The Doximity Paradox: Why Medical AI's Biggest Moat Is Its Biggest Vulnerability

Here is the reality: Doximity is not a traditional medical product company. It is an attention network that happens to have a clinical-grade LinkedIn layer underneath a billing machine. The moment people started calling it a medical AI stock, the market stopped asking the question that matters: what exactly is being optimized, and who owns the ground truth? I have spent the last nine years staring at ledgers. I have manually audited ERC-20 contracts in a co-working space in Austin while everyone around me was chasing ICO whitepapers. I backtested liquidity provision strategies on Uniswap V2 with $50,000 of my own capital when DeFi Summer felt more like a fever dream than an engineering discipline. I traced $2 billion in locked assets back to centralized oracle manipulation during the 2022 crash. That is not a flex. It is a confession: I have a pathological need to know whether the machine actually does what its narration says it does. When I look at Doximity, I do not see a medical AI company. I see a closed database wearing a stethoscope. It has network effects, doctor swag, fax-machine-adjacent messaging, and now a generative AI feature that writes clinical notes. The market reacts to the word AI. But the structural story underneath has nothing to do with transformer parameters. It has everything to do with who controls the provenance of medical truth. Doximity's adoption numbers are often quoted, but the exact figures are less important than the shape of the system. The platform has claimed, through public presentations and industry surveys, that a large percentage of U.S. physicians use some form of its services. I am deliberately not repeating the number as if it were a financial statement line item. Here is what the number represents: a concentrated directory of human beings who make life-or-death decisions, all linked to a centralized database that can be audited, subpoenaed, or silently censored by one entity. Auditing isn't about finding intent; it's about finding the boundary between what the system claims and what it can actually enforce. Doximity claims to be the professional network for doctors. Its AI claims to save time. But what does it enforce? It enforces a workflow that routes documentation through a proprietary schema, captures billing codes, and then uses that captured data to train or tune predictive features. The value loop is elegant, and closed. That loop is the moat. The problem is that walls designed to keep competitors out also keep the outside world from verifying what is inside. In the language of my world, this is the difference between a transparent state machine and a private database with an API. The former can be audited by anyone. The latter can only be audited by the party that controls the keys. Doximity is not malicious. I am not suggesting that it is. But the architecture of medical AI should not depend on the benevolence of one company. Let us be precise about the AI side. Every large language model that touches clinical workflow is trained on a corpus. That corpus contains textbooks, journal articles, clinical notes, discharge summaries, and possibly EHR exports. The model learns patterns from that corpus. The model also inherits its biases, its gaps, and its factual errors. In a centralized system, the training data is a black box. The FDA is trying to figure out how to regulate this. The European Union is passing laws about it. But regulation is a reactive layer. It is not a structural fix. Here is where blockchain stops being a speculative religion and starts being an engineering answer. A zero-knowledge proof can verify that a model output was derived from a specific dataset without revealing the raw dataset. You can assert: this clinical note was generated by a model that was trained on a versioned, hash-committed corpus, and no post-training variable was tampered with. That is not a fantasy. That is a cryptographic primitive that has been deployed in financial systems for years. The same primitive can be applied to medical AI. But let me pause and do the math that matters. ZK proving costs are absurdly high. If gas returns to bull-market levels, operators will bleed money. I said this about L2s, and I will say it again about any system that claims to put medical data on-chain without thinking about proof generation and verification costs. A hospital does not care about ZK if it costs $4 per interaction. Medical record keeping is a high-volume, low-margin operation. The only way a provenance layer survives is if the economic incentive is aligned with something larger than the fee. In blockchain, that alignment comes from token flows. In healthcare, that alignment has not yet been designed. This is the real reason the Doximity moment matters. It is not the AI. It is the possibility that the medical industry will sleepwalk into a centralized data provenance model, convinced that a secure database is the same as a verifiable one. It is not. A database can be encrypted, compliant, and completely untrustworthy as a source of ground truth for AI. Encryption protects against outsiders. It does not protect against the insiders who control the encryption keys. It does not prevent silent training-data poisoning. It does not make the model's lineage auditable by patients, physicians, or regulators. Let me give you a specific failure mode. Suppose Doximity, or any centralized medical AI platform, trains a model on a corpus that contains a systematic bias against a particular demographic. The bias is not introduced by an engineer typing "exclude Black patients". It is introduced by historical care patterns encoded in the source notes. The model learns that bias. It generates a recommendation. The recommendation appears neutral. No single human reviewer can catch it because the bias is distributed across millions of tokens. Now ask: who is accountable? The company. But the company can say the model is a black box, which is true in the narrow sense of interpretability. In the ledger world, accountability moves from intention to action. The data provenance trail makes it possible to trace the model's output back to the exact subset of training examples that contributed to the probability. That is not a legal filing. It is a computational fact. I have built a prototype at Verifiable Truth that uses zero-knowledge proofs to verify the origin of training data for large language models. We have not solved the bias problem. But we have solved the provenance problem. We can prove that a model was trained on a specific hash-committed version of a dataset, and that the model weights used for inference were the weights that went through the training pipeline. The industry calls this model lineage. The rest of the world calls