I put the "humanized" output through Pangram and it still comes out as 100% AI generated.

To be honest with you, I don't think I would be able to identify with high certainty that the bottom text is AI generated, so it definitely goes a long way to obscure the AI-generated nature of it, but I also think it still feels unnatural somehow. I realize my framing naturally calls into question whether I'm being honest, but I am being honest. Given my experience with similar "skills" (it's just chunks of prompt, nothing magical after all) I expected even less.

But still, this is all very strange because it wasn't that many generations of AI models ago that AI writing was a lot better - I'm talking GPT 4.1, Claude 4.5, that sort of era.

Anthropic newsroom posts on the other hand are carefully constructed and well-written in a way that I have not seen demonstrated by LLMs yet, past or present. I expect that they have well-paid staff who are careful with every detail of their public communications. When you put it that way, it almost feels unfathomable that they wouldn't, doesn't it?

	▲	sebmellen an hour ago \| parent [-]
		GPT 4.5 was really good.

▲

breezybottom 8 hours ago | parent | prev [-]

That's about as useful as saying you asked the magical sky fairy.

	▲	meowface 7 hours ago \| parent \| next [-]
		Pangram has an extremely low false positive rate. Even on adversarial examples. One trade-off is even some obviously LLM text won't get detected by them, but they work really hard to ensure false positives are rare since a false accusation is much worse for society than someone getting away with LLM meatpuppetry.
	▲	jchw 8 hours ago \| parent \| prev [-]
		I think you can't trust Pangram in a high stakes situation, but it is absolutely better than random noise at detecting AI-generated text. Which isn't surprising. If the distribution of probabilities can yield blatant Claudisms, it's not surprising it would also have more subtle deviations. (Addendum: As I recall, LLM-generated outputs roughly follow Zipf's law, but the distribution still tends to have some subtle distinctions vs human text; pretty interesting, but I don't know where I heard this, so nothing to cite. Sorry.)