There is a particular silence that follows a regulatory approval, and it is louder than any press release. I have been listening to it for the better part of a year now, ever since the FDA quietly crossed the one-thousandth mark in its clearances of AI-enabled medical devices โ a milestone that arrived with the fanfare of a spreadsheet and the moral weight of a tombstone. A thousand algorithms now sit inside the diagnostic pathway of American medicine. A thousand black boxes render judgments about whose tumor is benign and whose is not, whose scan warrants a biopsy and whose warrants a calendar reminder. And when I went looking for the audit trail โ the verifiable, reproducible record of how each of those judgments was validated, on which populations, with what documented failure modes โ I found, mostly, a void dressed in PDFs.
That void is the real story behind the headlines now bouncing across both crypto and mainstream wires: governments are racing to regulate artificial intelligence in medicine, and the stakes climb higher with every quarter. The race is real. The escalation is real. The urgency in the headlines is not manufactured. What almost no one is asking is whether the thing being built at the finish line is a shield or a stage prop โ a protective wall around patients, or a beautifully printed certificate that certifies nothing at all. That is where my work lives, in the uncomfortable gap between the promise of a system and the provenance of its trust. It is why I keep returning, obsessively, to a question the medical regulators have not yet answered: when an algorithm makes a decision about a human body, who holds the ledger?
The Three Races Nobody Is Winning
To understand what is happening, you have to see that there are three separate races running at once, and they are not friendly ones. They are geopolitical sprints disguised as public-health measures, and each of them is being run with one eye on the patient and one eye on the export market for standards.
The European Union has chosen the path of categorization. Under its AI Act, which entered into force in 2024 with obligations phasing in across the following years, medical AI is stamped "high-risk" almost by default. That stamp triggers a cascade of obligations: conformity assessments, technical documentation, human oversight requirements, risk-management systems, and a quality-management apparatus that would not look out of place in aerospace manufacturing. The philosophy is precautionary โ assume danger, demand proof, and make the proof expensive enough that only the serious survive. It is principled. It is also slow, and slowness is its own kind of clinical risk when the alternative is no tool at all.
The United States has chosen the path of the marketplace. The FDA's approach has historically favored post-market surveillance and voluntary standards, clearing many devices through pathways like the 510(k) that lean on "substantial equivalence" to prior predicates. The philosophy is iterative โ ship, monitor, correct. It is faster, and speed is a virtue the EU conspicuously lacks. Whether it is safer depends entirely on whether the monitoring is real or ceremonial, and I have audited enough "monitoring" programs in my own industry to know which way that usually leans. A post-market surveillance plan that generates reports no one reads is not oversight. It is paperwork performing oversight.
China has chosen the path of the gate. The NMPA treats AI-assisted diagnostic software as a Class III medical device, requiring pre-market registration, algorithmic transparency submissions, and โ critically โ ongoing data updates as the model evolves. The philosophy is central control: nothing moves until it is filed, and the filing is never finished. This has the perverse virtue of treating a model as a living thing that must be continuously re-declared to the state, which is closer to the truth than the West's snapshot approach in at least one important respect.
Three philosophies. Three different theories of where risk lives and who should bear it. And one shared blind spot. Each of these regimes is excellent at regulating the artifacts of AI โ the models, the certifications, the documentation packets, the intended-use statements. None of them is remotely equipped to regulate the behaviors of AI in deployment: the quiet drift of a model as patient populations shift, the silent update pushed to a million devices overnight, the data used to retrain it six months after approval in a jurisdiction nobody is watching. The regulatory apparatus is built to certify a snapshot. It is being asked to govern a river.
This is not a failure of effort. It is a failure of medium. You cannot hold a living system accountable with a dead document. And that โ precisely that โ is the problem distributed ledgers were invented to address, even if the crypto industry has spent a decade forgetting it in favor of number-go-up.
What Compliance Actually Buys
Let me be precise, because precision is the only honest form of generosity I know how to offer.
