The Benchmarking Gap FDA Hasn't Closed

The Benchmarking Gap FDA Hasn’t Closed

FDA’s operative instrument for AI medical devices is narrow by design. The agency’s December 3, 2024 final guidance on Predetermined Change Control Plans governs how a manufacturer updates a specific model within a specific device’s intended use, without a new marketing submission. It is a change-control mechanism, built for incremental drift in single-function tools. It says nothing about what happens when a model’s capability profile shifts from narrow diagnostic support to something closer to generalist clinical reasoning.

That is precisely the gap Nature Medicine’s 2026 Medical AI Superintelligence Framework is pointing at. The framework proposes benchmarking for AI systems whose capability tier, not just their update cadence, changes over time, a category FDA’s current guidance was never written to evaluate, as the same analysis notes.

This matters because the scale of narrow AI clearance is already large enough to obscure the coming problem. FDA has authorized more than 1,450 AI/ML-enabled devices, with nearly 300 clearances in 2025 alone. Every one of those clearances was evaluated against a fixed, intended-use claim. None of FDA’s current tooling asks whether a system’s reasoning capability has crossed a threshold that changes what oversight it requires. The volume of narrow clearances is not evidence the regulatory model scales to generalist systems. It is evidence the model has not yet been tested against one.

Practitioners closest to deployment are already compensating for this without waiting on the agency. Governance guidance for AI in medical imaging now recommends internal review structures layered on top of FDA clearance, precisely because clearance answers a narrower question than the one hospital risk committees are actually asking, per recent industry guidance. Health systems evaluating vendors are told to track FDA’s evolving framework alongside the EU AI Act’s implementation timeline and independent benchmark datasets, not either one alone, because no single regulatory instrument currently defines capability-tier risk.

There is a counter-argument worth taking seriously: that regulation is not actually the bottleneck on medical AI progress, and that attributing slow adoption to governance gaps overstates the case, as one recent critique argues. That may be true for adoption velocity. It is a separate question from whether current instruments classify risk correctly, and the benchmarking gap is a classification problem, not an adoption problem.

For sponsors and health systems, the decision is not whether to wait for FDA to define a superintelligence-tier benchmark. It is whether internal governance, ideally structured against ISO 42001’s system classification requirements, tracks a model’s capability tier independently of its regulatory clearance status. Clearance answers what a device was approved to do at a point in time. It does not answer what the same underlying model can do six updates later.

Treat FDA authorization as the floor for a narrow-AI compliance program, not the ceiling for a capability-aware one. The framework for the latter is being written in journals before it reaches the agency’s desk.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. The central argument—that FDA’s change-control framework addresses update cadence but not capability-tier shifts—is logically coherent and the distinction between classification problems and adoption
Source & Claim Verification	Qwen · local	cleared. All factual claims are supported by citations, but the claim about the number of AI/ML-enabled devices authorized by FDA (1,450) could be more specific about the source’s relevance to this exact figur
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately identifies a gap in FDA’s benchmarking framework but does not substantively engage with ISO 42001’s system classification requirements or MDR/IVDR’s risk-based tiers.
Technical Accuracy	Llama	cleared. The article accurately highlights the limitations of FDA’s current guidance on AI/ML-enabled medical devices and the need for a more comprehensive framework to address the evolving capabilities of the
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies and addresses a counterargument, clearly distinguishing it from the core thesis, and avoids vendor hype by focusing on regulatory and governance gaps rather than pr
Novelty & Non-Duplication	Grok	held. Core thesis and framing are lifted almost wholesale from the cited Clinical Trial Vanguard opinion on the Nature Medicine framework vs FDA PCCP, making this a repackage rather than a distinct or new a
Validation	DeepSeek	cleared. The central claim that FDA’s current framework lacks a mechanism to evaluate AI systems whose core capability tier fundamentally changes is validated by the agency’s own guidance, which is explicitly

Sources cited: 10. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.