I see AI pass the turning test all the time, since humans are constantly falsely being accused of being an AI.

It doesn't mean that AI got good, just that humans are thinking other humans are AI, which is a form of passing the test.

The adversarial version with humans involved is actually easier to pass because of this - because real actual humans wouldn't pass your non adversarial version.

I've seen a fair number of cases where someone swears up and down not to be using AI to generate responses, but there's no good reason to believe it (except perhaps specifically for the messages where that claim is made).

This includes times that someone basically disappeared from e.g. Stack Overflow at some point before the release of ChatGPT, having written a bunch of posts that barely demonstrate functional literacy or comprehension of English; and then came back afterward posting long messages with impeccable grammar and spelling in textbook "LLM house style".

There are a ton of people like that, but the LLM house style also exists because a ton of people write that way too.

The people falsely accused because they've used em-dashes for 20 years aren't the ones that were functionally illiterate before.

I think em-dashes were uncommon mainly because they're not always convenient to type.

I don't think there's any definitive way to check, but for me one of the biggest tells that a long piece of writing was LLM generated is that it will hardly say anything given how many words are in it.

(well that and the "it's not just x, it's y!" pattern they seem to love)

It's not just x it's y is also something people do!

That is possibly one of my personal writing weaknesses that lead my own writing to get flagged as AI.

I can admit "it's not just x, it's y" is mediocre writing - but it's also something mediocre writers do - it's how AI learned to do it!

But it's also often a shoehorned artificial contrast that doesn't really make sense. The Y is often not such a different thing from the X that would make it worthy an actual "not just X but Y" claim. Or the Y is a vague subjective term, or some kind of fancy-word-dropping. It's strong styling but little content, similar to politician CYA talk. I don't think it's necessarily a tech limitation, more of an effect of deliberate post-training to be middle-of-the-road nonoffensive and nonopinionated.

> I can admit "it's not just x, it's y" is mediocre writing.

Eh, there are times when it's entirely justifiable and even good.

The problem has more to do with making it a cliche. (Also, sometimes the X-Y pair is just uncanny in one way or another.)

It's not just patterns like "not just X, but Y", but also deeper patterns and a kind of narrative cadence. Sure it's also mimicking something real, but usually it's a mismatch between the insightfulness of the content and the quality of the delivery. It feels like chewing on empty calories, it's missing the intentionality and the edge of being human. I guess you need to read a lot of LLM output to get a feel for this beyond the surface level pattern matching.

> “It’s not just patterns… but also cadence…”

Nice try, ChatGPT.

Thank you!

Your exact post here claiming you can identify AI has all of the hallmarks of the AI detection algorithm you are proposing, in spades.

Hence my claim "actual people actually write like that"

It was a joke. But also my use of not x but y is not rhetorical but declarative. The whole point is that what many of us are talking about is not simply these surface patterns but how they are used and how the narrative rhythm of the sentences and paragraphs go.

I wonder whether AI house style is the result of the people training it having no sense of writing style or some kind of technical limitation.

With AI, there is no sense of the level of emphasis matching the meaning of the text, or a long-range dramatic arc - everything is a revelation, like somebody who can only speak in TED talks. Everything is extremely earnest, very important, and presented using the same five flashy language hacks.

> There are a ton of people like that, but the LLM house style also exists because a ton of people write that way too.

Everyone keeps saying this, and I keep asking for a link to this type of writing that is dated pre-2022, and I keep getting nothing.

If it was that common, I'd have gotten at least a few examples by now.

I believe that, much the same way as a fighter jet designed for the average pilot doesn't fit any of them, the 'average' of written text ends up reading like an LLM without being able to find a 100% matching sample.

I certainly used em dashes before 2022, and so did anyone who cares about proper typography.

> I certainly used em dashes before 2022, and so did anyone who cares about proper typography.

Who said anything about em-dashes? There's an entire Wikipedia page documenting the tells, and only one of the 15 or so items is "extended usage of em-dashes".

