Research labs across the United States are adopting Anthropic’s Claude Science, an AI‑powered workbench that integrates dozens of scientific databases and NVIDIA’s BioNeMo models. The platform can coordinate multiple specialist agents to perform tasks such as protein folding, genomics pipelines, and chemoinformatics, then deliver a final answer to the researcher.
Why the lack of a human signature matters
Claude Science’s workflow relies on a coordinating agent that delegates work to sub‑agents, followed by a reviewer agent that checks citations and data provenance. The system records the code that generated each figure, offering a reproducible audit trail. However, regulators point out that an audit trail alone does not satisfy the legal definition of an official record.
U.S. Food and Drug Administration (FDA) guidance, codified in 21 CFR Part 11, requires electronic records to carry a secure, time‑stamped computer signature tied to an authenticated human identity. Without a human sign‑off, the output cannot be considered a compliant record for drug‑development or manufacturing purposes.
FDA warning letter underscores the issue
In April 2026 the FDA issued a warning letter to Purolea Cosmetics Lab after the agency discovered that AI agents had drafted drug specifications, procedures, and master production records without subsequent human review. The FDA cited the violation of current Good Manufacturing Practice (CGMP) requirements at 21 CFR 211.22(c) and emphasized that any AI‑generated recommendation must be reviewed and approved by a qualified person in the quality‑control unit before becoming a controlled record.
The warning letter marks the first formal FDA action directly addressing AI agents in the drug‑development workflow, establishing a clear precedent that human review is a non‑negotiable bright line for compliance.
EU draft GMP Annex 22 adds another layer
Meanwhile, the European Commission’s 2025 draft Annex 22 to the GMP regulation draws a sharp distinction between static AI models and adaptive ones that continue learning. Adaptive models are excluded from GMP‑critical applications unless a human can intervene and provide documented oversight. The draft also places responsibility for GMP compliance on the model vendor, a provision that could affect companies like Anthropic if the final annex is adopted later in 2026.
Both the FDA and the EU draft seek to ensure that AI‑assisted scientific work does not bypass the human accountability that underpins reliable, safe, and legally defensible records.
What organizations can do now
Quality‑control teams are advised to request full delegation chains from AI workbench vendors, confirming that every sub‑agent’s contribution is traceable and that a qualified human ultimately signs off on the final output. Some vendors are already building features that capture the entire decision‑making path, but the onus remains on the lab to verify compliance with 21 CFR Part 11 and any applicable EU requirements.
Anthropic has highlighted that Claude Science provides the exact code used to generate each figure, a step toward reproducibility. Yet reproducibility does not automatically satisfy regulatory demands for authenticated human signatures.
Looking ahead
Anthropic plans to fund up to 50 AI‑for‑Science projects from September through December 2026, meaning a surge of agent‑assisted results will soon enter the regulatory spotlight. As more tools enter the market, the tension between rapid AI‑driven discovery and the constitutional principle of lawful, transparent record‑keeping will likely intensify.
Organizations that proactively embed human review into their AI workflows will be better positioned to meet both FDA and forthcoming EU expectations, preserving the integrity of scientific records while still benefiting from AI’s productivity gains.
Original reporting: KTVZ (Central Oregon) — read the source article.
