Millions of people now paste lab results or describe symptoms into a chatbot before they ever reach a doctor’s office. That habit raises an urgent question: what exactly do the major AI assistants permit, what does the evidence say about their accuracy in real hands, and how are those highly sensitive health details protected? The three leading providers — OpenAI, Google and Anthropic — all state in their terms that their assistants do not diagnose or treat. Their dedicated health features, the ones that connect to medical records, have launched in the United States and not in Spain. And a randomized trial published in Nature Medicine found that people consulting a language model were no more accurate than those using their usual resources. This article offers no medical advice; it maps the rules of the game.
- Scale of use: More than 300 million people ask ChatGPT health questions every week, according to OpenAI, up from 230 million in January 2026.
- Geography matters: Record-connecting health features are live in the US only; general health questions remain available elsewhere.
- Evidence gap: In a 1,298-person trial, model-assisted participants identified relevant conditions in under 34.5% of cases, no better than controls.
- Data protection: GDPR classifies health data as a special category requiring explicit consent, which helps explain the European delay.
What the companies actually offer — and where
As of official documentation available on 31 October 2026, no dedicated health feature that links a chatbot to clinical histories is available in Spain. OpenAI introduced ChatGPT Health in January 2026 as a separate space inside ChatGPT, explicitly excluding users in the European Economic Area, Switzerland and the United Kingdom in that first phase. On 23 July 2026 it opened the feature to US users over 18 on web and iOS across the Free, Go, Plus and Pro plans. From outside that market, general health questions still work, but the part that connects records and Apple Health does not.
Anthropic announced in January 2026 that Claude Pro and Max subscribers in the United States could grant access to lab results and histories through beta connectors (HealthEx and Function) and via Apple Health and Android Health Connect. Its documentation does not mention any expansion to Spain.
Google took a different route. Its Google Health Coach, built on Gemini, left testing on 19 May 2026 with global availability inside the Google Health Premium subscription, according to the company’s blog. But medical-record integration is, once again, US-only.
The fine print: no diagnosis, no treatment
All three firms hedge carefully in writing. OpenAI says its health feature is designed to support medical care, not replace it, and is not intended to diagnose or treat. Google is the bluntest: its Gemini privacy notice literally asks users not to employ Gemini apps to obtain diagnoses, treatments or medical, legal, financial or professional advice of any kind. Its generative AI terms add that such content does not replace the advice of a qualified professional.
Usage policies mostly target developers building services for third parties. OpenAI’s rules prohibit offering tools that give personalized advice requiring a license without appropriate involvement from a qualified professional. Anthropic’s treat health as a high-risk case: a qualified person must review the recommendation before delivery, and users must be told AI was involved.
If the company does not stand behind the answer as a diagnosis, neither should you treat it as one.
Does AI help when a real patient uses it?
Controlled evidence suggests it does not improve outcomes in public hands. The trial appeared in Nature Medicine on 9 February 2026, led by Andrew M. Bean of the Oxford Internet Institute at the University of Oxford, and is open access.
It enrolled 1,298 people in the United Kingdom across ten medical scenarios. Some used a language model (GPT-4o, Llama 3 or Command R+) and others used their usual resources, such as a search engine.
Tested alone, the models identified the correct condition in 94.9% of cases. When people used them, relevant conditions surfaced in fewer than 34.5% and the correct decision about what to do — call an ambulance, see a GP — in fewer than 44.2%. Neither figure beat the control group, according to Nature’s summary.
Reviewing some conversations, the authors found two failures: participants supplied incomplete or wrong information, and models sometimes generated misleading content. The models tested are no longer the newest, but the problem the study highlights is one of interaction, not medical knowledge, and a larger model alone will not fix it. A frightened patient who phrases the question poorly is the perfect setup for both hallucination and sycophancy.
Your health data under GDPR
The General Data Protection Regulation defines health data as information revealing a person’s physical or mental health (Article 4.15) and places it among the special categories in Article 9, where processing is prohibited unless exceptions apply — including the data subject’s explicit consent.
That reinforced protection aligns with record-connecting features staying out of Europe. It does not stop you from pasting lab results into a general chat, where each service’s rules apply. Google warns that human reviewers examine some Gemini conversations, that reviewed chats are kept for up to three years, and that users should not include confidential information they do not want a reviewer to see. If you disable Activity saving or use temporary chats, they are not used to train models, though they are stored for 72 hours.
OpenAI says conversations in its health space do not train its base models, and Anthropic says the same about connected health data. But both promises cover features unavailable in Spain; an ordinary chat follows each account’s general settings.
Using these tools with prudence
The reasonable use is the one the companies themselves describe: understanding terms, organizing information and preparing for an appointment, not deciding on treatment.
- Strip identifying details. Name, record number, national ID or clinic are unnecessary to explain what a parameter measures.
- Check privacy settings first: saved activity, temporary chats and whether conversations train models.
- Use it to prepare questions, exactly what OpenAI and Anthropic propose in their announcements.
- Do not use it to decide whether something is urgent. That is precisely the task where the Oxford study found the worst results.
- Take the answer to a professional, never as a substitute for one.
The European delay on these features looks more like protection than punishment. An AI that reads your full medical history can be useful, but it concentrates the most sensitive data that exists inside a company that accepts no responsibility for what it tells you. The chatbot works as a translator of medical jargon and a notepad for questions. Asking more means delegating a decision to a system whose makers, in writing, ask you not to.