When Doximity rolled out the next generation of its physician-facing AI tools, the market barely blinked. In a bull market that rewards any entity willing to attach two letters to its earnings narrative, a feature release from a mature clinical platform reads as noise. But the signal is structural, not semantic. Doximity quietly occupies a position that no token project has ever achieved: it is the default communication substrate for American medicine, with estimates suggesting more than four out of five physicians hold accounts there. It is where referrals move, where clinical questions are asked, where the fax machine culture of the industrialized world finally met a modern interface. The company's move into AI is not a pivot. It is the rent-collection mechanism on a twelve-year accumulation of professional behavior data. And the assumption buried inside that roadmap โ that the largest longitudinal map of physician judgment in the United States can be converted into an intelligence layer โ deserves the same audit discipline I applied in 2017, when I refused to sign off on a utility token's smart contract until a critical integer overflow in its vesting schedule was patched. Trust is a protocol, not a promise. No earnings call can compile a governance structure that was never written.
Doximity is frequently described as a software company with a physician network, yet that framing undersells the structural advantage. It manufactures no devices, develops no drugs, owns no hospitals. It is a layer between the actors of the healthcare system โ the referral switchboard, the messaging rail, the professional reputation graph. In blockchain terms, it is the settlement layer that everyone uses and no one governs. The interface does not need to own the value it routes; it needs only to control the routing.
Understanding why this matters to those of us building decentralized infrastructure requires shifting attention from the model to the pipeline. Doximity is not constructing medical AI from first principles. It is extracting value from a corpus assembled for other purposes: every message routed, every referral completed, every search query executed within the platform becomes a training signal. That is not an accident of architecture. It is the centralization model that the blockchain community has spent a decade failing to dislodge. We watched this playbook in social media, in search, in credit scoring. Healthcare is the most consequential arena yet, because the data is not a clickstream. It is the recorded judgment of clinicians under conditions of life and death.
Regulation does not prevent this concentration. It accelerates it. Compliance costs are fixed costs, and fixed costs favor incumbents. A startup proposing a decentralized alternative must clear the same regulatory bar with a fraction of the legal budget, and a governance token does not reassure a hospital's risk committee. The blockchain community has long claimed medical data as a natural candidate for distributed storage and privacy-preserving computation. That claim has produced elegant research and negligible adoption. The gap between our narrative and their reality is not technological. It is a governance gap, defined by liability, professional culture, and an inconvenient fact: the people generating the most diagnostically valuable data are not patients. They are physicians.
So let me be precise about what is actually being built inside Doximity's AI pipeline. The tool draws on three data classes. The first is the physician's own activity within the platform: messaging patterns, reading habits, search queries, response latency. The second is structured clinical artifacts flowing through its secure communications system: referral letters, discharge summaries, prior authorization requests. The third is aggregate behavioral signal across hundreds of thousands of clinicians, which allows the model to compare any individual physician's approach against a real-time national baseline. Combined, these produce something unprecedented โ a longitudinal map of how American medicine actually practices, as opposed to how guidelines say it should.
The engineering challenge is substantial. Fine-tuning a clinical language model on specialty-specific communication requires rigorous de-identification, careful evaluation of hallucination rates, and an understanding of the liability surface that consumer AI companies never touch. I have seen the opposite failure in my own industry. During the DeFi summer of 2020, I watched the DAO I coordinated prioritize velocity over verification, and we paid for it in governance attacks and treasury depletion. The medical AI industry is making the inverse error. It has verification discipline, but it is internalizing the velocity of centralization, which is a different flavor of the same disease: the illusion that control equals safety.
The blockchain community has spent years arguing that patients should own their clinical data through self-sovereign identity and encrypted vaults. That argument has not moved the needle, and I believe I know why. Patients do not generate the most valuable data. Physicians do. And physicians are not demanding tokenized ownership of their behavioral traces on professional networks. They are demanding efficiency, peer validation, and fewer administrative hours. Culture compiles where logic fails, and the culture of clinical medicine is built on reputation, licensing boards, and institutional trust โ not crypto-economic incentives. I learned this lesson managing governance token distribution for a Lagosian artist collective in 2021. We designed the fairest voting structure we could, and the artists, like physicians, mostly wanted to know whether the mechanism would protect their work and pay them on time. Incentive design matters only after professional trust is solved. It is never a substitute.
