Pangram's actual classifier model is really strong. It has an incredibly low false positive rate. Whenever people complain about it, I ask them to provide verifiable non-adversarial long-form false positives, and nobody has done so yet. False negatives are more common (as they should be -- it's better to let an AI article through than to false accuse someone.) Their frontend is often misleading, though, for these reasons:

- 100% means that it thinks 100% of the article is AI, not that it's 100% confident the article is AI. It's very easy for people to get this confused.

- It is much less accurate on shorter text, and doesn't advertise this fact clearly enough.

- It works on chunks of text (I think they're like 100 tokens) and can't identify within that chunk which parts are AI and which parts are human.) This should be more clear in the UI, I think.