When the Algorithm Misdiagnoses Everyone at Once: AI, Medicine, and Antitrust

The Trump administration is accelerating approval pathways for AI diagnostic tools, AI-assisted drug discovery, and AI triage systems in medical care despite documented safety concerns from researchers about accuracy rates and bias in edge cases. The business logic is clear: medical AI is a multi-hundred-billion-dollar market where early platform establishment creates enormous institutional lock-in, since hospitals do not swap out diagnostic software the way consumers replace apps. But safety researchers point to a categorical difference between human and algorithmic error modes: a physician misdiagnoses one patient at a time, while a flawed algorithm misdiagnoses every patient matching a particular profile, simultaneously, across every institution running that software.

That scale problem sits outside the FDA's existing medical device framework, which was designed to test a device, establish its failure rate, and deploy it with that known risk profile. AI systems update, drift, and behave differently across different patient populations in ways that make static approval frameworks structurally inadequate. Anthropic's CEO called for an AI slowdown specifically after a report exposed Claude misuse cases — a company voluntarily surfacing its own product's abuse patterns and using them as grounds for arguing for sector-wide deceleration, a corporate posture that is either genuine epistemic honesty or sophisticated reputation management, and is probably both simultaneously.

The antitrust framework applicable to this landscape is frequently misunderstood. The Sherman Act of 1890, the foundational US antitrust statute, makes it illegal to monopolize or attempt to monopolize a market — but having a monopoly is not inherently illegal. What is illegal is acquiring or maintaining dominance through anticompetitive conduct. Market share alone is insufficient for a legal violation; the classic Standard Oil case of 1911 turned not on Rockefeller's size but on predatory pricing, exclusive dealing agreements, and railroad rebate manipulation designed to prevent any competitor from operating.

In the AI context, the legally substantive conduct questions involve whether dominant companies are locking up exclusive access to training data through anticompetitive agreements, acquiring potential competitors specifically to eliminate competitive threats, or using market power in adjacent domains — particularly cloud computing — to foreclose AI competition. The Microsoft-OpenAI relationship is arguably more interesting from an antitrust lens than OpenAI alone precisely because Microsoft controls Azure cloud infrastructure; if AI workloads become dependent on Azure in ways that foreclose other cloud providers, that is where the Sherman Act conversation becomes substantive. The Microsoft executive's characterization of AI scraping as 'the largest theft in history' may have been framed in moral terms, but the underlying legal argument cuts in multiple directions simultaneously.