Researchers across the United States are watching a new regulatory push that could change how AI‑assisted scientific work is documented. Anthropic’s Claude Science, launched in June 2026, lets a coordinating AI agent delegate tasks to specialist sub‑agents for everything from genomics pipelines to protein‑folding predictions. While the platform automatically records the code behind each figure, neither the coordinating nor the reviewer agents are required to provide a human signature on the final record.
Regulators demand human accountability
The U.S. Food and Drug Administration issued a warning letter in April 2026 to Purolea Cosmetics Lab, citing a violation of current Good Manufacturing Practice (CGMP) rules. The FDA said the lab used AI agents to draft drug specifications and master production records without a qualified human reviewing and signing the results, a breach of 21 CFR 211.22(c) and the Federal Food, Drug, and Cosmetic Act. The agency made clear that any AI‑generated recommendation must be reviewed and approved by a human in the company’s quality‑control unit before it becomes an official record.
Across the Atlantic, the European Commission’s 2025 draft of Annex 22 to the GMP guidelines draws a line between static AI models and those that continue to learn. The draft excludes adaptive models from GMP‑critical applications unless a human can intervene and provide documented oversight. The vendor that supplies the model would bear GMP compliance responsibility. Although still under revision, the final annex is expected by the end of 2026.
What the gap means for R&D teams
Claude Science integrates over 60 databases and connectors, including NVIDIA’s BioNeMo models such as Evo 2, Boltz‑2, and OpenFold 3. Its multi‑agent workflow mimics a research team: a generalist agent receives a request, assigns sub‑agents, a reviewer agent checks citations and data provenance, and the coordinator delivers the final answer. The platform’s audit trail shows the exact code that produced each figure, allowing a researcher to reopen the work months later.
However, compliance experts note that an audit trail alone does not satisfy 21 CFR Part 11, which requires a secure, time‑stamped computer stamp tied to an authenticated human identity. Without a human signature, the output cannot be considered an official electronic record under FDA rules. The same principle applies under the EU draft, where documented human oversight is mandatory for any GMP‑critical function.
Industry response
Anthropic has highlighted the reproducibility of Claude Science, but companies adopting the tool must now address the sign‑off question. Kivo, a provider of AI compliance solutions, recommends that organizations map the full delegation chain of agents and assign a qualified individual to review and sign the final output before it enters any regulated record‑keeping system.
In response to the regulatory climate, several life‑science software vendors are adding features that capture human approval steps, such as electronic signatures and role‑based access controls. The goal is to bridge the gap between the platform’s technical auditability and the legal requirement for human accountability.
Looking ahead
With Anthropic pledging funding for up to 50 AI‑for‑Science projects from September through December 2026, the volume of agent‑assisted research is set to increase dramatically. As more labs adopt AI workbenches, the pressure on regulators and vendors to define clear compliance pathways will grow.
For R&D teams, the practical takeaway is clear: identify who will be the responsible signatory for each AI‑generated record, ensure that the platform can capture a secure, time‑stamped human signature, and maintain a transparent delegation log. Doing so will keep scientific innovation moving forward while staying within the bounds of FDA and forthcoming EU regulations.
Original reporting: KRDO (Colorado Springs metro) — read the source article.
