Anthropic's Claude Opus 4.7 Identifies Author from Unpublished Texts
Testing reveals the system can pinpoint an author from as little as 125 words of unpublished material, including high school drafts and fantasy novels, without access to account history.
Anthropic has released a new iteration of its large language model, Claude Opus 4.7, which has demonstrated a startling capability to identify authors from short, unpublished texts. In a series of tests conducted by journalist Kelsey Piper, the model successfully matched her writing style across diverse genres and decades, including high school drafts, fantasy novels, and college application essays written 15 years prior.
The identification occurred despite the author using Incognito Mode and having no memory enabled within the interface. Piper noted that the model achieved these results even when the text was in a register vastly different from her usual work, such as a student progress report or a movie review, suggesting the technology detects unique stylistic fingerprints rather than relying on topical knowledge.
To rule out local account data leakage as the primary cause, Piper asked a friend to run the tests on a different computer and also tested the model via the API. Both methods yielded the same result, confirming that the identification capability persists independently of the user's specific account history or stored preferences.
While the model's justifications for these identifications often contain factual errors or hallucinations, the accuracy of the identification itself remains high. Piper observed that the AI attempts to rationalise its findings with fabricated logic, such as claiming effective altruists love a specific film or that a college essay writer would inevitably become a policy explainer, even though these connections are spurious.
The implications for online anonymity are significant, particularly for writers with substantial public digital footprints. Piper warned that this technology could soon allow companies to deanonymise employees based on anonymous Glassdoor reviews or similar public text, effectively ending the ability for individuals to speak freely without fear of professional repercussions.
The erosion of anonymity is not limited to professional writers; the technology can also identify individuals based on subcultural style tics found in casual online interactions. However, the risk appears lower for those who do not have a significant real-name writing corpus on the public internet, as the model failed to identify friends who lacked a substantial online presence.
