I didn't want to shell out $20 a month for the general thing, so I spent $90 on data collection and built my own for code comments specifically. It runs locally in your browser with a relatively small classification model trained on old-school stylometric features. You can try that before turning to Pangram for uncertain cases, if you wish.[1]
It's easy to get high accuracy numbers if you're testing on long (50+ words) texts. Much harder when the documens are short, as code comments tend to be.[2]
I’m surprised at the current sentiment in the comments. Pangram is amazing and has really interesting engineering too. I would have guessed that reliably identifying LLM generated text was not possible without watermarks.
> I would have guessed that reliably identifying LLM generated text was not possible
It depends what you mean by "reliably." If you mean, "we should be comfortable relying on this kind of tool at scale to identify and punish students, professionals, and writers who may have used AI," absolutely not.
If you take "reliably" to mean "1 in 200 false positive rate" as they disclose on their front page, absolutely that is possible (they are doing it today!). If you think there are more than 200 assignments turned in over a given year at university, you probably do not consider a tool like this fit for purpose. It's an open question whether those procuring said tool are aware of this
Unfortunately their marketing is really insisting on the former, and trying to push it into the zeitgeist that detection of AI-generated or edited text is reliable-type-1 now and long-term. They fail to make it clear that this is merely a tool that strongly suggests text follows patterns known to us at the present time of known LLMs. However, that fingerprint will drift over time, as LLMs get better, human writing style evolves, and the line between human and "smart autocorrect" becomes even blurrier (does speech-to-text push the model into "AI assisted" mode, because it tidied up your punctuation, for example?)
they also publish great tech reports. their founder is so confident in their model that he's regularly on social media offering bounties for false positives
People expect a binary response, is it AI generated yes or no. But it's more complicated than that. For example, if you see an emdash, it's probably AI generated. But it can also mean the author used it for fixing grammar or tenses. LLMs can't help but try to help. The same for it's not X, but Y. Sure it's a known pattern, but it's not like people don't use this trope all the time.
In my experience, Pangram is great for detecting an author who is trying to pass someone else's work as theirs, or if they are tackling a subject they have little to no knowledge in.
Based on their methodology it looks like the accuracy figures (99.82% for Opus 5) are the true positive rate rather than a combined metric that factors in the false positive rate as well. They claim 1 in 10,000 but it would be nice if we had a per-model breakdown for that specific test.
I don't work for Pangram, but I do a lot of writing at work. API docs, blogs, code, tweets, linkedin, all that.
If you've tried AI detectors a couple years ago, they're basically in the same position that coding agents were a few years ago, where everyone was skeptical at first, but the tech has gotten a lot better. Give it a shot, it's quite good. They are slightly tuned a bit towards classifying things as AI, but I imagine that's deliberate.
The only thing is that their models are pricey, but, very useful.
It's easy to get high accuracy numbers if you're testing on long (50+ words) texts. Much harder when the documens are short, as code comments tend to be.[2]
[1]: https://xkqr.org/aicomment
[2]: https://entropicthoughts.com/better-ai-comment-classifier
reply