Yeah, I don't think that there's going to be a viable future in distinguishing between LLM and human text either, certainly not for people who have an interest in defeating it. That being said:

• Some people are just going to use the most-common models.

• Some models are derived from other models and will share statistical similarities.

I still think that trying to detect LLMs is largely going to be a dead-end road, though.

There are also other wrinkles. I'd bet that an increasing amount of not-trying-to-maliciously-impersonate-a-human text out there is going to come from LLMs. Like, okay, Google Translate uses an LLM. I remember listening to an interview with a senior British EU translator (this was back before Brexit). This is going to be a quality human translator. He was saying that machine translation had gotten so good that these days, that now he always did his first pass with Google Translate and just cleaned it up by hand. Okay, so you figure that this is probably widespread, probably even more so now, right? So if most of the EU's documents are coming from an LLM, even if you could hypothetically detect all LLMs, you probably wouldn't want to trigger on some of them.

Not to mention that people adopt mannerisms from text that they read or hear spoken. If someone is using, say, automated browser translations, they're seeing a lot of the world through an LLM, and you'd expect them to pick up characteristics of that.

and for that matter, looking into it, Pangram seems to have an extremely low false-positive rate these days: https://www.theatlantic.com/technology/2026/05/pangram-ai-detection-accuracy/687381/ . the only complaints seem to be about its false negative rate. the FP rate of 1 in 10,000 is maybe higher than you want but whatever of Eicker's text you throw at it it always comes up positive.

The third screenshot is me asking if it was written by AI and not receiving a reply. The user pretty much does not reply to people, just obviously-LLM summaries to his own posts. Almost all such comments are the same amount of lines. I used the first result for an LLM detector just to give a number; Eicker's language always ticks the "this is LLM" neurons to me.

Now it's 2026 and we have hundreds of thousands of AI models.

And yet they still all write the same. And the vast majority are only using of a handful of models anyway.