When a hospital purchases an AI diagnostic tool today, it receives a model, a license, and a folder. That folder contains the training-data summary, the validation metrics, the intended-use statement, and a great deal of language written by lawyers to survive lawyers. It does not contain a cryptographic proof that the version deployed in the radiology suite is the version that was validated. It does not contain a tamper-evident record showing the model has not been fine-tuned since approval. It does not contain a mechanism for verifying, independently, that the manufacturer's claimed sensitivity and specificity actually hold on this hospital's patient population โ a population that is, statistically speaking, almost guaranteed not to resemble the training set.
I have spent my professional life inside systems that claimed to be decentralized while quietly concentrating power in the hands of a few. I recognize the pattern the instant it appears. The medical AI industry is about to become a masterclass in a genre I know too well: the decentralization of responsibility coupled with the centralization of profit. The regulators will approve. The vendors will ship. The liability will scatter like light โ everywhere and nowhere, present in every failure and attributable to no one โ landing finally, as it always does, on the patient who had no seat at the table and no access to the folder.
Compliance, as currently designed, does not produce safety. It produces the documentation of safety, and the two are not the same artifact. They are not even the same category of thing. One is a property of a system. The other is a property of a filing cabinet.
This is the same insight that should have saved us during DeFi Summer in 2020, when audited protocols folded overnight because the audit certified the code and not the governance, not the incentives, not the thousand human decisions surrounding and shaping that code. Alpha hides in the boredom of due diligence โ in the unglamorous work of reading the terms nobody reads, checking the wallet nobody checks, tracing the token nobody traces. The medical AI market is about to be handed a folder full of beautiful, boring, unread assurances. The question is who will actually read them, and whether reading will even be enough when the underlying systems are designed to be unreadable.
The Data That Cannot Cross the Border
Here the blockchain world has something genuinely useful to contribute, and I want to be careful not to oversell it, because my credibility lives or dies on the distinction between a blueprint and a prayer.
Medical AI is voracious for data. The strongest models need the widest, most diverse patient populations in order to generalize โ otherwise you get exactly the failures we already know by heart. There was the Google retinal model that underperformed on underserved populations. There were the pulse oximeters that overestimated oxygen saturation in patients with darker skin, a device-level bias that took a pandemic to make visible. There was IBM Watson for Oncology, which on closer inspection was found to suggest unsafe and sometimes outright incorrect treatments. Each of these was a system that worked beautifully on the data it was trained on and catastrophically on the humans it actually met.
But that same data is exactly what cannot move. GDPR in Europe, HIPAA in the United States, and a patchwork of patient-privacy regimes everywhere else make cross-border training data a legal minefield. The result is a fragmentation that quietly sabotages the very goal regulators claim to be pursuing: safe, generalizable medicine. You cannot build a model that works for everyone if the law prevents you from learning from everyone.
Federated learning โ training models across institutions without centralizing the data โ is the obvious technical answer, and it is already being deployed in pockets. But federated learning has a governance problem, not merely a machine-learning problem. Who decides which institutions participate? How is a participant's contribution verified without exposing the underlying data? What happens when a hospital's local model is poisoned, whether by malice, by a data-pipeline error, or by something as mundane as a mislabeled batch? These are questions of trust among parties who do not trust each other โ which is, definitionally, the problem that distributed ledgers were designed to solve.
Zero-knowledge proofs, in particular, offer the shape of a solution. They allow one party to demonstrate that a computation was performed correctly over private inputs, without ever revealing those inputs. A hospital could contribute to a shared model and prove its contribution was legitimate without exposing a single patient record. A regulator could verify that a model was trained on a compliant dataset without the dataset leaving its sovereign jurisdiction. This is not science fiction. The cryptographic primitives exist today. What does not yet exist is the regulatory imagination to demand them โ the willingness to specify infrastructure instead of outcomes.
Notice what is missing from the current conversation. The regulators are racing to write rules about outcomes while ignoring the infrastructure of accountability that would make those rules verifiable. They are drafting laws on sand and calling it architecture. The difference between a rule you can enforce and a rule you can only hope for is the difference between a ledger and a promise.
