I’m very skeptical of these results. I run a pretty large group paper website on top of arXiv and we run pangram on papers. The numbers are not nearly this high.
One thing I see a lot is papers flagged as AI because they include llm rollouts in the paper as examples.
Pangram is legit. I don't work at pangram, we integrated it in our paper website and one of the cool emergent behaviors I've seen is that on AI papers with example rollouts, it will accurately mark the paper's main text as human generated and the rollouts as AI generated.
My understanding is that they strongly believe in no false positives, so it's definitely possible to slip something by them but if it marks something as AI, it very likely is.
> My understanding is that they strongly believe in no false positives
Who cares what they "believe" (or, more accurately, say they believe). What are the underlying processes that actually guarantee this, and what data supports it?
2: the third study they link there is based entirely around the assumption that Pangram is correct, and seems to have been a collaboration or something as they're included in the credits area.
> In particular, Qwen3.5-Plus is the hosted version corresponding to Qwen3.5-397B-A17B with more production features, e.g., 1M context length by default, official built-in tools, and adaptive tool use. For more information, please refer to the User Guide.
Well written piece on an attack vector I'd never thought too hard about before. Thanks for elaborating on why sending an email or two to a random person helps an attacker achieve their goal. A lot of similar articles skip over details like that.
reply