Why Foundational Models Break Old Rules
Frontier models no longer fit the tidy categories healthcare governance was built around. A single system can now draft clinical notes, interpret imaging, answer patient questions, and surface billing codes, blurring the line between medical device, decision support, and administrative tool. Traditional approval pathways assume a fixed purpose and a stable model, yet frontier systems are updated continuously and repurposed by users in ways regulators never anticipated.
Also worth reading: Is Healthcare Agentic AI Governance Ready for Benefits Consultant Workflows? · What Is Clinical AI Governance and How Should Healthcare Organizations Implement It in 2026? · How Should Modern Health Systems Navigate Healthcare AI Risk Governance Effectively?
Governance layers must therefore separate the foundational model from the applications built on it, assigning accountability at each level rather than to one opaque vendor. WHO’s recent discussion paper and the emerging “AI Ethics 2.0” agenda both point this way, urging lifecycle oversight, transparency, and clinician responsibility. Whether governance can actually keep pace remains uncertain, but healtho.io exists to help organizations translate these principles into practical, ethical AI adoption.
Separating Model Power From Policy
The core question facing health systems today is whether governance frameworks can evolve as quickly as the models they're meant to regulate. Frontier models now demonstrate capabilities—multimodal reasoning, clinical summarization, diagnostic support—that outpace the review cycles of institutional boards, regulatory agencies, and ethics committees. A model deployed today may behave differently six months later after fine-tuning or architectural changes, yet most governance structures assume a static product evaluated once at approval. This mismatch between model velocity and policy velocity is the central governance challenge of the decade.
One promising approach, echoed in recent WHO guidance and emerging accountability research, is to separate the foundational model layer from the governance layer entirely. Rather than treating the model as the unit of oversight, organizations can build accountability infrastructure—audit trails, usage monitoring, escalation protocols—that persists across model versions. This decoupling lets health systems swap or upgrade models without rebuilding compliance from scratch, and it shifts the ethical burden from predicting every failure mode to ensuring failures are detected, attributed, and corrected quickly. Governance, in other words, should be designed for change rather than against it.
Accountability Gaps in Clinical AI
Frontier models in healthcare are moving faster than the governance structures meant to oversee them, and the gap is widening. Traditional regulatory frameworks were built around discrete, auditable tools—a diagnostic algorithm with a fixed function, a device with a defined intended use. Large multi-modal models defy that logic: they generate outputs across tasks their developers never explicitly designed for, adapt to context in unpredictable ways, and blur the line between general-purpose platform and clinical intervention. When a model trained on broad data assists with triage, documentation, and diagnosis simultaneously, who bears responsibility when it errs—the developer, the deployer, the clinician, or the institution?
The WHO's recent discussion paper on AI in health policy acknowledges both the promise and the risk, but acknowledgment is not accountability. What's needed is a layered approach that separates foundational model development from downstream clinical deployment, assigning distinct obligations at each stage. Developers should warrant transparency about training data and known failure modes; health systems should validate performance in their own populations; clinicians should retain meaningful override authority. Governance that keeps pace will look less like a single gatekeeper and more like distributed, verifiable responsibility across the entire chain of care.
WHO and Bioethics Governance Signals
The World Health Organization's recent discussion paper on AI in evidence-informed health policy marks a turning point: regulators are no longer asking whether frontier models belong in healthcare, but how to govern systems that evolve faster than oversight frameworks can. The WHO emphasizes both the opportunities—faster evidence synthesis, broader access to clinical knowledge—and serious risks, including biased training data, hallucinated recommendations, and opaque decision pathways that clinicians cannot audit. Meanwhile, bioethics scholars writing on "AI Ethics 2.0" argue that governance designed for narrow diagnostic tools simply does not scale to large multi-modal models that generate, reason, and adapt across clinical contexts.
The gap between capability and accountability is widening. Discussions on separating foundational models from governance layers highlight a structural problem: when a base model is fine-tuned, wrapped, and deployed by third parties, responsibility diffuses across the stack, and no single actor owns patient safety. Some developers experimenting with AI accountability frameworks report that even well-intentioned systems drift toward unexamined authority. Sustainable governance will require layered oversight—model-level transparency, deployment-level validation, and continuous post-market surveillance—rather than one-time approvals.
A Consultant's Framework for Balance
The core tension is structural: frontier models evolve on commercial timescales of weeks, while healthcare governance moves through committees, public consultation, and regulatory review that can take years. A model that passed clinical validation in January may have a successor by March with emergent capabilities no one anticipated, leaving governance frameworks perpetually chasing a moving target. This gap is not merely administrative; it translates into real clinical risk when unvetted multimodal systems are quietly deployed for triage, documentation, or diagnostic support before ethical safeguards catch up.
Yet pace alone is the wrong metric. Governance can keep pace if it shifts from static approval to continuous, tiered oversight, embedding ethics into procurement, deployment, and post-market surveillance rather than treating it as a one-time gate. The WHO’s recent discussion paper and the emerging “AI Ethics 2.0” agenda both point this way, urging adaptive regulation that distinguishes foundational models from the governance layers built atop them. Consultants and health systems must therefore design accountability mechanisms that scale with capability, not merely react to it.
Governance Layer vs. Model Layer
| Dimension | Frontier Model Layer | Governance Layer | Can Governance Keep Pace? |
|---|---|---|---|
| Development speed | Rapid release cycles, weeks between capability jumps | Regulatory review cycles measured in months or years | No — governance lags deployment |
| Adaptability | Models retrained and fine-tuned continuously | Static frameworks, slow guideline updates | Partially — needs adaptive oversight |
| Transparency | Opaque weights, emergent behaviors | Requires auditable documentation and disclosure | No — asymmetry favors developers |
| Accountability | Diffuse across labs, vendors, clinicians | WHO and bioethics bodies still defining liability | Emerging — multi-stakeholder efforts underway |