The Ledger as Conscience
In 2026, I worked with a small, stubborn team of philosophers and engineers on a protocol we called Veritas Chain โ an attempt to put the provenance of AI-generated content on-chain so that any claim could be traced back to its source, its model, and its moment of creation. We built it partly out of idealism and partly out of fear, because the deepfake era had already begun to dissolve the shared factual substrate that any functioning society requires in order to argue with itself productively. The essay I wrote around that work was cited in a European regulatory whitepaper on AI transparency, and I confess the citation flattered me more than it should have. Truth is coded in transparency, not promises โ and a citation is not a truth. It is a signal that someone, somewhere, found your argument convenient to repeat.
But the medical case makes the content case look like a parlor game. When a synthetic document is misattributed, the damage is reputational. When a synthetic diagnosis is misattributed โ or when an AI's recommendation is silently overridden, or silently followed, without any record of the choice โ the damage is a body. And here is the part that keeps me up at night: in the overwhelming majority of clinical AI deployments today, there is no immutable record of the model's decision, the clinician's override, or the eventual outcome. The data exists in log files that get rotated, in EHR fields that get overwritten, in vendor systems the hospital does not own and cannot inspect. Listening to the silence between the code lines, you begin to hear the patients who were told nothing โ the ones whose care was shaped by an algorithm whose own history has already been deleted.
A ledger would not fix the medicine. But it would fix the memory. It would make it possible, three years after a bad outcome, to reconstruct exactly which version of which model advised which clinician at which moment, and what that clinician did with the advice. That is not a technical luxury. It is the precondition of any meaningful accountability regime, and it is precisely the thing the current regulatory race is not building. We have designed a system that can approve a diagnostic algorithm but cannot prove, on demand, that the algorithm being used is the one that was approved. That is not a regulatory framework. It is a rumor with letterhead.
The first serious malpractice suit against a medical AI vendor will make all of this suddenly legible. When plaintiff's counsel asks the hospital to produce the exact model version, the exact training data lineage, and the exact decision log for the date in question, the folder will be revealed for what it is: a snapshot, taken once, of a system that has changed every week since. I have seen this movie. In crypto, it was called "trust me, the audit passed." The audit did pass. The money was still gone.
The Explainability Trap
There is a seductive answer that regulators keep reaching for, and I want to name why it is a trap, because it is the most consequential technical decision hiding inside the policy debate.
The demand is for explainability: an AI must be able to justify its decisions in terms a human can understand, so that physicians can trust it and patients can contest it. The intention is unimpeachable. The consequence may be perverse. The most powerful modern models โ deep neural networks, the very architectures that outperform humans on medical imaging โ are, by their nature, difficult to interpret. Where they cannot be made interpretable, regulators push toward simpler, more transparent models. And simpler models are, on average, less accurate. Less accurate means more missed tumors, more false positives, more downstream harm. The mandate intended to protect patients may, in the aggregate, harm them.
This is the transparency-versus-performance tradeoff, and I do not think regulators have fully internalized that it is not free. A rigid explainability mandate does not produce trustworthy medicine. It produces transparently mediocre medicine, while the best models migrate to jurisdictions with looser rules โ a regulatory arbitrage that Europe, ironically, set the stage for by regulating first and hardest. Skepticism is the shield; empathy is the sword. We should be skeptical of black boxes, yes. But we should also be empathetic enough to recognize that a slightly opaque model that saves lives may be more ethical than a transparent one that does not.
The better path runs through the ledger, not through the model. If we cannot explain how a model thinks, we can at minimum prove what it was, what it was trained on, what it decided, and who acted upon that decision. Provable provenance is not the same as interpretability, but it is the substrate on which accountability can be constructed. We are debating the wrong axis. We keep asking the model to explain itself when we should be asking the system to prove itself.
The Compliance Moat
Here is the contrarian turn, and I want it stated plainly, because it is the part most coverage of this regulatory race refuses to say.
The race to regulate medical AI is not primarily a race to protect patients. It is a race to establish moats โ and the moats will be dug by whoever can afford the paperwork.
