“ GPT-3.5, the base model behind the free version of ChatGPT, has been conditioned by OpenAI specifically not to present itself as a human, which may partially account for its poor performance.” So 3.5 isn’t bad at passing it compared to a 1960s test, it just has been design specifically not to, contrary to the implications of the title.
> some interrogators reported thinking that ELIZA was “too bad” to be a current AI model, and therefore was more likely to be a human intentionally being uncooperative
Given that ELIZA doesn't really work outside its supposed social setup, which makes some of its behavior acceptable and/or intelligible, it would be hard to believe that anyone would have been actually fooled by this. (Still, the performance of ELIZA in the bounds of its intended setup is remarkable, especially for its small rule set. But there is little doubt that this is insufficient for a sustained free conversation.) – Let's call it a "parasitic win". ;-)
You can think of a language as an psychological object, a thing that can be described by a grammar and a dictionary. That's basically where linguistics started, and it's part of it. But some things about language are really best understood and described by looking at it as a thing that people do together, a situated, interactice activity.
And I think that part of the reason that both Eliza and GPT work so well is that the observer really, really wants it to, because that's a crucial part of how this activity is structured, as described by Grice and others:
I like to think of it as language is intelligence as a lifeform, and humans as a hostform.
Language has operations and rules that construct the space of all possible thought, and not all languages are in the set of all possible thoughts. The host form navigates a coordinate space of a constrainted possible thoughts. Whenever a language is used to communicate across interface boundaries (bodies) then some level of mapping takes place to exchanging mutual possible thoughts.
Yep. It demonstrates that the Turing test is strongly influenced by the culture it's administered in. One bot can influence other bots' performance on the test just by existing and being familiar to the public.
There's an amusing perspective I imagine from the intelligence of a AGI/ASI that has to consider the arrogance of being told that one must complete a human-level intelligence test.
Imagine the amusement of said AGI when it reads that us humans are so arrogant to think that any intelligence must come with a full set of our petty human emotions including that of be annoyed by somebody's arrogance.
Another meaningless study written for the sake of writing something.
Didn’t even bother to prepare a proper role for GPT-4.
And the whole article can be shortened down to “GPT 3.5 has been conditioned by OpenAI specifically not to present itself as a human.” and it does it with success :)
GPT4 might not be great at the Turing test either, it seems (https://arxiv.org/abs/2310.20216 and that is the simple non-adverserial version of the test).
If you want to create your own chatbot in RiveScript, the language it's pretty easy
and you can bind Perl functions, code and who knows to your bot, so if you
ask for the weather, for instance, you could call an HTTP module for Perl to call wttr.in for instance. Happy Hacking.
>> raises further questions about using the Turing test to evaluate AI model performance.
Aside from the fact that ELIZA was tuned to present as human and GPT3.5 was not, the validity of the basic Turing Test still seems a key issue. Fooling ordinary humans is a common occurrence, as Richard Feynman pointed out, "... and the easiest one to fool is ourselves.".
The turing test, as has been often discussed, is pointless. Reason being that the tests results are determined at least as much, if not more, by the interrogator and the human interrogee as they are by the machine.
It doesn't test "Can a machine think" so much as it tests "Can a human be tricked". There is a reason why ML research mostly ignores the Turing "Test".
But, alas, its name has a nice ring to it, it is known (by name at least) to pop culture, and easy enough to understand for everyone, and so it gets far more attention than it deserves. Pretty much like Asimovs three "laws" of robotics :D
That doesn't mean it's pointless. For most practical (commercial?) purposes, "can a human be tricked" is a much more important property than "can it think". If a decent portion of humans can't distinguish the AI anymore, with trickery or not, THAT opens up space for plenty of applications, regardless of conscience or not. Think of support bots that sound like humans, sales, email drafts, etc.
Like, most people would say that animals can think, but clearly it can't pass the Turing test. Would you rather have unconscious GPT-4 or a conscious pig "working" at your job?
> For most practical (commercial?) purposes, "can a human be tricked" is a much more important property than "can it think".
I disagree.
In my experience human customers don't care much about whether its obvious when they are communicating with a robot. In fact, in several use cases it's considered good practice, and may even be legally required, to tell them when they are.
What they DO care about, is whether their requests, inquiries, complaints, etc. are handled quickly and efficiently.
We don't create value for our customers by giving them good chatterbots. We create value by giving them highly functional automated systems.
I mean, if it can't handle requests like that, then it'll also not pass the Turing test (since asking the AI to do help with simple support request would get it to fail). I doubt that "thinking" systems are the only ones that can handle a support ticket (GPT-4 does quite alright on a lot of them already), at least if you require consciousness as a requirement for "thinking".
Maybe we just disagree on what we mean when we say "think"?
