We find that all models are very strong at accurately labeling human text, with the top specialized and general models having perfect performance across our tests. The leaderboard below shows human-detection accuracy.

Weak overall accuracy is driven by mislabeling AI rewrites as human-written, a task for which models are less capable overall.

Additionally, we tested whether models were capable of modifying text to fool other AI-detectors and themselves. We found that Opus 5.5 and Astra can often rewrite more than 50% of a document without being detected as Mixed or AI by Pangram. These models can also fool themselves, though less successfully.

Human writing samples were sourced from private, pre-LLM documents sourced by the Vals team. AI writing samples were produced using a diverse pool of generator models. For our main evaluation, OpenAI and Anthropic latest-generation models are excluded from the generators for this data, because they’re tested as evaluators.