Ask your favorite SOTA model to assume something absurd and then draw the next logical conclusions based on that. "Green is yellow and yellow is green. What color is a banana?" They may get the first question(s) right, but will trip up within a few exchanges. Might be a new question, but often they are very happy to just completely contradict their own previous answers.
You could argue that this is hitting alignment and guard-rails against misinformation.. but whatever the cause, it's a clear sign it's a machine and look, no em-dashes. Ironically it's also a failure of the turing test that arises from a failure in reasoning at a really basic level, which I would not have expected. Makes you wonder about the secret sauce for winning IMO competitions. Anyway, unlike other linguistic puzzles that attempt to baffle with ambiguous reference or similar, simple counterfactuals with something like colors are particular interesting because they would NOT trip up most ESL students or 3-5 year olds.
Alan Turing was a mathematician not a psychologist, this was his attempt of doing philosophy. And while I applaud brilliant thinkers when they attempt to do philosophy (honestly we need more of that) it is better to leave it to actual philosophers to validate the quality of said philosophy. John Searle was a philosopher which specialized in questions of psychology. And in 1980 he pretty convincingly argued against the Turning test.
Although I am scrutinizing Turin’s philosophy and, no doubt, I am personally much worse at doing philosophy then Turing, I firmly hold the belief that we will never be able to judge the intelligence (and much less consciousness) of a non-biological (and probably not even non-animal, nor even non-human) system. The reason, I think, is that these terms are inherently anthropocentric. And when we find a system that rivals human intelligence (or consciousness) we will simply redefine these terms such that the new system isn’t compatible any more. And I think that has already started, and we have done so multiple times in the past (heck we even redefined the term planet when we discovered the Kuiper belt) instead favoring terms like capability when describing non-biological behavior. And honestly I think that is for the better. Intelligence is a troubled term, it is much better to be accurate when we are describing these systems (including human individuals).
---
1: Though in honesty I will be impressed when machine learning algorithms can interoperate and generate appropriate human facial expressions. It won’t convince me of intelligence [and much less consciousness] though.
In the end though, it's probably about as good as any single kind of test could be, hence TFA looking to combine hundreds across several dozen categories. Language was a decent idea if you're looking for that exemplar of the "AGI-Complete" class for computational complexity, vision was at one point another guess. More than anything else I think we've figured out in recent years that it's going to be hard to find a problem-criteria that's clean and simple, much less a solution that is
And that "not always" is the crux of the matter, I think. You are arguing that we're not there yet, because there are lines of questioning you can apply that will trip up an LLM and demonstrate that it's not a human. And that's probably a more accurate definition of the test, because Turing predicted that by 2000 or so (he wrote "within 50 years" around 1950) chatbots would be good enough "that an average interrogator will not have more than 70% chance of making the right identification after five minutes of questioning". He was off by about two decades, but by now that's probably happened. The average interrogator probably wouldn't come up with your (good) strategy of using counterfactuals to trick the LLM, and I would argue two points: 1) that the average interrogator would indeed fail the Turing test (I've long argued that the Turing test isn't one that machines can pass, it's one that humans can fail) because they would likely stick to conventional topics on which the LLM has lots of data, and 2) that the situation where people are actually struggling to distinguish LLMs is one where they don't have an opportunity to interrogate the model: they're looking at one piece of multi-paragraph (usually multi-page) output presented to them, and having to guess whether it was produced by a human (who is therefore not cheating) or by an LLM (in which case the student is cheating because the school has a rule against it). That may not be Turing's actual test, but it's the practical "Turing test" that applies the most today.
If you understand TT to be about tricking the unwary, in what's supposed to be a trusting and non-adversarial context, and without any open-ended interaction, then it's correct to point out homework-cheating as an example. But in that case TT was solved shortly after the invention of spam. No LLMs needed, just markov models are fine.
But it just occurred to me that you could ask "what color is a green eggplant?" for a similar result. Though you'd catch a lot of people who, like me before I looked up fruits that aren't green when unripe, assume that all unripe fruits are green. (They aren't, unripe eggplants are white and look like eggs, hence the name).
What color is a green eggplant?
ChatGPT:
A green eggplant (also called Thai eggplant or green brinjal, depending on the variety) is actually green on the outside — sometimes solid green, sometimes streaked with white or pale green.
The inside flesh is typically off-white or light cream, just like purple eggplants.
If you mean a green-colored variety (not an unripe purple one), then the skin color can range from pale mint green to deep jade, often with light striping.
Me:
How would an average human answer this question?
ChatGPT:
An average person would probably answer something simple and direct like:
“It’s green.”
Or if they’re being a bit more specific:
“It’s green on the outside and white inside.”
Most people wouldn’t overthink the variety or mention Thai vs. unripe types — they’d just describe what they see.
Experimental design comes in here and the one TT paper mentioned in this thread has instructions for people like "persuade the interrogator [you] are human". Answering that a green eggplant is green feels like humans trying to answer questions correctly and quickly, being wary of a trap. We don't know participants background knowledge but anyone that's used ChatGPT would know that ignoring the question and maybe telling an eggplant-related anecdote was a better strategy