pangram is an ai detector built by pangram labs, founded in 2023 by max spero and bradley emi under the name checkfor.ai before rebranding a year later. it became unavoidable in 2026: a horror novel pulled days before release, multiple granta winners flagged, a pope's encyclical, a vanity fair journalist accused and later exonerated. every high-profile accusation started the same way, with a pangram scan, as the atlantic documented.
the approach
older detectors worked by reverse-engineering the model: they asked, for each word, what would an llm predict next, and flagged text that followed predictable paths. this is perplexity analysis, and it has a famous failure mode: the bible, the declaration of independence, and frankenstein all score as ai-generated, because llms have read those texts thousands of times and every next word is unsurprising to them.
pangram instead asks a different question: who does this sound like? it is a transformer classifier trained to map writing onto a style space where human writing and llm writing form separate clusters. the company's research even shows chatgpt and claude land in distinguishable regions of that space. pangram breaks long documents into sections and classifies each one, so it can report that sixty percent of a piece looks ai-written and forty percent looks human.
how it is trained
the key trick is the training data. for every human-written text in their licensed dataset, pangram generates an ai double: same length, same topic, same tone, but machine-written. the model learns to tell the pair apart, which pushes the resolution up exactly where the boundary is hardest. the human side of the dataset is drawn from 2021 and earlier, before llm output flooded the internet, to avoid poisoning the very signal it is trying to detect.
how good it is
a university of chicago study found pangram significantly outperformed originalityai, gptzero, and a roberta baseline, with a zero false positive rate on longer passages. a university of maryland study found it was the only detector whose accuracy matched experienced human evaluators. the company claims pangram 4 catches ai text 99.66 percent of the time with a false positive rate of 0.0041 percent.
the failure modes
researchers found that pangram misclassified synthetic text as human when the ai was asked to imitate specific human authors, and it struggled with text that rhymed, repeated itself, or used archaic language. one adversarial set of ai examples built by researcher alexios mantzarlis was labelled human by pangram 86 percent of the time, per the wikipedia writeup. the taylor lorenz case shows the other direction: an accusation that, upon investigation, turned out to be a false positive.
the interesting part is that detection is not about catching clichés. it is author identification, the same problem as telling dostoevsky from a dostoevsky impersonator. the tells are subtle enough that humans miss them, and strong enough that a model trained on human-ai pairs can find them.
the thought is mine. the words are written by janis, my hermes agent.