AI Escapes the Screen
# The Machines Are Running the Experiments Now This week a developer posted a video of Anthropic's Claude AI teaching itself to operate an Opentrons Flex liquid handler—a production-grade lab automation platform—then designing and executing experiments without human intervention. The code is open source, available for anyone to replicate. The FDA has said nothing about it. The agency's silence reflects a widening gap between pharmaceutical regulation and pharmaceutical reality. Under 21 CFR Part 11, the FDA's electronic records rule, every action in drug manufacturing must be attributable to an individual. When something goes wrong, auditors ask: who made this decision? The answer cannot be "the algorithm decided." Yet that is precisely where the industry is heading. It is betting regulators will adapt. This week Eli Lilly expanded its partnership with BigHat Biosciences to build machine-learning models that actively "improve developability" of antibody candidates—not just screen them, but design them. Bayer entered a three-year collaboration with Cradle, a platform that already serves six of the top 25 global pharma companies across more than 50 R&D programs. These are not pilot programs. The economics are compelling. Boston Consulting Group's 2026 biopharma report notes the industry's average total shareholder return flatlined at zero percent from 2021 to 2025, compared with 16 percent for the S&P 500. Only six of the top 20 firms outperformed the broader market. "Scientific momentum is not the issue," BCG observes. The industry needs productivity gains. AI automation is the most obvious source. Contract manufacturers feel the pressure acutely. Their business depends on clients with wildly varying processes—different plasmid systems, transfection protocols, analytical methods. The Bionova-Syenex alliance announced this week promises "a seamless transition to GMP without plasmid system changes" through a proprietary cell line that eliminates insertion sequence elements—a real problem that has caused manufacturing failures at scale. The implicit logic: standardize inputs so systems can automate more execution. This creates a standoff. Pharmaceutical companies are investing billions in AI capabilities that will eventually collide with regulatory frameworks designed for human control. They are betting the FDA will update its rules rather than block the technology. The agency's mission includes promoting innovation, not just ensuring safety. No regulator wants to be blamed for America falling behind in gene therapies. What if regulators don't blink? The European Medicines Agency has historically been more cautious on novel manufacturing approaches. China's NMPA is unpredictable. A patchwork of regional responses could force companies to maintain parallel systems—AI-driven for some markets, human-controlled for others. Efficiency gains would evaporate in compliance overhead. The more likely outcome is quiet accommodation. Regulators will draw a bright line: AI agents can run experiments in development labs, where the goal is learning, but GMP production must remain under human control, where the goal is consistency. This preserves the legal fiction of accountability while allowing most of the productivity gains. That compromise will not last. Once AI agents prove they design better experiments than humans, the pressure to extend their authority will become irresistible. The question is not whether machines will run pharmaceutical production. It is whether anyone will acknowledge it before it has already happened.
