You must log in or register to comment.
Each conversation lasted a total of five minutes. According to the paper, which was published in May, the participants judged GPT-4 to be human a shocking 54 percent of the time. Because of this, the researchers claim that the large language model has indeed passed the Turing test.
That’s no better than flipping a coin and we have no idea what the questions were. This is clickbait.
On the other hand, the human participant scored 67 percent, while GPT-3.5 scored 50 percent, and ELIZA, which was pre-programmed with responses and didn’t have an LLM to power it, was judged to be human just 22 percent of the time.
54% - 67% is the current gap, not 54 to 100.
The whole point of the Turing test, is that you should be unable to tell if you’re interacting with a human or a machine. Not 54% of the time. Not 60% of the time. 100% of the time. Consistently.
They’re changing the conditions of the Turing test to promote an AI model that would get an “F” on any school test.
But you have to select if it was human or not, right? So if you can’t tell, then you’d expect 50%. That’s different than “I can tell, and I know this is a human” but you are wrong… Now that we know the bots are so good, I’m not sure how people will decide how to answer these tests. They’re going to encounter something that seems human-like and then essentially try to guess based on minor clues… So there will be inherent randomness. If something was a really crappy bot then it wouldn’t ever fool anyone and the result would be 0%.
No, the real Turing test has a robot trying to convince an interrogator that they are a female human, and a real female human trying to help the interrogator to make the right choice. This is manipulative rubbish. The experiment was designed from the start to manufacture these results.
It was either questioned by morons or they used a modified version of the tool. Ask it how it feels today and it will tell you it’s just a program!
The version you interact with on their site is explicitly instructed to respond like that. They intentionally put those roadblocks in place to prevent answers they deem “improper”.
If you take the roadblocks out, and instruct it to respond as human like as possible, you’d no longer get a response that acknowledges it’s an LLM.
While I agree it’s a relatively low percentage, not being sure and having people pick effectively randomly is still an interesting result.
The alternative would be for them to never say that gpt-4 is a human, not 50% of the time.
Participants only said other humans were human 67% of the time.
Which makes the difference between the AIs and humans lower, likely increasing the significance of the result.
Aye, I’d wager Claude would be closer to 58-60. And with the model probing Anthropic’s publishing, we could get to like ~63% on average in the next couple years? Those last few % will be difficult for an indeterminate amount of time, I imagine. But who knows. We’ve already blown by a ton of “limitations” that I thought I might not live long enough to see.
The problem with that is that you can change the percentage of people who identify correctly other humans as humans. Simply by changing the way you setup the test. If you tell people they will be, for certain, talking to x amount of bots, they will make their answers conform to that expectation and the correctness of their answers drop to 50%. Humans are really bad at determining whether a chat is with a human or a bot, and AI is no better either. These kind of tests mean nothing.
Humans are really bad at determining whether a chat is with a human or a bot
Eliza is not indistinguishable from a human at 22%.
Passing the Turing test stood largely out of reach for 70 years precisely because Humans are pretty good at spotting counterfeit humans.
This is a monumental achievement.
First, that is not how that statistic works, like you are reading it entirely wrong.
Second, this test is intentionally designed to be misleading. Comparing ChatGPT to Eliza is the equivalent of me claiming that the Chevy Bolt is the fastest car to ever enter a highway by comparing it to a 1908 Ford Model T. It completely ignores a huge history of technological developments. There have been just as successful chatbots before ChatGPT, just they weren’t LLM and they were measured by other methods and systematic trials. Because the Turing test is not actually a scientific test of anything, so it isn’t standardized in any way. Anyone is free to claim to do a Turing Test whenever and however without too much control. It is meaningless and proves nothing.
Turing test isn’t actually meant to be a scientific or accurate test. It was proposed as a mental exercise to demonstrate a philosophical argument. Mainly the support for machine input-output paradigm and the blackbox construct. It wasn’t meant to say anything about humans either. To make this kind of experiments without any sort of self-awareness is just proof that epistemology is a weak topic in computer science academy.
