PharmTech AI Pulse Check: Who Owns Authorship, Oversight, and Trust in AI-Driven Evidence

In the latest installment of the PharmTech AI Pulse Check video series, Associate Editorial Director Christopher Cole is joined by Richard Jaenisch, Florin Muraru, and Gourav Pandey to unpack three regulatory developments reshaping how evidence gets generated in pharma: FDA's new discussion paper on generative AI in medical devices, ARPA-H's SURPASS program, and a global push through ICMRA to phase out animal testing starting with batch release.

What Will Industry Push Back On in FDA's AI Paper?

Pandey expects comments—due October 19—to cluster around three themes: how generative AI's non-deterministic behavior fits within existing tools like predetermined change control plans and good machine learning practices; how to validate systems built on continually changing frontier models; and how to manage AI's tendency toward “confident, fluent, but quietly wrong” outputs. Jaenisch adds that because the paper is still a discussion document, “this is the cheapest moment to influence a framework”—a rare chance for real two-way dialogue with FDA rather than reacting to a finished rule.

Who Is Responsible When an AI Agent Goes Rogue?

That question of oversight extends to authorship and accountability. Jaenisch argues that the word “rogue” is misapplied “when a company is initiating the rules and pushing the button and leading the pieces,” meaning liability should sit with the company deploying the tool, not the tool itself. He frames authorship as a scale question, spanning development, deployment, and ongoing monitoring, and says the industry needs clarity on where responsibility attaches at each stage.

On ARPA-H's SURPASS program, Muraru calls regulatory acceptance “realistic,” pointing to EMA's 2022 qualification of prognostic covariate adjustment as precedent for using model-based predictions to shrink control-arm sizes. But he cautions that sponsors must lock models before enrollment, version every change, and justify adaptations against independent data, and that the program's four-year timeline looks tight once follow-up periods are added. Muraru also notes a structural limit: digital twins can reduce how many patients are needed to show efficacy, but safety-monitoring requirements don't shrink alongside them, meaning chronic-therapy trials may hit efficacy answers early while still needing to collect long-term safety data.

How Will Pharma Build Trust as NAMs Take Hold?

The panel also discusses new approach methodologies (NAMs) advancing through ICMRA, starting with batch release. Jaenisch expects the shift to demand new skills in digital biology and large-dataset literacy across QC and regulatory teams, alongside significant retraining since few current employees have hands-on NAMs experience. Muraru stresses that patient trust won't hinge on patients understanding the models themselves, but on knowing “a qualified person reviews the evidence and signs it,” someone the company can name when something goes wrong. Pandey adds that method validation and in silico literacy are becoming core QC competencies rather than “nice to haves.”

Together, the discussion points to a regulatory landscape where simulation, real-time analysis, and non-animal evidence carry increasing weight, and where accountability, validation rigor, and transparent authorship will determine how quickly that evidence gets trusted.