New "computational Turing test" reportedly catches AI pretending to be human with 80% accuracy.
See full article...
See full article...
Being too nice online is a dead giveaway for AI bots, study suggests
- Thread starter JournalBot
- Start date
What about spouting absolutely bullshit with perfect confidence, never using one word when an inaccuracy followed by a dozen rationalizations will do? The great similarity between AI driven chat and a human being is the almost complete inability of either to tell the truth. The most significant characteristic of chat AI is its perfect resemblance to a dirty politician.
Upvote
47
(52
/
-5)
"Did you run that through the AI checker?"
"Didn't have to. It's human."
"What? Look at it. Never uses the same synonym twice, em-dashes all over the place....what makes you think it's human written?"
"In the essay I was told where to go and how to get there. Twice. And conjectured about certain proclivities that my matrilineal family line engage in with paying customers."
"Ah. Yes, that would do it."
"A+ paper otherwise. That student's got promise."
"Didn't have to. It's human."
"What? Look at it. Never uses the same synonym twice, em-dashes all over the place....what makes you think it's human written?"
"In the essay I was told where to go and how to get there. Twice. And conjectured about certain proclivities that my matrilineal family line engage in with paying customers."
"Ah. Yes, that would do it."
"A+ paper otherwise. That student's got promise."
Upvote
115
(118
/
-3)
The thumbnail shows a frowny face but the article image shows a smiley face. I can't tell if that's intentional or just caching gone wrong. Either way, I love it.
Upvote
70
(70
/
0)
I'd really like to see whether this holds up for models whose training data includes a healthy dose of Russian vernacular.
EDIT: I was being kind of flippant, but actually maybe this would be an legitimately interesting question. I do sort of wonder whether they would get the same results testing across languages. Russian is an interesting test case, because some of the boundaries around what might be considered "polite" speech are a little different (at least, such was the manner of speech in the post-Soviet country in which I lived many years ago).
And then I think also of testing this out in a language like Korean, which has some layers of formality in terms of verb endings and pronouns -- I mean, this is actually kind of interesting to think about, especially when you consider the gobbledygook that most training data actually consist of...
EDIT: I was being kind of flippant, but actually maybe this would be an legitimately interesting question. I do sort of wonder whether they would get the same results testing across languages. Russian is an interesting test case, because some of the boundaries around what might be considered "polite" speech are a little different (at least, such was the manner of speech in the post-Soviet country in which I lived many years ago).
And then I think also of testing this out in a language like Korean, which has some layers of formality in terms of verb endings and pronouns -- I mean, this is actually kind of interesting to think about, especially when you consider the gobbledygook that most training data actually consist of...
Last edited:
Upvote
40
(40
/
0)
Please consider that this may be one of the most polite comment threads in the history of Ars. Thank you for participating.
Upvote
49
(52
/
-3)
That smiley faced image appears to show a computer not connected to the Internet. I think we have a correlation. (and what a pristine clean looking classic Mac that is)
I just asked ChatGPT to insult me - very clearly not a human. Note - I don't pay for AI, I just occasionally waste their resources.
I just asked ChatGPT to insult me - very clearly not a human. Note - I don't pay for AI, I just occasionally waste their resources.
Upvote
14
(16
/
-2)
In Fallout 4, there is a random encounter with a military robot who is looking for spies and demands that you prove you aren't one.
There are a number of different answers which will get you out of this, but saying "Go Fuck yourself, you crazy robot" will convince it that you couldn't possibly be a spy.
There are a number of different answers which will get you out of this, but saying "Go Fuck yourself, you crazy robot" will convince it that you couldn't possibly be a spy.
Upvote
86
(86
/
0)
J.C. Helios
Ars Scholae Palatinae
??? What does this mean? Isn't "retrieving relevant context" table-stakes for having a conversation?
Upvote
18
(20
/
-2)
As someone who tries desperately to avoid toxicity (due to the anxiety it causes me—not that I don't get anxious about all interactions…), let me unleash things for just a moment:
As someone who also took up em-dashes with the ease of starting to use compose keys a few years ago, I'm fucking sick and tired of learning everything I do is what LLMs are tuned to do. This is fucking bullshit.
As someone who also took up em-dashes with the ease of starting to use compose keys a few years ago, I'm fucking sick and tired of learning everything I do is what LLMs are tuned to do. This is fucking bullshit.
Upvote
87
(89
/
-2)
That must be one of those magical pre-Bluetooth desktop computers where neither mains nor keyboard need connecting.
Upvote
10
(10
/
0)
Hmmmm… duh? Aren't most LLMs (definitely ChatGPT) trained and augmented to be more helpful and nicer than the typical response you'd get online? That's the whole point of why people veer towards AI chatbots instead of asking questions in online forums — (perceived) helpfulness (and speed).
After all, when you post a question on Reddit you know what you'll get: One third snark, with the wittiest responses earning top votes; one third helpful/empathetic responses; and one third tell you that you're the worst/dumbest/weirdest person ever to even pose such a question. Sort of doesn't matter if it's a question about coding, interior design, or AITAH.
After all, when you post a question on Reddit you know what you'll get: One third snark, with the wittiest responses earning top votes; one third helpful/empathetic responses; and one third tell you that you're the worst/dumbest/weirdest person ever to even pose such a question. Sort of doesn't matter if it's a question about coding, interior design, or AITAH.
Upvote
70
(70
/
0)
Why would "sound like a human on social media" be a design goal for AI? Not only is that hardly the gold standard for desirable behavior by humans, it also doesn't sound even close to the "killer app" that makes AI worthwhile or profitable.
Upvote
68
(68
/
0)
This matches my expectations, I see lots of accounts that strike me as inauthentic, saying things like, "you have an interesting point of view, I would like to hear more." It doesn't advance conversation or even seem to be aware that there is a conversation happening.
Upvote
26
(26
/
0)
Why is this unexpected? Personally, it's exactly what I would expect. I mean the next sentence literally says that they undergo training to do something which is not mimicking humans. I'm 99% sure that Anthropic and OpenAI are not including "Dude, why would you even ask that" and "Just google it!!" as good reponses in their instruction tuning datasets.
Upvote
51
(51
/
0)
YetAnotherAnonymousAppellation
Ars Praefectus
The one really easy to spot clue in my experience is posts with mixed up gender references. Either that or mixing up gender is a real problem for some humans.
Upvote
8
(8
/
0)
give it time, they'll make chatbots start sounding like jerks online. those social media sites need it else they'll lose their precious engagement.
Upvote
24
(25
/
-1)
Upvote
35
(35
/
0)
Wait until they start launching campaigns to claim that Gore won in Florida, we never landed on or went to the Moon, JFK was shot by Russian agents, the Earth is flat, the Vegas Mandalay Bay shooting was a false flag, storm systems Katrina and Harvey were caused by HAARP and whatever else triggers rage responses.
Upvote
12
(12
/
0)
Or whatever all the troll bots on Reddit are using. I'm wondering what bots they ran into. Karma bots?
Upvote
4
(4
/
0)
Well, that’s easy enough, everyone just has to end their correspondence with “and I fucked your mother last night”.
At long last, we have a way to identify the humans from the clankers.
At long last, we have a way to identify the humans from the clankers.
Upvote
21
(22
/
-1)