E.g., https://news.ycombinator.com/item?id=33755016 from 2022. Only N-gram frequency analysis, no modern ML techniques needed.

Probably time for somebody to revisit this topic with an actual model.

That's not even close to "reliable and definitive". I'll give you "suggestive". And it presumably has certain volume requirements.

Not to put too fine a point on it, but do you consider handwriting comparisons "reliable and definitive"?

Because juries regularly convict people in courts based on such evidence.

Maybe a better question is, what is the level of "reliable and definitive" that evidence would need to reach for you to believe it should be admissible in court?

I suspect that maybe that's the misunderstanding. The level for you is much higher than maybe it is in the justice system. Which is fine. But that doesn't change the justice system. In court, the currently accepted levels are still enough.

People will be free to try to refute the attributions during the proceedings if they think it will do any good. I mean the adversarial system is also part of the justice system.

What on Earth makes you think that the people who will be using this kind of tech will care how "reliable" or how "definitive" it is? False positives aren't a bug with ICE, they're a perk of the job!