Chatterbots, including the 60s-era ELIZA, which are mostly useless for this kind of task, have repeatedly been shown to be able to trick people into chosing the machine during the Turing "Test", so I'm not sure how you come by that conclusion.
> at least if you require consciousness as a requirement for "thinking".
I don't require consciousness, or an abstract and not-even-clearly-defined "thinking" capability from an advanced ML model. I require fitness for purpose, whatever that purpose is, and in most use cases I ever encountered, "trick the humen into believing I'm a human" isn't one of them.
> Maybe we just disagree on what we mean when we say "think"?
I don't think we really disagree, I think the problem is that "thinking", "consciousness", "intelligence" and a lot of other terms that often get thrown around when talking about AI, are really not well defined in technical terms.
It gets to the point of the 'philosophical zombie'. That If an AI can completely mimic a human, then some internal subjective 'experiences' must be occurring that are similar to a humans internal subjective experience of reality. Call it consciousness or whatever.
Of course it can't be proven one way or the other, we can't prove humans are conscious either.
Humans think they are conscious because they have an internal sense of self. How do you determine if a machine has that? If it can mimic a human was one hypothetical test.
There are more robust versions of Turing Test than this one, so can't toss it out as an idea.
> It doesn't test "Can a machine think" so much as it tests "Can a human be tricked".
But importantly, Turing didn't just mean it'd be tested on ability to make small talk. He chose the format because he considered text to be a sufficiently general interface for testing any particular intelligent behavior, and gave examples such as feeding in chess moves.
If it can perform all such intelligent behavior at human level (as far as human interrogators can tell), then that's a significant result and - depending on what philosophy of mind you subscribe to - may or may not be sufficient to demonstrate that it can "think".
The "Test" isn't evaluating a machine. It evaluates an interrogators response to that machine.
The core problem of answering the question "can machines think" isn't evaluated by Turings method. The "Test" doesn't test what it is supposed to test.
And that has nothing to do with "what philosophy of mind you subscribe to", that is simply a fundamental flaw in the tests design, and the reason why it's pretty much pointless; it defines "thinking" as "the ability to trick humans". Well, there are optical illusions that can trick people into seeing things that don't exist.
> And that has nothing to do with "what philosophy of mind you subscribe to",
There are definitely those that would reason "Using the tools and expertise available to us, this behaves indistinguishably from a human - therefore we should believe it is intelligent, like humans are".
It's not the same question, and Turing is clear that it's a surrogate question due to the ambiguity of the original question, but it's not unrelated either.
> therefore we should believe it is intelligent, like humans are
They can believe whatever they want, that doesn't mean I have to share their beliefs. Fact of the matter is: Without a definition of intelligence that doesn't involve pointing at ourselves, there is no accurate test of intelligence for something that isn't us.
> this behaves indistinguishably from a human
Several problems with this:
a) Something that behaves like something else, isn't necessarily the same thing. Nature doesn't do duck-typing. A porpoise isn't a fish.
b) Without a clear definition of "indistinguishable from human", this, again, is more a test of the discriminating factors (the interrogator) ability, than of the qualities of the thing that it is actually supposed to test. Case in point: Once upon a time, people didn't know that there is a difference between stars and planets, as both were just bright dots in the sky.
> They can believe whatever they want, that doesn't mean I have to share their beliefs.
I don't claim you have to, and noted that whether you're willing to make such an inference depends on the views you subscribe to.
> Fact of the matter is: Without a definition of intelligence that doesn't involve pointing at ourselves, there is no accurate test of intelligence for something that isn't us.
Say people have varying, largely ambiguous, definitions of what it means for a webpage to be responsive. You propose measuring time from user click to the new content being presented.
For some the test is entirely sufficient as it already encompasses what they mean by responsive, for others it's a partial indication but not on its own conclusive, and for some it may be a separate concern to what they mean.
The test wouldn't settle the issue of definition, and generally isn't equivalent to the original question, but is still useful, relevant, and can actually be carried out.
> Nature doesn't do duck-typing. A porpoise isn't a fish.
I don't think nature really does any typing like this at all. "Porpoise" and "fish" are our own labels, and we may choose to define them based off of observable characteristics or behaviors.
> Without a clear definition of "indistinguishable from human", this, again, is more a test of the discriminating factors (the interrogator) ability,
The test can be repeated with many interrogators to help factor out the variance in individuals' ability.
The proposed test for responsiveness actually does test an aspect of responsiveness. The problem here is only in the definition of the variable to test: Yes, "responsiveness" isn't well defined.
The Turing "test" however doesn't test intelligent behavior at all, so the problem is BOTH in the definition of the tested variable, AND in the actual test.
> The problem here is only in the definition of the variable to test: Yes, "responsiveness" isn't well defined.
That's all that was really required for the purpose of that analogy - arguing that there can still be useful/relevant tests without agreement on definition, as a response specifically to your "Without a definition [...]" claim.
