Notes
Short governance notes. Each one takes a single technical question and translates it into governance language: what the risk is, who is accountable, and how it would actually be deployed. Each note is written to teach one idea clearly, turning technical work into governance reasoning that others can follow and reuse.
-
28 August 2026
I ran an identifier audit across my own medical AI governance pack, expecting to find stale cross-references. There were none. What the audit found instead, only because it forced every registered document to be opened, was that seven of them did not contain the document they were registered as: two had become accumulated working notes and five were empty files. Two claims elsewhere in the pack depended on those documents and were therefore false, including a self-assessment row marked Covered on the strength of an artifact that did not exist in the form claimed. The wider finding was worse and more useful: no document in the pack has an assigned approver, so every artifact is written, owned and approved by the same person. This note is the method, not the confession, because the check generalises: existence is not content, and a register that has never moved in the unfavourable direction is not being checked.
-
07 August 2026
The Financial Stability Board names third-party provider concentration and market correlation as two of the AI vulnerabilities most able to turn into systemic risk in finance, and its 2026 sound-practices consultation is moving that concern from monitoring toward expected practice. Read from the hospital side, this exposes a blind spot in the vendor-dependency map I already build: a per-institution review can be done honestly and well and still cannot see the risk that many institutions quietly depend on the same model, cloud, or chip. The finance frame does not just transfer my discipline. It hands back a column my hospital-only version was missing.
-
07 August 2026
The OECD's AI Incidents and Hazards Monitor tracks reported AI harms and sorts them by sector, and government, security, and defence is a category that keeps recurring, from an official body's AI-generated materials misfiring in public to AI-generated legal advice straining a country's social courts. In my clinical AI work I am often both the builder and a large part of the control that says the system is safe, and that is tolerable partly because a clinician mediates the output and the patient is present. When the deployer instead holds public power over citizens who are absent, did not consent, and often do not know AI was involved, self-certification stops being enough. The from-promise-to-evidence discipline still transfers, but the evidence now has to be produced for someone outside the deploying authority.
-
27 July 2026
Google DeepMind published its approach to bioresilience, using AI to strengthen defenses against biological threats while openly acknowledging that the same frontier models can lower barriers to biological misuse. For someone who governs patient-facing medical AI, the useful part is not the biosecurity headline but a flip in accountability geometry: patient-facing governance protects the person in front of the system, while dual-use governance has to protect people who are nowhere near it. This note works through why that flip changes which safeguards count, and why deployment safeguards for bio-capable models cannot just be clinical-style guardrails.
-
27 July 2026Access, not only quality: what the 2026 state laws on AI in prior authorization are really governing
Through 2026 a wave of US states enacted laws on health insurers' use of AI in prior authorization and claims, converging on one rule: AI can assist, but it cannot be the sole basis for denying care, and a licensed professional must make the adverse determination. Most medical AI governance, including most of mine, is about the quality and safety of care for the patient in front of the system. These laws govern something else: access. The useful question they raise is whether the human oversight they require is real or a rubber stamp, and that turns out to be the same from-promise-to-evidence problem, now on the access axis.
-
23 July 2026
Agentic AI governance is moving from voluntary pledges toward institutionalized controls, with national AI safety bodies formalizing testing agreements and operational guidance. The practical question that shift creates is narrow and answerable: for each way an AI agent can cause harm, where is the test that proves it is controlled, and where is the log that proves it held. This note walks through the agent risk-control matrix I built to answer exactly that.
-
22 July 2026AI governance is also infrastructure governance: what model and cloud dependency means for hospitals
The EU's tech chief, Henna Virkkunen, told the Financial Times that AI has become a geopolitical weapon, pointing to Washington's brief June export-control cut-off of Anthropic's models as the case in point. For medical AI, the practical version of that warning is a governance question hospitals rarely put in a risk register: what happens to a patient-facing system when the foreign model it depends on stops being available, for reasons that have nothing to do with clinical performance.
-
19 July 2026
A medical AI can rank patients well and still be wrong about probabilities. If '80% risk' does not mean roughly 80 in 100 similar patients, clinicians and patients are being miscalibrated, not informed.
-
16 July 2026
Most medical AI outputs a risk score, and someone draws a line: above it, act; below it, do not. That line silently allocates false alarms, missed cases, cost, and responsibility. Deciding where it sits is governance, not engineering.
-
14 July 2026
Average performance is a comfortable number. It is also where inequity hides. A medical AI can improve the aggregate while quietly getting worse for the patients who can least afford it.
-
11 July 2026
If before-and-after cannot prove a medical AI worked, what can? Comparison: a control group, a natural experiment, or a difference-in-differences design that isolates the AI's effect from everything else that changed.
-
09 July 2026
After a medical AI goes live, results often look better. But 'better after' is not 'better because'. Mistaking one for the other is how systems get credit they have not earned.
-
06 July 2026
A model that passes validation at launch can quietly decay as patients, equipment, and workflows change. The world moves; the model does not move with it unless someone makes it.
-
04 July 2026
Medical AI is usually validated where it was built. But the sample a model learned from is not the world it will be deployed into, and the gap between them is a governance problem, not a technical footnote.
-
01 July 2026
A cancer-screening model that finds 90% of patients can still be a poor bet for deployment. The governance question is not how accurate it is, but what its errors cost, who bears them, and whether anyone will notice when it drifts.