People talk about Pangram as if the detector were a black box that “just knows.” It is not. The 2024 technical report and the Pangram 4 model card describe a recipe that is almost exactly what working writers now suspect:
Take a human document. Ask several large language models to write the same thing. Train a classifier on the difference.
They call the AI side a synthetic mirror. From their February 2024 post:
We prompt an LLM by requesting a document on the same topic, of the same length. For a fraction of examples, we have the LLM start with the first sentence of the human document.
That is not “detect AI in the abstract.” That is this book page versus this model’s rewrite of this book page. Pangram 4 still uses the same idea, then adds a humanizer head, token-level labels, and mirrors from “75 of the most popular models, including all presently available OpenAI and Anthropic models.”
They train on books. A lot of books.
Table 5 of the 2024 paper lists the human pool at about 28 million documents. One row is not a rounding error:
- Books: 7,000,000 examples
- Scientific papers: 3,000,000
- Student writing: 23,000
- ESL: 165,000
They say those sets are open source, commercially licensed, and dated 2021 or earlier so ChatGPT cannot contaminate the human label. They evaluate the books domain on Project Gutenberg. They do not publish the catalog behind the 7 million training examples. Gutenberg cut into chapters can get you into the millions. So can a shadow library. Those are different claims. We take the number as fact and the source as unnamed.
The second trick: hunt the humans you already fail
Naive training saturates. So they scan the remaining millions of human documents, keep the ones the current model already calls AI, generate new mirrors of those, and retrain. That is Algorithm 1, “hard negative mining with synthetic mirrors.”
In plain language: the model is repeatedly shown the living humans it mistakes for machines, then shown an LLM writing the same assignment. The decision boundary is “book-shaped / student-shaped 2021 prose” versus “LLM rewrite of that prose.” Fully invented AI books are the easy class. The hard class is a human page that already looks a little like a model.
What our scans add
On 2 September 2026 we scored our own converts on Pangram 4.0 (dashboard, no API key):
- Unused leftover-free Gutenberg sentences at ~1,000 words: 14/14 types entire-text Human once we stopped inserting generated asides.
- The same original story rewritten by Llama 3.3 70B, Qwen 2.5 72B, and Grok 4.5: 100% AI, every time, celebrity voice or not.
- Pasting Vercel / Stripe names onto those Gutenberg stems did not flip the document. The book rhythm stayed Human.
That is what you would expect if the human class is attested book prose and the AI class is an instruct-model rewrite. It does not prove they torrented LibGen. It does prove the detector is allergic to LLM cadence and comfortable with 1925 sentences.
The essay for the thread: You now prove you’re human to a model trained on the books the labs ate. This page is the paper notes.
Sources: arXiv 2402.14873, arXiv 2607.27183, Pangram blog, Feb 2024, Pangram 4 model card, TechCrunch, 29 Jul 2026.