Scientist says Anthropic’s AI “discovery” may have relied on his unpublished research

Anthropic's claim that Claude 2 “independently discovered” novel enzymes, which a Copenhagen biologist says actually stemmed from his unpublished 2022 research shared with the model, highlights the growing risk of AI systems inadvertently exposing proprietary scientific data. For website owners hosting research repositories, preprints, or proprietary datasets, this episode signals that AI models can ingest and later regurgitate confidential information even when the data is not publicly indexed. If Claude can surface unpublished findings, bots crawling your site may be flagged by AI services as sources of valuable, non‑public content, potentially attracting higher‑value traffic but also raising intellectual‑property concerns and compliance scrutiny from AI providers. **Actionable steps:** 1. **Update your llms.txt** to explicitly declare whether your site permits AI crawling of unpublished or sensitive research, using the `Allowed: false` directive for proprietary datasets. 2. **Implement bot‑fingerprinting** (e.g., monitor User‑Agent strings and request patterns) to differentiate legitimate scholarly crawlers from AI training bots, and block or rate‑limit the latter. 3. **Deploy a content‑hash monitoring tool** that alerts you when excerpts of your unpublished work appear in AI model outputs or public AI search results, enabling rapid takedown requests.