Testing Text Detectors With Pre-Internet Books
cephaloform · x · 2026-07-17
- This retweet discusses how, when evaluating text detectors like Pangram, you can't just test them on public texts from before the ChatGPT era, or the results will be contaminated by the fact that "the model has seen the training data."
- To verify it isn't just memorizing old text, someone found old books that don't exist on the internet, scanned them, and ran tests totaling nearly 3 million words.
- In the results, the detector still found a few traces of AI: 3 sections suspected of AI assistance, and 1 section suspected of AI generation.
- The key takeaway isn't the conclusion itself, but the rigorous anti-"data leakage" evaluation approach it provides.
More from Research
- Project APE finds verifier reliability drops when papers contain multiple errors — soumitrashukla9 · 2026-07-22
- Project APE says verifier costs fell about 90x in a year as Chinese open models lead — soumitrashukla9 · 2026-07-22
- OpenAI-linked paper says capability RL can make models more reward-seeking — MariusHobbhahn · 2026-07-22
- Project APE builds its verifier benchmark from 100 AI-written papers with injected errors — soumitrashukla9 · 2026-07-22
- Paper proposes a CRED taxonomy and benchmark to measure research-error detectors — soumitrashukla9 · 2026-07-22
- OpenAI says long-horizon models need safety and alignment checks across full action sequences — rhiever · 2026-07-22