I would never claim that we can reliably detect all AI generated text. There are many ways to write text with LLM assistance that is indistinguishable from human output. Moreover, models themselves are extremely bad at detecting AI-generated text, and it is relatively easy to edit these tells out if you know what to look for (one can try to prompt them out too, though success is more limited there). I am happy to make a much narrower claim, however: each particular set of models, when not heavily prompted to do otherwise, has a "house style" that's pretty easily identifiable by humans in long-form writing samples, and content written with that house style has a very high chance of being generated by AI. When text is written in this house style, it is often a sign that not only were LLMs used in its generation, but the person doing the generation did not bother to do much editing or use a more sophisticated prompt that wouldn't result in such obvious tells, which is why the style is commonly associated with "slop."

I find it interesting that you believe this claim is wildly conspirational, or that you think the difficulty of reliably detecting AI generated text at scale is evidence that humans can't do pretty well at this much more limited task. Do you also find claims that AIs are frequently sycophantic in ways that humans are not, or that they will use phrases like "you're absolutely right!" far more than a human would unless prompted otherwise (which are the exact same type of narrow claim) similarly conspirational? i.e., is your assertion that people would have difficulty differentiating between a real human's response to a prompt and Claude's response to a prompt when there was no specific pre-prompt trying to control the writing style of the response?

On the other fork where I responded to your claims with a direct and detailed response, you insisted that my comment “isn't really that interesting” and just disengaged. I’m not going to write another detailed explanation of why your “slop === AI” premise is flawed. Go reread the other fork if you’ve decided you’re interested.

> I find it interesting that you believe this claim is wildly conspirational

I don’t believe it’s wildly conspiratorial. I believe it’s foolishly conspiratorial. There’s some weird hubris in believing that you (and whatever group you identify as “us”) are able to deterministically identify AI text when experts can’t do it. If you could actually do it you’d probably sell it as a product.

> believing that you (and whatever group you identify as “us”) are able to deterministically identify AI text

I think you will find the OP said no such thing. They instead said they identified a mixture of writing styles consistent with a human author and an LLM. The OP says nothing about deterministically identifying LLMs, only that the style of specific sections is consistent with LLMs leading to the conclusion.

I think you find OP absolutely did say that.

> Parts of it were 100% LLM written. Like it or not, people can recognize LLM-generated text pretty easily

https://news.ycombinator.com/item?id=45868782

Thanks for adding the quote, that is a different part of the post than I was focusing on.

I still think that's a far cry from deterministically recognizing LLM-generated text. At least the way I would understand that would be an algorithmic test with very low rates of both false positives and false negatives. Instead I understood the OP to be saying that people have an intuitive sense of LLM generated text with a relatively low false negative rate.

I am certain that the skill varies widely between individuals, but in principle there is no reason to suspect that with training humans could not become quite good at recognizing low effort (no attempt at altering style) LLM generated content from the major models. In principle it is no different than authorship analysis used in digital forensics, a field that shows fairly high accuracy under similar conditions.

I am pretty much certain that parts of it were LLM-written, yes. This doesn't imply that the entire blog post is LLM-generated. If you're a good Bayesian and object to my use of "100%" feel free to pretend that I said something like "95%" instead. I cannot rule out possibilities like, for example, a human deliberately writing in the style of an LLM to trick people, or a human who uses LLMs so frequently that their writing style has become very close to LLM writing (something I mentioned as a possibility in an earlier reply; for various reasons, including the uneven distribution of the LLM-isms, I think that's unlikely here).

Human experts can reliably detect some kinds of long-form, AI-generated text using exactly the same sorts of cues I've outlined: https://arxiv.org/html/2501.15654v1. You may take issue with the quality of the paper, but there have been very few studies like this and this one found an extremely strong effect.

I am making an even more limited claim than the article, which is only that it's possible for "experts" (i.e. people who frequently interact with LLMs as part of their day jobs) to identify AI generated text in long-form passages in a way that has very few false positives, not classify it perfectly. I've also introduced the caveat that this only applies to AI generated text that has received minimal or no prompting to "humanize" the writing style, not AI generated text in general.

If you would like to perform a higher-quality study with more recent models, feel free (it's only fair that I ask you to do an unreasonable amount of work here given that your argument appears to be that if I don't quit my lucrative programming job and go manually classify text for pennies on the dollar, it proves that it can't be done).

The reason this isn't offered as a service is because it makes no economic sense to do so using humans, not because it's impossible as you claim. This kind of "human" detection mechanism does not scale the way generation does. The cues that I rely on are also pretty easy to eliminate if you know someone is looking for them. This means that heuristics do not work reliably against someone actively trying to avoid human detection, or a human deliberately trying to sound like an LLM (I feel the need to reiterate this as many of the counterarguments to what I'm saying are to claims of this form).

> I’m not going to write another detailed explanation of why your “slop === AI” premise is flawed.

This isn't a claim that I made. I believe that text written with LLM assistance is not necessarily slop, and that slop is not necessarily AI generated. The only assertion I made regarding slop is that being written with LLM assistance with minimal prompting or editing is a strong predictor of slop, and that the heuristics I'm using (if present in large quantities) are a strong predictor of an article being written with LLM assistance with minimal prompting or editing. i.e. I, I am asserting that these kinds of heuristics work pretty well on articles generated by people who don't realize (or care) that there are LLM "tells" all over their work. The fact that many of the articles posted to HN are being accused of being LLM generated could certainly indicate that this is all just a massive witch hunt, but given the acknowledged popularity of ChatGPT among the general population and the fact that experts can pretty easily identify non-humanized articles, I think "a lot of people are using LLMs in the process of generating their blog posts, and some sizable fraction of those people didn't edit the output very much" is an equally compelling hypothesis.

That’s a really interesting study. Thanks for sharing that.

This seems like the kind of thing to share when making a bold claim about being able to detect AI with high confidence. This is a lot more weighty than not so subtly asserting that I’m too dumb to recognize AI.

> a human deliberately trying to sound like an LLM (I feel the need to reiterate this as many of the counterarguments to what I'm saying are to claims of this form).

I assume this is a reference to me. To be clear, I was never referring to humans specifically attempting to sound like AI. I was saying that a lot of formulaic stuff people attribute to AI is simply following the same patterns humans started, and while it might be slop, it’s not necessarily AI slop. Hence the AITA rage bait example.

Thanks for engaging thoughtfully! FWIW I actually looked this article up because I was interested in your claim that even experts couldn't perform these tasks, something I hadn't heard before--I'm not actually ignoring what you're saying. It's actually very nice to have a productive conversation on HN :)