PharmTech AI Pulse Check: Episode 2 - Shadow AI, Trade Secrets, the API Race

In the second episode of PharmTech's new biweekly video series, “PharmTech AI Pulse Check,” host Chris Cole, associate editorial director, PharmTech, sits down with three expert co-hosts—Richard Jaenisch, Open Biopharma Research and Training Institute, Florin Muraru, independent regulatory affairs advisor, and Gourav Pandey, R&D quality lead, Takeda—to unpack three fast-moving flashpoints in pharma AI: data governance, the shift toward API-gated scientific literature, and the limits of AI-content detection.

The conversation opens with OpenAI's public clarification that a user's Codex prompts couldn't have leaked into the system, alongside Anthropic's new life-sciences verification program. Pandey argues the industry is "arguing with the wrong verb" by fixating on train-versus-retain distinctions. What actually matters, he says, is usage policy: how long data are held, who can access the data, where model weights end up, and how those answers hold up against evolving EU data regulations. He also warns that once an industry's own data feed a benchmark, models tested against that benchmark have effectively already seen the answers, undermining their validity over time. Muraru adds a sobering data point—a report finding roughly 77% of employees paste data into AI tools, often from personal accounts outside company control—and draws a sharp line between batch records (GMP records where any leak risks data-integrity violations) and formulation data (a trade secret that, once pasted into a personal AI account, may lose its legal protection under the EU Trade Secrets Directive). His minimum policy: enterprise accounts with contractual zero data retention, personal accounts blocked at the network level, and on-prem realistically limited to the largest manufacturers. Jaenisch pushes back gently on the trust framing around Anthropic's program, suggesting it's less about trust and more about managing liability around hazardous chemical queries and shadow AI usage. He predicts federated/on-prem deployment will become pharma's norm the way it's emerging in hospitals.

The second segment tackles a prediction that non-API scientific journals will be "dead in five years" as publishers like Reaxys and Springer Nature build AI-facing API access. Muraru warns this risks an "availability bias," with research locked behind PDFs or without API/MCP access simply dropping out of agent-driven literature searches, which is especially dangerous for regulatory and pharmacovigilance work where completeness matters. Jaenisch counters that PDFs are already extractable regardless of access method, so publishers gain little real protection; the practical upside of APIs is curated, publisher-vetted retrieval. Pandey shares original research he calls "citation laundering," showing that retrieval architecture matters more than the underlying model for answer trustworthiness, and warns of compounding availability bias as agents gravitate toward whatever's easiest to retrieve quickly, a Goodhart's Law problem once citation counts become an optimization target.

The final segment addresses new research showing AI watermarking can increase vulnerability to adversarial prompts. Jaenisch argues that detection is the wrong lens entirely, adding that the real question is whether content is attributable, reviewed, and traceable to a source, regardless of whether a person or model drafted it. Watermarking, he notes, mainly flags "slop" and fails on edited or mixed content anyway. Pandey and Muraru agree from a GxP standpoint: validation has never depended on guessing who typed something, but on signatures and audit trails, and detection tools will always lag behind the newest models. As long as a document follows approved procedures and someone takes responsible ownership and signs off, authorship, human or AI, is secondary.