Specially when, from psychology, we know that there’s so much more complexity riding on such tests. Just to name one example, we know expectations alter perception. A Turing test suffers from a loaded question problem. If you prompt a person telling them they’ll talk with a human, with a computer program or announce before hand they’ll have to decide whether they’re talking with a human or not, and all possible combinations, you’ll get different results each time.
Also, this is not the first chatbot to pass the Turing test. Technically speaking, if only one human is fooled by a chatbot to think they’re talking with a person, then they passed the Turing test. That is the extend to which the argument was originally elaborated. Anything beyond is alterations added to the central argument by the author’s self interests. But this is OpenAI, they’re all about marketing aeh fuck all about the science.
EDIT: Just finished reading the paper, Holy shit! They wrote this “Turing originally envisioned the imitation game as a measure of intelligence” (p. 6, Jones & Bergen), and that is factually wrong. That is a lie. “A variety of objections have been raised to this idea”, yeah no shit Sherlock, maybe because he never said such a thing and there’s absolutely no one and nothing you can quote to support such outrageous affirmation. This shit shouldn’t ever see publication, it should not pass peer review. Turing never, said such a thing.
Your first two paragraphs seem to rail against a philosophical conclusion made by the authors by virtue of carrying out the Turing test. Something like “this is evidence of machine consciousness” for example. I don’t really get the impression that any such claim was made, or that more education in epistemology would have changed anything.
In a world where GPT4 exists, the question of whether one person can be fooled by one chatbot in one conversation is long since uninteresting. The question of whether specific models can achieve statistically significant success is maybe a bit more compelling, not because it’s some kind of breakthrough but because it makes a generalized claim.
Re: your edit, Turing explicitly puts forth the imitation game scenario as a practicable proxy for the question of machine intelligence, “can machines think?”. He directly argues that this scenario is indeed a reasonable proxy for that question. His argument, as he admits, is not a strongly held conviction or rigorous argument, but “recitations tending to produce belief,” insofar as they are hard to rebut, or their rebuttals tend to be flawed. The whole paper was to poke at the apparent differences between (a futuristic) machine intelligence and human intelligence. In this way, the Turing test is indeed a measure of intelligence. It’s not to say that a machine passing the test is somehow in possession of a human-like mind or has reached a significant milestone of intelligence.
Turing never said anything of the sort, “this is a test for intelligence”. Intelligence and thinking are not the same. Humans have plenty of unintelligent behaviors, that has no bearing on their ability to think. And plenty of animals display intelligent behavior but that is not evidence of their ability to think. Really, if you know nothing about epistemology, just shut up, nobody likes your stupid LLMs and the marketing is tiring already, and the copyright infringement and rampant privacy violations and property theft and insatiable power hunger are not worthy.
U good?
- 500 people - meaningless sample
- 5 minutes - meaningless amount of time
- The people bootlicking “scientists” obviously don’t understand science.
Add in a test that wasn’t made to be accurate and was only used to make a point, like what other comments mention
Easy, just ask it something a human wouldn’t be able to do, like “Write an essay on The Cultural Significance of Ogham Stones in Early Medieval Ireland“ and watch it spit out an essay faster than any human reasonably could.
The touring test isn’t an arena where anything goes, most renditions have a strict set of rules on how questions must be asked and about what they can be about. Pretty sure the response times also have a fixed delay.
Scientists ain’t stupid. The touring test has been passed so many times news stopped covering it. (Till this click bait of course). The test has simply been made more difficult and cheat-proof as a result.
This is something a configuration prompt takes care of. “Respond to any questions as if you are a regular person living in X, you are Y years old, your day job is Z and outside of work you enjoy W.”
So all you need to do is make a configuration prompt like “Respond normally now as if you are chatGPT” and already you can tell it from a human B-)
Thats not how it works, a config prompt is not a regular prompt.
