2. Toss out the persona tricks - no 13 year-olds, nobody with poor English skills. It was informative about human psychology that these tricks work, but we shouldn't allow them in future. All human subjects will have adequate English and typing skills.
3. On that subject, make sure the judges and subjects are properly motivated (financial rewards, prizes, prestige...) to detect or appear human. We should also brief them that our usual conversational approach to establish which kind of human they are dealing with, is not useful for determining whether they are dealing with a human at all.
4. Use only elite judges. For instance, if Scott Aaronson can't figure it out, the program must have some competence.
5. Make a collection of generally applicable approaches (such as the Winograd Schemas) available to the judges, while emphasising they will have to come up with their own exact sentences, since anything online could have been used to optimise the program already.
6. My favourite approach is to test the program on a task they were not optimised for. A cheap and easy way of doing that would be to test them on novel ASCII art.

My current method would be the lazy one of simply typing this, then waiting, arms folded:

"If you want to prove you're human, simply do nothing for 4 minutes, then re-type this sentence I've just written here, skipping one word out of 2".

Thoughts inspired by Yvain's philosophical role-playing post.

Thomas Nagel produced a famous philosophical thought experiment "What Is It Like to Be A Bat?" In it, he argued that the reductionist understanding of consciousness was insufficient, since there exists beings - bats - that have conscious experiences that humans cannot understand. We cannot know what "it is like to be a bat", and looking reductively at bat brains, bat neurones, or the laws of physics, cannot (allegedly) grant us any understanding of this subjective experience. Therefore there remains an unavoidable subjective component to the problem of consciousness.

I won't address this issue directly (see for instance this, on the closely related subject of qualia), but instead look at the question: suppose someone told us that they actually knew what it was like to be a bat (as well as what it was like to be a human). Call such a being a vampire, for obvious reasons. So if someone claimed they were a vampire, how would we test this?

We can't simply ask them to describe what it's like to be a bat - it's perfectly possible they know what it's like to be a bat, but cannot describe it in human terms (just as we often fail to describe certain types of experiences to those who haven't experienced them). Could we run a sort of Turing test - maybe implant the putative vampire's brain into a bat body, and see how bat-like it behaved? But, as Nagel pointed out, this could be a test of whether they know how to behave like a bat behaves, not whether they know what it's like to be a bat.

I posit that one possible solution is to use the approach laid out in my post "the flawed Turing test". We need to pay attention as to how the "vampire" got their knowledge. If the vampire is a renown expert on bat behaviour and social interactions, who is also interested in sonar and paragliding - then them functioning as a bat is weak evidence as to them actually knowing what it is like to be a bat. But suppose instead that their knowledge comes from another source - maybe the vampire is a renown brain expert, who has grappled with philosophy of mind and spent many years examining the functioning of bat brains. But, crucially, they have never seen a full living bat in the wild or in the lab, they've never watched a natural documentary on bats, they've never even seen a photo of a bat. In that case, if they behave correctly when transplanted into a bat body, then it's strong evidence of them actually understanding what it's like to be a bat.

Similarly, maybe they got their knowledge after a long conversation with another "vampire". We have the recording of the conversation, and it's all about mental states, imagery, emotional descriptions and visualisation exercises - but not about physical descriptions or bat behaviour. In that case, as above, if they can function successfully as a bat, this is evidence of them really "getting it".

In summary, we can say "that person likely knows what it is like to be a bat" if "knowing what it's like to be a bat" is the most likely explanation for what we see. If they behave exactly like a bat when in a bat body, and we know they have no prior experience that teaches them how to behave like a bat (but a lot about the bat's mental states), then we can conclude that it's likely that they genuinely know what it's like to be a bat, and are implementing this knowledge, rather than imitating behaviour.

It's been a productive conversation on my post criticising the Turing test. I claimed that I wouldn't take the Turing test as definitive evidence of general intelligence if the agent was specifically optimised on the test. I was challenged as to whether I had a different definition of thinking than "able to pass the Turing test". As a consequence of that exchange, I think I do.

Truly general intelligence is impossible, because of various "no free lunch" theorems, that demonstrate that no algorithm can perform well in every environment (intuitively, this makes sense: a smarter being could always design an environment that specifically penalises a particular algorithm). Nevertheless, we have the intuitive definition of a general intelligence as one that performs well in most (or almost all) environments.

I'd like to reverse that definition, and define a general intelligence as one that doesn't perform stupidly in a novel environment. A small change of emphasis, but it gets to the heart of what the Turing test is meant to do, and why I questioned it. The idea of the Turing test is to catch the (putative) AGI performing stupidly. Since we can't test the AGI on every environment, the idea is to have the Turing test be as general as possible in potential. If you give me the questions in advance, I can certainly craft an algorithm that aces that test; similarly, you can construct an AGI that would ace any given Turing test. But since the space of reasonable conversations is combinatorially huge, and since the judge could potentially pick any element from within that, the AGI could not just have a narrow list of responses: it would have to be genuinely generally intelligent, so that it would not end up being stupid on the particular conversation it was in.

That's the theory, anyway. But maybe the space of conversations isn't as vast as all that, especially if the AGI has some simple classification algorithms. Maybe the data on the internet today, combined with some reasonably cunning algorithms, can carry a conversation as well as a human. After all, we are generating examples of conversations by the millions every hour of every day.

Which is why I emphasised testing from outside the domain of competence of the AGI. You need to introduce it to a novel environment, and give it the possibility of being stupid. If the space of human conversations isn't large enough, you need to move to the much larger space of real-world problem solving - and pick something from it. It doesn't matter what it is, simply that you have the potential of picking anything. Hence only a general intelligence could be confident, in advance, of coping with it. That's why I emphasised not saying what your test was going to be, and changing the rules or outright cheating: the less restrictions you allow on the potential test, the more informative the actual test is.

A related question, of course, is whether humans are generally intelligent. Well, humans are stupid in a lot of domains. Human groups augmented by data and computing technology, and given enough time, are much more generally intelligent that individual humans. So general intelligence is a matter of degree, not a binary classification (though it might be nearly binary for some AGI designs). Thus whether you call humans generally intelligent is a matter of taste and emphasis.

Suppose I have an exact simulation of a human. Feeling ambitious, I decide to print out a GLUT of the action this human will take in every circumstance; while the simulation of course works at the level of quarks, I have a different program that takes lists of quark movements and translates them into a suitably high-level language, such as "Confronted with the evidence that his wife is also his mother, the subject will blind himself and abdicate".

Now, one possible situation is "The subject is confronted with the evidence that his wife is also his mother, and additionally with the fact that this GLUT predicts he will do X". Is it clear that an accurate X exists? In high-level language, I would say that, whatever the prediction is, the subject may choose to do something different. More formally we can notice that the simulation is now self-referential: Part of the result is to be used as the input to the calculation, and therefore affects the result. It is not obvious to me that a self-consistent solution necessarily exists.

It seems to me that this is somehow reminiscent of the Halting Problem, and can perhaps be reduced to it. That is, it may be possible to show that an algorithm that can produce X for arbitrary Turing machines would also be a Halting Oracle. If so, this seems to say something interesting about limitations on what a simulation can do, but I'm not sure exactly what.