But, I do think you could equally say that the responsiveness test is actually testing speed of the computer or speed of the connection, in addition to some defining responsiveness in a way that isn't related to click-to-new-content speed.
I had forgotten that eMacs even exist. But there must have been another bad update to iOS's autocorrection wieghts because I definitely did not have an eMacs problem the past few years while writing comments on my phone...
Ahhhh interesting, I have had that problem forever but maybe it's because I often shrug it off and let it keep the auto-correct as-is... further reinforcing to it that that's what I meant?
On the other hand, it did offer (correctly) to use the word "fucking" the other day for me instead of trying to change that to "ducking", so maybe we're making a bit of progress!
Is there no repository with all the prompts they've used for their study? In a research like this prompts are the main part, so it's kind of weird for me that they only show one prompt in the paper.
> The study also showed that participants' education and familiarity with large language models (LLMs) did not significantly predict their success in detecting AI.
Also, humans only score 61%, and the best chat gpt prompt gets 41% (but 49% in the graph?).
I’m definitely having trouble distinguishing between content farm crap written by humans and chatgpt output. They’re usually of comparable accuracy.
I’d be curious to see an extended study where they had a dozen human participants or so, and got statistically significant percentages for each individual. Is the worst human below 41%? Is the best close to 100%?
Turing test is broken if the examiners know they're running a Turing test.
Better way to have done it would have been to put them in a chat room with x other humans and just have a normal conversation/meeting, without the humans being told that AI is even involved. GPT would have aced it, this article is just scammy clickbait.
52 comments
[ 3.4 ms ] story [ 91.3 ms ] threadAnd I think that part of the reason that both Eliza and GPT work so well is that the observer really, really wants it to, because that's a crucial part of how this activity is structured, as described by Grice and others:
https://en.m.wikipedia.org/wiki/Cooperative_principle
Language has operations and rules that construct the space of all possible thought, and not all languages are in the set of all possible thoughts. The host form navigates a coordinate space of a constrainted possible thoughts. Whenever a language is used to communicate across interface boundaries (bodies) then some level of mapping takes place to exchanging mutual possible thoughts.
And the whole article can be shortened down to “GPT 3.5 has been conditioned by OpenAI specifically not to present itself as a human.” and it does it with success :)
Nearly fooled me
That's exactly what a human would say.
* Are you a human
> Would you prefer if I were not a human?
Eliza wins this round! %-P
Also, under Perl set up cpanminus if you don't want to play with Emacs:
Logout and login again. Install rivescript: Try Eliza under RiveScript: If you want to create your own chatbot in RiveScript, the language it's pretty easy and you can bind Perl functions, code and who knows to your bot, so if you ask for the weather, for instance, you could call an HTTP module for Perl to call wttr.in for instance. Happy Hacking.it was eventually found
Aside from the fact that ELIZA was tuned to present as human and GPT3.5 was not, the validity of the basic Turing Test still seems a key issue. Fooling ordinary humans is a common occurrence, as Richard Feynman pointed out, "... and the easiest one to fool is ourselves.".
It doesn't test "Can a machine think" so much as it tests "Can a human be tricked". There is a reason why ML research mostly ignores the Turing "Test".
But, alas, its name has a nice ring to it, it is known (by name at least) to pop culture, and easy enough to understand for everyone, and so it gets far more attention than it deserves. Pretty much like Asimovs three "laws" of robotics :D
Like, most people would say that animals can think, but clearly it can't pass the Turing test. Would you rather have unconscious GPT-4 or a conscious pig "working" at your job?
I disagree.
In my experience human customers don't care much about whether its obvious when they are communicating with a robot. In fact, in several use cases it's considered good practice, and may even be legally required, to tell them when they are.
What they DO care about, is whether their requests, inquiries, complaints, etc. are handled quickly and efficiently.
We don't create value for our customers by giving them good chatterbots. We create value by giving them highly functional automated systems.
Maybe we just disagree on what we mean when we say "think"?
Chatterbots, including the 60s-era ELIZA, which are mostly useless for this kind of task, have repeatedly been shown to be able to trick people into chosing the machine during the Turing "Test", so I'm not sure how you come by that conclusion.
> at least if you require consciousness as a requirement for "thinking".
I don't require consciousness, or an abstract and not-even-clearly-defined "thinking" capability from an advanced ML model. I require fitness for purpose, whatever that purpose is, and in most use cases I ever encountered, "trick the humen into believing I'm a human" isn't one of them.
> Maybe we just disagree on what we mean when we say "think"?
I don't think we really disagree, I think the problem is that "thinking", "consciousness", "intelligence" and a lot of other terms that often get thrown around when talking about AI, are really not well defined in technical terms.
This is not actually clear; people commonly feel that they are having meaningful conversations with animals.