Follow the incentives. Every conformity assessment, every technical file, every ongoing audit and post-market surveillance program costs money โ millions, at scale, and years of waiting. Established players with in-house legal teams, government-relations shops, and hard-won experience navigating the MDR labyrinth will absorb these costs as a line item. A well-funded startup might survive them. A small research lab, a university spinout, an open-source consortium working on a rare-disease model โ they will not. The consequence is a quiet consolidation that the regulators will never be blamed for and the patients will never be told about. The institutions that can afford compliance will define medicine. The ones that cannot โ particularly those serving rare diseases, low-prevalence conditions, and historically underserved populations, where the economics never quite worked anyway โ will simply vanish. Regulation, in the name of safety, will thin the field until safety is whatever the survivors say it is.
I have watched this exact dynamic play out in my own industry, and the parallel is not metaphor. It is pattern recognition. The projects that shout decentralization the loudest are the ones whose team wallets and foundation holdings you can trace on-chain with a few keystrokes. DAOs become compliance shields, and community governance becomes a theatre of quorum. On-chain voter turnout is perpetually below five percent, and the decisions that actually matter are made by whales and VCs in rooms that never appear on any block explorer. The medical AI industry is about to discover what the crypto industry already learned the hard way: transparency demanded is not transparency delivered, and the ledger ultimately remembers who actually acted. The ledger remembers, but the community forgives โ and here the forgiving community is the hospital board that signs the contract, not the patient whose diagnosis was decided by a model nobody can audit.
The Blueprint Nobody Asked For
I am not content to critique. My training, and my temperament, push me toward the constructive โ toward blueprint rather than lament. So let me sketch what an honest regulatory regime would actually demand, borrowing directly from what worked when I helped design the governance of a five-million-dollar arts foundation transitioning into a DAO.
First, provenance over promise. Every approved model should carry a cryptographically signed manifest of its architecture, training lineage, and validation population, anchored to a ledger the regulator can independently read. Re-training should require a new manifest and, where the change is material, a new approval. The model in the clinic should be verifiable against the manifest, the same way a deployed contract can be verified against its source. Second, contribution without exposure. Federated training should be mandated as the default for multi-institution datasets, with zero-knowledge proofs attesting to each contribution's legitimacy and a shared ledger recording participation without revealing a single record. Privacy and generalization are not in tension here; they are jointly solvable, and the failure to solve them jointly is a choice.
Third, the override must be recorded. Every clinician decision to accept or reject an AI recommendation should be logged immutably, not to surveil the physician but to build the accountability substrate that makes contestation possible. Fourth, minority protection. In the arts foundation, the central design problem was preventing whales from steamrolling small voices, and we solved it with a hybrid mechanism that weighted participation breadth, not just capital. The same logic applies to a regulator's advisory bodies and to the standards committees that will write the rules: if a single institution or vendor can dominate the drafting, the rules will serve the drafter. Inclusivity is not a soft virtue; it is an anti-capture mechanism.
Governments are racing. That much is true, and it is not nothing that the race is happening at all. But a race to write rules that cannot be verified is a race to the wrong finish line โ and the people who will pay the price are not the ones running.
The Finish Line Nobody Is Guarding
So let me return to the silence.
The headlines tell you that governments are racing to regulate AI in medicine. They are right about the racing. But a race toward a finish line drawn in policy prose, enforced by self-report, and verified by nothing is not a race toward safety. It is a race toward reassurance โ and reassurance, purchased without verifiable infrastructure, is the most expensive product in the world. It is paid for later, in misdiagnoses, in eroding trust, in the slow discovery that the beautiful folder was never quite true.
I do not believe the answer is less regulation. I have seen what unregulated systems do to people, and I have no appetite for that sermon. I believe the answer is a different kind of regulation โ one that stops certifying snapshots and starts governing rivers; one that demands cryptographic provenance instead of PDF promises; one that builds the accountability ledger into the mandate instead of bolting it on after the first scandal. The infrastructure exists. It is being built, quietly, by people who understand that truth is coded in transparency, not promises.
Because the algorithm will keep making decisions about our bodies regardless of what any regulator writes. The only open question is whether, years from now, when we finally need to know what it decided and why, there will be something left to read โ or only another silence, dressed in a beautiful folder, and a promise that nobody kept.