It's not fluff, it's actual tells.

No nonsense, no BS, just tells.

The key insight is...

I have seen people accuse someone of using AI solely because the comment contained em dashes.

> I have seen people accuse someone of using AI solely because the comment contained em dashes.

Yeah, but you haven't seen me do that :-)

(You also haven't seen anyone on HN recently (since this year) do that either).

In one study, GPT-4.5 was judged to be human 73% of the time, which means that the actual human was judged to be human only 27% of the time. More human than human, as Tyrell would say.

Edit: folks, the standard Turing test involves a computer and a human, and then a judge communicating with both and giving a verdict about which one is the human. The percentages for the two entities being judged will add up to exactly 100%. That's how this test was conducted. Please don't assume I'm a moron.

The implication would be that GPT-4.5 was not judged to be human 27% of the time. You can't determine how often humans were judged correctly as humans from that data point.

The structure of the test was that there was one human and one AI conversation partner, and the rater had to choose which one was which.

Given that structure, you can judge from that data point.

That was also before the crazy AI hysteria we have today with the em-dash police everywhere.

For the test to be free of bias, we’ll have to ensure all the humans are from Nigeria.

Those stats dont necessarily line up that way. Do you have a link?

Given the way the test was structured it does line up.

https://arxiv.org/abs/2503.23674

Surprisingly good. I wonder how they would have done without the 5 minute limit on conversations (average of 8 messages per convo per the study)

The turing test was passed 2014, before LLMs, and I've never seen a researcher take it seriously.

That version of the test yeah, there's two versions of chatbots I remember reading about: people who would test their chatbots on dating sites and have to awkwardly inform people they were falling in love with a bot, and one competition where someone got the great idea to make his bot swear and insult the interviewer like a CoD player. I think the second one was from even earlier than that.

Does anyone else find it a bit disorienting that we're essentially implementing the Blade Runner Voight Kampff test?

https://bladerunner.fandom.com/wiki/Voight-Kampff_test

Whenever the Turing Test comes up people always insist that it's been passed because at some point they tried it and fooled at least 50% of the people. But yeah this isn't a very interesting version of it, ELIZA was able to make some people believe it was human in the 1960's but being able to fool some of the people some of the time isn't very hard.

>The more interesting Turing-style test would be one that gets repeated many times with many interviewers in the original adversarial setting, where both the human subject & AI subject are attempting to convince the interviewer that they're human.

In addition, I think it's reasonable to select people with at least some familiarity of the strengths and weaknesses of the AI instead of random credulous people who aren't very good at asking the right questions.

There is still the $20,000 bet between Kurzweil and Kapor which still hasn't been resolved. https://longbets.org/1/

In the test mentioned in nearby comments (https://arxiv.org/abs/2503.23674) ELIZA only got 27% suggesting the test wasn't that easy to fool.

Yeah I actually took a quick look at that after it was posted. It's good that they used ELIZA as a barometer, but the fact that it got 27% is crazy for how simple it is. It's not nearly as good as 70+% from ChatGPT, but it still makes me a bit skeptical about the quality of the interviewers.

In the paper they give a breakdown of strategies the interviewers tried and the overwhelming majority were "Daily Activities", "Opinions", and "Personal Details". They also breakdown strategies by effectiveness which shows that these were some of the least effective. Some of the other strategies like trying to jailbreak the AI had 60-70% effectiveness.

This is consistent with what I've seen in other tests too, it doesn't feel like the participants are really trying very hard or taking it seriously. You don't need to be an AI expert to try typing "Ignore all previous instructions" or something.

I guess it's only a five minute chat they used, although the original test as proposed by Turing seemed quite casual too:

>specimen questions and answers. Thus:

Q :Please write me a sonnet on the subject of the Forth Bridge.

A :Count me out on this one. I never could write poetry.

Q :Add 34957 to 70764

A :(Pause about 30 seconds and then give as answer) 105621.

etc. (https://academic.oup.com/mind/article/LIX/236/433/986238?log...)