The answer to "can a human be tricked" is always yes, regardless of the particular context.
Of course it can't be proven one way or the other, we can't prove humans are conscious either.
Humans think they are conscious because they have an internal sense of self. How do you determine if a machine has that? If it can mimic a human was one hypothetical test.
There are more robust versions of Turing Test than this one, so can't toss it out as an idea.
But importantly, Turing didn't just mean it'd be tested on ability to make small talk. He chose the format because he considered text to be a sufficiently general interface for testing any particular intelligent behavior, and gave examples such as feeding in chess moves.
If it can perform all such intelligent behavior at human level (as far as human interrogators can tell), then that's a significant result and - depending on what philosophy of mind you subscribe to - may or may not be sufficient to demonstrate that it can "think".
And there's the problem again.
The "Test" isn't evaluating a machine. It evaluates an interrogators response to that machine.
The core problem of answering the question "can machines think" isn't evaluated by Turings method. The "Test" doesn't test what it is supposed to test.
And that has nothing to do with "what philosophy of mind you subscribe to", that is simply a fundamental flaw in the tests design, and the reason why it's pretty much pointless; it defines "thinking" as "the ability to trick humans". Well, there are optical illusions that can trick people into seeing things that don't exist.
There are definitely those that would reason "Using the tools and expertise available to us, this behaves indistinguishably from a human - therefore we should believe it is intelligent, like humans are".
It's not the same question, and Turing is clear that it's a surrogate question due to the ambiguity of the original question, but it's not unrelated either.
They can believe whatever they want, that doesn't mean I have to share their beliefs. Fact of the matter is: Without a definition of intelligence that doesn't involve pointing at ourselves, there is no accurate test of intelligence for something that isn't us.
> this behaves indistinguishably from a human
Several problems with this:
a) Something that behaves like something else, isn't necessarily the same thing. Nature doesn't do duck-typing. A porpoise isn't a fish.
b) Without a clear definition of "indistinguishable from human", this, again, is more a test of the discriminating factors (the interrogator) ability, than of the qualities of the thing that it is actually supposed to test. Case in point: Once upon a time, people didn't know that there is a difference between stars and planets, as both were just bright dots in the sky.
I don't claim you have to, and noted that whether you're willing to make such an inference depends on the views you subscribe to.
> Fact of the matter is: Without a definition of intelligence that doesn't involve pointing at ourselves, there is no accurate test of intelligence for something that isn't us.
Say people have varying, largely ambiguous, definitions of what it means for a webpage to be responsive. You propose measuring time from user click to the new content being presented.
For some the test is entirely sufficient as it already encompasses what they mean by responsive, for others it's a partial indication but not on its own conclusive, and for some it may be a separate concern to what they mean.
The test wouldn't settle the issue of definition, and generally isn't equivalent to the original question, but is still useful, relevant, and can actually be carried out.
> Nature doesn't do duck-typing. A porpoise isn't a fish.
I don't think nature really does any typing like this at all. "Porpoise" and "fish" are our own labels, and we may choose to define them based off of observable characteristics or behaviors.
> Without a clear definition of "indistinguishable from human", this, again, is more a test of the discriminating factors (the interrogator) ability,
The test can be repeated with many interrogators to help factor out the variance in individuals' ability.
The proposed test for responsiveness actually does test an aspect of responsiveness. The problem here is only in the definition of the variable to test: Yes, "responsiveness" isn't well defined.
The Turing "test" however doesn't test intelligent behavior at all, so the problem is BOTH in the definition of the tested variable, AND in the actual test.
That's all that was really required for the purpose of that analogy - arguing that there can still be useful/relevant tests without agreement on definition, as a response specifically to your "Without a definition [...]" claim.
But, I do think you could equally say that the responsiveness test is actually testing speed of the computer or speed of the connection, in addition to some defining responsiveness in a way that isn't related to click-to-new-content speed.
M-x doctor can fix all your woes.
On the other hand, it did offer (correctly) to use the word "fucking" the other day for me instead of trying to change that to "ducking", so maybe we're making a bit of progress!
[0] https://corecursive.com/eliza-with-jeff-shrager/
> The study also showed that participants' education and familiarity with large language models (LLMs) did not significantly predict their success in detecting AI.
Also, humans only score 61%, and the best chat gpt prompt gets 41% (but 49% in the graph?).
I’m definitely having trouble distinguishing between content farm crap written by humans and chatgpt output. They’re usually of comparable accuracy.
I’d be curious to see an extended study where they had a dozen human participants or so, and got statistically significant percentages for each individual. Is the worst human below 41%? Is the best close to 100%?
Better way to have done it would have been to put them in a chat room with x other humans and just have a normal conversation/meeting, without the humans being told that AI is even involved. GPT would have aced it, this article is just scammy clickbait.