By Martin Scott

. 6 min read

Four health AI products now on offer do different jobs. ChatGPT Health and Amazon’s Health AI are consumer assistants that read your medical records if you connect them. MedGemma is a set of open models for developers, not an app. Hippocratic AI sells voice agents to hospitals, which use them to call patients.

The real differences are who holds your data and which privacy law covers it. This explainer uses each maker’s own documentation. We have not tested the tools and do not compare their answers, and nothing here is medical advice.

ChatGPT Health: a records layer inside a general chatbot

OpenAI introduced ChatGPT Health on 7 January 2026 for a small group of early users. That first wave excluded the European Economic Area, Switzerland and the UK. On 23 July OpenAI said Health was rolling out to logged-in US users aged 18 and older on web and iOS, across the Free, Go, Plus and Pro plans.

According to the OpenAI Help Center page for Health, you can connect Apple Health (iPhone only), supported US provider portals, One Medical and Function Health. ChatGPT can then compare a new lab result with earlier tests or summarize changes since a last appointment. OpenAI says Health “is not intended for diagnosis or treatment.”

Where the data goes

Synced records sit with OpenAI. The January post says medical records reach ChatGPT through b.well, a US health data network OpenAI partners with. OpenAI says connected data and the chats that use it are not used to train its foundation models or target ads. If you disconnect a source, its synced data is deleted from OpenAI’s systems within 30 days, but chats that already used it stay until you delete them.

One setting changed after launch. The July post said ChatGPT asks permission by default before using connected records. The Help Center now says that from September 2026, new Health connections follow your global plugins setting, which is “Allow low-risk actions” if you never changed it. You can switch this in Settings, then Plugins, then Health.

The Help Center says Health “does not offer a Business Associate Agreement,” the contract HIPAA (the US health privacy law) requires when a vendor handles patient data for a provider. Clinics are pointed to ChatGPT for Healthcare and ChatGPT for Clinicians.

Amazon Health AI: an assistant attached to a clinic

Amazon’s assistant began inside One Medical, the primary care chain it owns. CNBC reported the launch for One Medical members on 21 January 2026. Amazon’s own One Medical announcement says the assistant explains lab results, books appointments and handles medication renewals, and that conversations “are not automatically added to your medical record.”

On 10 March, TechCrunch reported that Amazon was opening Health AI on Amazon.com and in the Amazon app. Amazon’s expansion post says the assistant is free and needs neither Prime nor a One Medical membership. Eligible Prime members get up to five free direct-message consultations with a One Medical provider for more than 30 common conditions. Outside that offer, a pay-per-visit consultation costs $29.

Where the data goes

The difference from ChatGPT is that a care provider sits behind the chat. You give One Medical permission to pull records through a health information exchange, the network US providers use to share patient files. Amazon says Health AI can also see relevant health purchases on Amazon, such as vitamins or blood pressure monitors.

Amazon says all interactions happen “within a HIPAA-compliant environment.” It says protected health information is not used to market general merchandise or by Amazon Ads. It also says it trains Health AI models “on abstracted patterns without directly identifying information.” It runs on Amazon’s Bedrock cloud service and, Amazon says, is not intended for diagnosis or treatment without a care provider’s support.

MedGemma: models for builders, with strict terms

MedGemma is not something a patient opens. Google publishes it as part of Health AI Developer Foundations, a set of open models developers can download. The MedGemma 1.5 model card lists a 4 billion parameter version created on 13 January 2026. It can read CT and MRI scans, pathology slides, chest X-rays alongside earlier images, lab reports and text records.

The model card says outputs “are not intended to directly inform clinical diagnosis, patient management decisions, treatment recommendations, or any other direct clinical practice applications.” Google’s benchmark results come from its own tests and a technical report that is an arXiv preprint, so not yet peer reviewed.

Where the data goes

Because the weights are downloadable, patient data goes wherever the developer runs the model, on a local GPU or in Google Cloud’s Model Garden. Responsibility moves with the data. The Health AI Developer Foundations terms of use bar any use “that could cause a Health Regulatory Authority to deem Google to be a ‘manufacturer’ of a medical device.” Anyone who distributes a product built on it must seek regulatory authorization where that applies.

Hippocratic AI: voice agents sold to hospitals

Hippocratic AI sells to health systems, insurers and drug companies, not to the public. Its Nurse Co-Pilot page describes the flow. A nurse orders a call from the electronic health record (EHR), the agent phones the patient, and a structured note goes back into the EHR, with a full transcript for review. It covers four tasks, including admission education and medication adherence.

The company’s overview page says its agents do not diagnose, prescribe, handle hospice care or mental health disorders, or deal with children under two. It says they hand off to a human nurse when a call needs one. Hippocratic AI says it is HIPAA compliant and holds HITRUST and SOC 2 certifications. Its safety and accuracy figures are its own and we could not check them independently.

For patients, the data path is the hospital’s. Records start and end in the provider’s EHR. The patient never signs up with Hippocratic AI because the hospital orders each call.

What regulators have said so far

We found no FDA statement about any of these four products, only general guidance. The FDA’s clinical decision support guidance, reissued on 29 January 2026, covers software that advises patients or caregivers rather than health professionals. It says such software meets the definition of a device. FDA says it will handle these tools under existing policies, including its general wellness guidance updated on 6 January 2026.

Privacy law splits along the same seam. HIPAA covers providers, insurers and their business associates, which is the position Amazon and Hippocratic AI describe.

Consumer health apps outside HIPAA can fall under the FTC’s Health Breach Notification Rule, which requires breach notices within 60 calendar days. That rule does not apply to HIPAA covered entities or their business associates. OpenAI has not said publicly whether it treats ChatGPT Health as covered by that rule, and we could not confirm it.

State law is moving too. New York’s chatbot rules for minors, which start in June 2027, reach general assistants. They limit reuse of health details from earlier sessions until a user is confirmed as an adult.

Both consumer assistants are US only for now. In Europe, AI Act duties already apply to new text generators, as our explainer on the EU’s 2 December deadline for marking AI output sets out.

Who each one is for

ChatGPT Health suits US adults who want their own records explained inside a general assistant, not clinics. Amazon Health AI fits people happy for One Medical, an Amazon-owned provider, to hold the link to their records. MedGemma is for developers who take on validation and any regulatory filing. Hippocratic AI is for hospitals and insurers that want calls made under their own HIPAA duties.

For anyone weighing artificial intelligence in healthcare at work, the better question is which contract and which law sit behind the data.

What to do this week

If you use ChatGPT Health, open Settings, then Plugins, then Health. Check whether it asks before using your records. If you run a clinic or agency, do not put patient data into consumer ChatGPT Health, since OpenAI offers no Business Associate Agreement for it. If you build on MedGemma, read section 3.2 of the Health AI Developer Foundations terms before you ship anything to patients.

Sources

Martin Scott

Martin Scott is the editor for business and policy at AI Observer, covering startups and deals, regulation, health and science, and AI agency profiles. He follows the money and the legal text, and separates what a rule says from what people claim it says. His style is analytical and plain-spoken.

Spotted a mistake? Tell us on the corrections page.

Conversation

Every comment is reviewed before it appears, so there can be a short delay.