On January 7, OpenAI launched ChatGPT Health, a dedicated health section of ChatGPT that can connect to users’ medical records, wellness apps and wearables. Gizmodo reported that OpenAI spent more than two years on the product with over 260 physicians from 60 countries, and that more than 40 million people ask ChatGPT for health advice every day, according to the company.
Medical records come through b.well, a health data network. In a press release the same day, b.well said users can explicitly authorise ChatGPT to access records from US healthcare providers across a network of more than 2.2 million providers and 320 health plans, labs and other sources.
/ 01What OpenAI launched
According to Gizmodo, users can connect apps including Apple Health, MyFitnessPal, AllTrails, Peloton and Instacart, and OpenAI pitched uses such as helping people “understand recent test results” and “prepare for appointments with your doctor.” Medical record integration is US-only; the rest of the product is available more widely but not in the European Union, Switzerland or the United Kingdom. At launch, access was limited to a small group of early users.
On privacy, Gizmodo reported that health chats have separate memories, that health conversations will not be used to train OpenAI’s foundation models, and that OpenAI describes its protections as “purpose-built encryption and isolation.” OpenAI used HealthBench, a benchmark it released in 2025, in developing the product.
b.well says its SDK for Health AI converts connected health data, both structured and unstructured, into AI-ready inputs for language model applications, using FHIR-based APIs. Its chief AI and technology officer, Imran Qureshi, said in the release: “Today, AI is only able to work across a small part of the patient record.”
/ 02Why it matters
People were already asking ChatGPT about their health; OpenAI’s Ashley Alexander called it “one of the most common and meaningful ways people use ChatGPT today.” Connecting the actual record is a logical next step, because generic answers to personal medical questions are of limited use. An answer about a lab value means more when it knows the patient’s medications, history and previous results.
It also raises the stakes. A general health answer that is slightly wrong is a nuisance. A personalised answer that misreads a value in a patient’s own record, or misses an interaction between two of their medications, is a different kind of error, and one the patient is less likely to question because it appears to be about them.
/ 03The traceability angle
The b.well quote points at the real problem. A medical record is not a document; it is a set of linked facts: conditions, medications, encounters, results, each with dates and sources, often duplicated or contradictory across providers. Feeding that into a language model as text and asking it to reason is the easiest approach and the least verifiable one.
A more robust approach models the record explicitly. When conditions, drugs and observations are entities in a graph linked to clinical knowledge, such as guideline recommendations and known interactions, a question about a patient becomes a traversal: which facts apply, which rules fire, which sources support them. The language model can then explain the result in plain language, and every statement in that explanation can point back to a specific entry in the record or a specific line in a guideline. That is what lets a clinician, or a cautious patient, check it.
The same structure helps with the record’s known messiness. When two providers list a medication differently, or a lab result appears twice with different units, a graph can represent the conflict explicitly and flag it, instead of leaving a model to pick one version silently. Surfacing uncertainty is a feature in clinical settings, not a weakness.
In health, an answer you cannot trace back to the chart is an answer you should not act on.
/ 04What to watch
- How answers cite the record, and whether users can see which entries an answer relied on.
- Regulatory treatment. Gizmodo noted FDA Commissioner Marty Makary’s view that software “simply providing information” can avoid FDA regulation; that line will be tested as products get more personal.
- Provider-side responses, as health systems decide whether to build their own grounded tools rather than leave patients to consumer apps.
Connecting the record is the easy half. The hard half is that a chart assembled from many providers through FHIR is full of duplicates, unit mismatches and medication lists that disagree, and a language model reading it as text will smooth those conflicts over without saying so.
Modelled as a graph, each medication statement, lab observation and encounter becomes its own node, tied to the provider that recorded it and the date it was recorded. Two conflicting entries stay visible as two entries. Clinical knowledge sits in the same graph: the cardiovascular build described on our site turned guidelines into 145 linked concepts and 284 recommended actions, each pointing back to its source page. A question such as “does my latest potassium result change what the guideline says about my current dose” then resolves to named record entries and a named recommendation that a clinician can open and check. Deployed on premises, the protected health information stays inside the provider’s network.
How we’d model it
Lab observation—recordedBy→ProviderMedication statement—conflictsWith→Medication statementRecommendation—derivedFrom→Guideline pageAnswer—cites→Record entry
- OpenAI Launches ChatGPT Health, Wants Access to Your Medical Records — Gizmodo gizmodo.com ↗
- OpenAI Selects b.well to Power Secure Health Data Connectivity for AI-Driven Health Experiences in ChatGPT — b.well resources.icanbwell.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.