it the difference between a white lie and a hardened audit. Doximity has something that most startups would kill for: actual physicians entering real clinical notes every day. That is a gold mine for AI training. But it is also a toxic asset if the source of that data is fragmented across different institutional workflows and then aggregated into a single private corpus. I am not talking about privacy in the HIPAA sense. I am talking about epistemic integrity. If two hospitals spell the same diagnosis differently, the model will learn to treat those spellings as different things. If no one rewrites the schema, the model amplifies the inconsistency. A ledger would not fix the spelling. But a hash-committed schema would at least make the inconsistency visible and versioned. Let me address the elephant in the room: why does this need to be a blockchain? Why not just a strongly audited database with signed hashes? Because a database is controlled by the entity that hosts it. Even if Doximity promises to publish signed hashes, the signature is only as good as the key management. A government can compel the key holder to sign a different document. A malicious insider can leak the key. A company going bankrupt can let the domain and the keys expire. A blockchain does not require trust in any single key holder. It distributes the state across a network of independent validators. That distribution is what makes the provenance claim checkable after the original publisher loses interest or gets bought out. The ledger doesn't lie, but it also doesn't interpret. A blockchain can anchor a hash of Doximity's training data, and it can prove that the hash did not change. What it cannot prove is that the data was collected ethically, that the labels are correct, or that the model outputs are clinically safe. Those are qualitative judgments. The ledger provides the substrate for those judgments. It turns the question from "can I trust the CEO" to "can I verify the merkle root." Here is what my 2022 crash analysis taught me. The failure of Celsius and FTX was not a smart contract bug. It was an off-chain oracle problem. The ledger said one thing, and the world said another. Decentralization is meaningless without decentralized data integrity. The same lesson applies to medical AI. A model can be open-sourced, decentralized, and honestly trained, but if the labels are poisoned, the output is poison. Doximity's centralized corpus is the oracle problem in medical form. Silence is the loudest audit trail in the market. Doximity does not publish its training data distribution. It does not publish the versioned hashes of its clinical note datasets. It does not publish a public audit of what its AI model learned from the physicians using the network. That silence is not evidence of wrongdoing. But it is evidence that the company's moat is based on opacity, not transparency. If medical AI is to be trusted, that opacity becomes a liability. Not a regulatory liability, necessarily. A structural liability that will eventually be exploited by a more transparent competitor. The contrarian view is that Doximity does not need on-chain provenance because it has brand trust and regulatory clearance. That view treats trust as a permanent feature rather than a depreciating asset. Let me stress-test it. Traditional medical trust is built on licensing boards, peer review, and face-to-face interactions. Doximity has digitized the professional network but not the peer review layer. AI-generated clinical notes are not peer reviewed before they are returned to the physician. The physician edits them, signs them, and sends them to the insurance company. The only audit is the claim denials. That is not a quality pyramid. That is a billing loop. The more Doximity pushes AI into that billing loop, the more its database becomes a closed training set for future models. The company is not selling software. It is selling a predictive billing machine wrapped in a clinical utility layer. That works until the machine produces an output that harms a patient. Then everyone will ask for the training data. And Doximity will say it cannot share the training data because it contains protected health information. That is not a technical limitation. It is a design failure. ZK proofs were created for exactly this scenario: prove the property without revealing the secret. If Doximity does not use them, it is because the market has not yet forced it to. We didn't build cryptographic integrity for fun. We built it because centralized systems keep failing in structured, predictable ways. The failure is always the same. The real-world data contradicts the company narrative. The company responds with legal pressure or a public relations statement. The ledger, if there is one, shows the true state. The speed of the resolution is determined by how many independent watchers can verify the ledger. In the crypto world, we call this "stronger contracts." In the medical world, it is just called "accountability." Let me bring this back to Doximity's immediate situation. The stock has likely been bid up on AI enthusiasm. I do not have access to the exact ticker price in the source document, and I will not fabricate a number. What I know is structural: when a centralized platform becomes the default AI documentation layer for a large cohort of physicians, its value to the public market is not its revenue. Its value is the exclusive access to that physician-generated clinical text. That text is the most valuable training data in medicine because it is the raw record of how doctors actually think, write, and decide. But value without provenance is fragile. The market is starting to understand that AI models trained on unverified data collapse. Synthetic data injected into a training set produces model drift. Adversarial text can be crafted to induce specific outputs. If a competitor wants to damage Doximity, it does not need to hack the EHR. It just needs to poison the model's training corpus through some public channel. Without a hash-committed data provenance chain, Doximity cannot even prove that the poisoning did not occur. That is the paradox. Doximity's moat is the network of doctors. The network feeds the AI. The AI feeds the network. The stronger the loop, the larger the centralized data target. The larger the target, the more likely an adversary will attack the training pipeline rather than the API. In my 2024 work with institutional clients, I saw this play out repeatedly with AI models in finance. They all looked robust until someone published a carefully designed prompt-injection on a public forum. The damage was not the prompt itself. The damage was that the companies could not prove what the model had learned from that prompt. Code is the only law that doesn't ask for permission. That is why I am