That said, the governance question does not disappear because physicians are indifferent to token models. It merely relocates to a less glamorous site: the model update pipeline. Every medical AI deployment is a continuous policy decision. When a model adjusts its recommended pathway for a given presentation, when it triages a message to the top of a queue, when it summarizes a patient's chart for a specialist, it is exercising authority over care. The entity controlling that adjustment is not merely operating software. It is writing de facto clinical policy. In a centralized architecture, that policy is written by a product team accountable to shareholders, reviewed by compliance officers, and revised on a quarterly roadmap. There is no physician veto. There is no mechanism for contributing clinicians to audit why their own behavioral data produced a particular output. The people who generate the training signal have no standing in the governance of the model they train.
This is the infrastructure opportunity the market is missing. Not patient-facing data wallets. Not medical NFTs. A verifiable audit layer for clinical AI updates: a public, append-only record of model versions, training data lineage, and evaluation results, anchored to a blockchain and governed by a multi-stakeholder committee that includes physician representatives, hospital systems, patient advocates, and independent auditors. I built something comparable at smaller scale, and the lesson was strategic, not sentimental. Diverse governance structures make attacks more expensive. A model registry governed by a balanced committee would make hidden bias more expensive too. If a competitor can prove that its training data included rural emergency physicians while Doximity's lineage shows a skew toward academic medical centers, that is not an abstract ethics complaint. It is an auditable, falsifiable claim โ the only kind that survives a bear market and a regulatory inquiry.
The second layer worth watching is federated learning with cryptographic verification. The reason centralized platforms dominate medical AI is not technical superiority; it is data gravity. The EHR monopolies hold the data, so the models are built on their servers. Advances in zero-knowledge proofs and secure multi-party computation have lowered the cost of proving that a model was trained on private data without exposing that data. A protocol allowing a hospital network to contribute local model updates to a global diagnostic model, while verifiably proving the contribution, would break the gravity loop. The hospital keeps the data; the network captures the intelligence; the contributor is compensated in tokens tied to measured model improvement. This is not decentralized medicine in the ideological sense. It is decentralized contribution with centralized inference. We govern the gray areas between blocks, not the binary fantasy of total decentralization or total control. The realistic path to physician data sovereignty looks less like a revolution and more like an opt-in protocol with cryptographic receipts. Every time a physician's interaction contributes to training, an on-chain receipt records the contribution, the model version, and the eventual evaluation outcome. The receipt exposes no patient data. It exposes the physician's informational and economic stake.
Having spent the past year negotiating real-world asset tokenization for a Layer-2 protocol, I have learned to translate between compliance and cryptography. The translation for healthcare is awkward but inevitable. Hospital procurement teams will not adopt a decentralized model registry because it is beautiful. They will adopt it because it reduces liability and satisfies an auditor. The framing must shift from user ownership to verifiable accountability.
Now the uncomfortable part. The contrarian position I must hold, with the same skepticism I apply to a Layer-2 carrying a hundred million in funding and a thousand daily active users, is that blockchain may not be the answer for medical AI at all โ at least not in the form our echo chambers imagine. The sober risk analysis cuts both ways. Medical AI is dangerous enough with a single accountable operator. A decentralized update pipeline governed by a token-weighted DAO could introduce precisely the latency and ambiguity that kills patients. When a model drifts, someone must be legally responsible. A smart contract cannot be summoned to a medical board hearing. Regulators will demand a named entity, a compliance officer, a paper trail of human judgment. Vision without verification is just hallucination, and in medicine, hallucination has a body count. Moreover, the physicians we would ask to participate in such governance are the least able to do so. Their working hours are consumed by documentation, examination, and the impossible economics of clinical practice. Asking a practicing physician to audit a zero-knowledge circuit is not inclusive design; it is malpractice. We who advocate for participatory governance must accept that professional expertise is itself a form of decentralization that tokens cannot replicate.
The endgame is therefore hybrid. Centralized operators like Doximity will continue to build and deploy, but they will be compelled, by regulation, procurement requirements, and competitive pressure, to publish auditable lineage for their data and models. The blockchain's role is not to replace the operator. It is to make the operator's claims falsifiable. Trust becomes a protocol, not a promise โ but the judge must be a licensed professional, not a tokenholder. This is the synthesis our industry keeps refusing to name: cryptography for verification, institutions for accountability, and a transparent record for everyone else. Building cathedrals in the bear market means building them where incumbents can see them.
The medical AI story is being written now, and the pen is held by centralized platforms. That is not a tragedy; it may even be a mercy. But for those of us who believe governance determines resilience, the relevant question is not whether Doximity's models are accurate. It is whether the architecture of accountability can survive the scale of deployment. The physician network has achieved what decentralized protocols only dream of: ubiquity, workflow capture, institutional trust. What it lacks is a receipt. Silence in the chain speaks louder than noise. The question is whether the medical establishment will allow that silence to be broken by an auditable protocol, or whether it will compress it until something cracks.