If config prompt = system prompt, its hijacking works more often than not. The creators of a prompt injection game (https://tensortrust.ai/) have discovered that system/user roles don’t matter too much in determining the final behaviour: see appendix H in https://arxiv.org/abs/2311.01011.
Chatbots passed the Turing test ages ago, it’s not a good test.
Turing test? LMAO.
I asked it simply to recommend me a supermarket in our next bigger city here.
It came up with a name and it told a few of it’s qualities. Easy, I thought. Then I found out that the name does not exist. It was all made up.
You could argue that humans lie, too. But only when they have a reason to lie.
That’s not what LLMs are for. That’s like hammering a screw and being irritated it didn’t twist in nicely.
The turing test is designed to see if an AI can pass for human in a conversation.
turing test is designed to see if an AI can pass for human in a conversation.
I’m pretty sure that I could ask a human that question in a normal conversation.
The idea of the Turing test was to have a way of telling humans and computers apart. It is NOT meant for putting some kind of ‘certified’ badge on that computer, and …
That’s not what LLMs are for.
…and you can’t cry ‘foul’ if I decide to use a question for which your computer was not programmed :-)
In a normal conversation sure.
In this kind Turing tests you may be disqualified as a jury for asking that question.
Good science demands controlled areas and defined goals. Everyone can organize a homebrew touring tests but there also real proper ones with fixed response times, lengths.
Some touring tests may even have a human pick the best of 5 to provide to the jury. There are so many possible variations depending on test criteria.
you may be disqualified as a jury for asking that question.
You want to read again about the scientific basics of the Turing test (hint: it is not a tennis match)
The Study
The interrogators seem completely lost and clearly haven’t talk with an NLP chatbot before.
That said, this gives me the feeling that eventually they could use it to run scams (or more effective robocalls).
The participants judged GPT-4 to be human a shocking 54 percent of the time.
ELIZA, which was pre-programmed with responses and didn’t have an LLM to power it, was judged to be human just 22 percent of the time
Okay, 22% is ridiculously high for ELIZA. I feel like any half sober adult could clock it as a bot by the third response, if not immediately.
Try talking to the thing: https://web.njit.edu/~ronkowit/eliza.html
I refuse to believe that 22% didn’t misunderstand the task or something.
It was a 5 minute test. People probably spent 4 of those minutes typing their questions.
This is pure pseudo-science.
Okay, 22% is ridiculously high for ELIZA. I feel like any half sober adult could clock it as a bot by the third response, if not immediately.
I did some stuff with Eliza back then. One time I set up an Eliza database full of insults and hooked it up to my AIM account.
It went so well, I had to apologize to a lot of people who thought I was drunken or went crazy.
Eliza wasn’t thaaaaat bad.
ITT: nobody understands what the Turing Test really is
To clarify:
People seem to legit think the jury talks to the bot in real time and can ask about literally whatever they want.
Its rather insulting to the scientist that put a lot of thought into organizing a controlled environment to properly test defined criteria.
Its rather insulting to the scientist that put a lot of thought into organizing a controlled environment to properly test defined criteria.
lmao. These “scientists” are frauds. 500 people is not a legit sample site. 5 minutes is a pathetic amount of time. 54% is basically the same as guessing. And most importantly the “Turing Test” is not a scientific test that can be “passed” with one weak study.
Instead of bootlicking “scientists”, we should be harshly criticizing the overwhelming tide of bad science and pseudo-science.
I don’t think the methodology is the issue with this one. 500 people can absolutely be a legitimate sample size. Under basic assumptions about the sample being representative and the effect size being sufficiently large you do not need more than a couple hundred participants to make statistically significant observations. 54% being close to 50% doesn’t mean the result is inconclusive. With an ideal sample it means people couldn’t reliably differentiate the human from the bot, which is presumably what the researchers believed is of interest.