evangelizing a standard, not a token. A few of us in Texas recently drafted a Proof of Decentralization standard for the Texas State Blockchain Council. We did it because regulators wanted to certify decentralization, and they had no technical framework to do it. The same problem exists in medical AI. Regulators cannot certify that an AI model was trained on a transparent, verifiable, and tamper-evident dataset because there is no common standard for data lineage. Doximity could lead that standard. It could become the first major medical network to anchor its training corpus to a public ledger. It could make its AI auditable without exposing a single patient record. It will not do that. Not because the engineering is impossible. The engineering is straightforward. It will not do it because the opacity is structurally profitable. The less you can audit a billing recommendation, the harder it is to challenge the billing recommendation. The harder it is to challenge the recommendation, the more complicated it is to deny a claim. Those are the economics of a closed system. The blockchain answer is a threat to those economics, which is why it will be resisted. But the resistance does not make the need less real. I have watched three cycles of this. In 2017, people said tokens were the answer to everything. They were not. In 2020, people said liquidity pools were the answer to everything. They were not. In 2025, people said ETFs and regulation would make crypto institutional. That also was not the answer. The answer has always been the same: verifiable state across independent nodes. That is the only way to prevent the oracle problem, the insider problem, and the model-poisoning problem. Let me give you a concrete architecture for how Doximity could do it. Take the clinical note corpus. Hash the raw text. Commit the hash to a cheap, cold-data chain. Then generate a ZK proof that the model weights were produced using that exact corpus and a specific training algorithm. Publish the proof commitment on the same chain. Anyone with the model binary can verify the proof and compare the hash. That is the beginning. Next, add a schema version hash. If Doximity updates its note templates, the schema hash changes. The model version is bound to the schema version. Now every model output can be traced back to a specific schema version and a specific corpus hash. That is not a toy. That is a database design principle that I have used in production for enterprise data contracts. Will physicians care? The early majority will not. But the people who pay the bills care. Insurance companies care. Hospital compliance officers care. Patients who have been harmed by AI recommendations care. The legal system cares the day someone files a deposition asking for the model's training data. The problem is they are not the influencers who move Doximity's stock. The stock moves on an AI narrative. The narrative fades when the first major AI-related medical lawsuit drags a closed ledger into discovery. My takeaway is not that Doximity is a sell. I do not make price predictions. My takeaway is that the conversation around medical AI is deeply mis-specified. We are arguing about whether the AI is accurate, biased, or safe. We should be arguing about whether the system can produce a verifiable audit trail. Accuracy is a model property. Bias is a data property. Safety is a deployment property. Auditability is a state machine property. The first three can be fixed after the fact. The last one has to be designed in from the beginning. I have been asked many times why I spend so much time thinking about provenance. The answer is in the 2022 ledger analysis. I mapped two billion dollars of locked assets to a single centralized oracle failure. That moment killed any residual romanticism I had about big-brained founders or benevolent platforms. The only thing that saved my capital was the ability to read the chain directly. No one had to give me permission. No one could spin the data. That is the property I want for medical AI. I want to read the chain and know whether the model was trained on a consistent, versioned source of truth. Flow follows fear, but only if the protocol holds. In the current sideways market, that phrase applies to crypto capital. In the medical AI market, it applies to adoption. Hospitals will not move their workflows to Doximity's AI features because of a compelling product demo. They will move when they are afraid of the liability of not having an audit trail. The protocol either holds that trail or it does not. Doximity's centralized ledger will hold for now. It will hold until a patient, a regulator, or a competitor asks to see the provenance. Then the conversation will change. The next stage of medical AI will not be measured by the size of the model. It will be measured by the integrity of its lineage. Integrity is not a marketing term. It is a property of a system that cannot be edited by a single insider and cannot be hidden behind a legal firewall. A secure database is not enough. A compliant cloud is not enough. A governance committee is not enough. The system must produce cryptographic proof that its claims are anchored in a state that nobody can rewrite after the fact. That is what Doximity is missing. Not a better AI. Not a bigger network. Not a faster integration with electronic health records. It is missing the willingness to make its own ground truth public. The market will eventually reward the platform that is brave enough to open that door. I am not holding my breath. But I have learned that the ledger is patient. It will outlast every narrative, including the ones that currently make the stock sing. Here is the forward-looking thought I want to leave you with. Imagine a second Doximity, built on the same network effects but with a verifiable data provenance layer as its default. It would offer the same clinical note generation, the same billing workflow, the same fax integration, but every model output carries a proof receipt. A physician can review the note. A hospital can verify the lineage. A regulator can audit the corpus without seeing a single patient. A patient can ask, "why did the model recommend this test?" and get a cryptographically signed answer that points to the exact clinical evidence. That is not a hundred-billion-dollar company. That is a new category of trust infrastructure. Doximity can be that company. It has the doctors. It has the data. It has the AI momentum. The only missing piece is a philosophical commitment to verifiability. Without that commitment, its data moat becomes a liability. With that commitment, it would redefine what medical AI means for a generation. The choice is not technical. It is a choice about who owns the truth. The truth should not be owned by anyone. It should be anchored everywhere.