Health in ChatGPT

Health in ChatGPT, first introduced as ChatGPT Health, is a consumer health and wellness feature inside ChatGPT from OpenAI. It can connect information from supported medical records and Apple Health, then use that context to explain records, compare results over time, prepare questions for an appointment, and discuss habits such as sleep, activity, nutrition, and exercise. OpenAI describes the feature as support for understanding and navigating health, not as a replacement for medical care or a tool for diagnosis or treatment.^[1][2]

The name reflects two stages of the product. OpenAI announced "ChatGPT Health" on January 7, 2026 as a separate, protected space available to a small early-access group. On July 23 it began a broader U.S. rollout under the interface name "Health in ChatGPT." The later design keeps a Health tab for managing records and accounts but, with permission, can use connected health context in conversations elsewhere in ChatGPT.^[1][3]

Name and rollout

The January announcement described a waitlist open to Free, Go, Plus, and Pro users outside the European Economic Area, Switzerland, and the United Kingdom. Access was initially limited, medical-record integrations were U.S.-only, and Apple Health required iOS. That was an early-access policy, not the current availability statement.^[3]

By July 25, 2026, an archived copy of OpenAI's help documentation said Health was rolling out gradually to logged-in users who were at least 18 years old and located in the United States. It was included with Free, Go, Plus, and Pro accounts and was supported on the web and on iOS. Connecting Apple Health required an iPhone. The help page said the rollout could take several weeks and that Health was not supported in Voice mode or Codex.^[2]

The redesign followed early testing in which OpenAI said more than 70 percent of health-related conversations by people with access took place outside the dedicated Health space. Users can now permit connected information to be considered in any ChatGPT conversation. The Health tab remains the place to connect accounts, inspect synced records, view trends, and return to health conversations. Chats and uploaded files from the earlier version remain in a legacy Health Project and do not move into the new tab automatically.^[1][2]

Connections and workflows

Current supported sources include Apple Health, supported U.S. provider portals, One Medical, and Function Health. A user may connect multiple accounts, but availability varies by provider and application. Fitness, sleep, and nutrition services such as Whoop, Oura, Garmin, Strava, or MyFitnessPal can contribute information indirectly when they share it with Apple Health and the user grants ChatGPT access to the relevant categories. Some app-specific or proprietary scores may not transfer.^[2]

OpenAI identified b.well as the connectivity provider for the medical-record integration introduced in January. b.well said its network uses FHIR-based APIs and can supply authorized records including laboratory and imaging reports with clinician interpretations, problems and diagnoses, medications and immunizations, clinical notes and hospital discharge summaries, care plans, and follow-up instructions. That partner description concerns the data pipeline, not independent validation of ChatGPT's medical conclusions.^[3][5]

After syncing, a user can ask ChatGPT to compare a new laboratory result with prior tests, summarize changes since a visit, explain a clinical note in plain language, identify questions to ask a clinician, or examine sleep and activity patterns. Health can also use relevant context in everyday requests, such as considering an allergy during restaurant planning or an injury during exercise planning. Users can review conditions, medications, and family history, correct missing context inside Health, and browse their imported records. The connection is read-only: ChatGPT cannot change a provider's record or write information back to Apple Health.^[1][2]

The experience is intended for the account holder's own records. OpenAI advises using a separate account or a separate project with project-only memory when exploring another person's information. This matters because mixed records could otherwise produce misleading context.^[2]

Permissions, memory, and data handling

Connected information is optional. By default, ChatGPT asks before using medical records or Apple Health information to personalize a response. A user can approve one request, choose to allow access generally, change that choice in Settings, or add @Health to a prompt to request health context explicitly.^[1][2]

OpenAI says connected medical records, Apple Health information, and conversations that use that data are not used to train its foundation models or target advertising, regardless of the user's ordinary model-training setting. Conversations that do not use connected Health data continue to follow the standard ChatGPT data-control setting. OpenAI's Health Privacy Notice separately says that, by default, it does not use connected medical data, ChatGPT Health conversations, or conversations with Connect Health enabled to improve foundation models.^[1][4]

The training exclusion does not mean that no human or service provider can ever access the information. The privacy notice says a limited number of authorized OpenAI personnel and trusted service providers might access Health Features data to improve model safety unless the user has opted out. It also permits processing for support, fraud prevention, security, legal compliance, and other purposes described in OpenAI's general privacy policy. The notice says OpenAI does not sell personal data collected through Health Features.^[4]

ChatGPT conversations are encrypted at rest and in transit, and OpenAI says connected Health information receives additional encryption protections. Users choose which accounts to connect. Disconnecting a source causes synced data from that source to be deleted from OpenAI systems within 30 days, but information already copied into conversation history remains until those conversations are deleted.^[1][2]

Memory behavior changed with the July design. When main ChatGPT memory is enabled, conversations that use Health may create memories, but OpenAI says memories are not created directly from synced medical records or Apple Health data. A user can turn memory off, use Temporary Chat, or place health conversations in a project with project-only memory. The earlier Health Project retains its older segmented memory controls.^[2][4]

Before another connected plugin takes an action that could disclose health information, OpenAI says extra safeguards check whether the sharing matches the request and may ask for confirmation. The company reports conducting red-team exercises focused on connected Health information. These are company safety claims rather than an external certification.^[1][2]

Medical and regulatory boundaries

OpenAI states that Health is not intended for diagnosis or treatment, does not replace professional care, and should not be used for urgent symptoms or emergencies. Important information should be checked against original records and discussed with a qualified healthcare professional.^[1][2]

Health in ChatGPT is a consumer feature, separate from OpenAI's organizational healthcare offering. The archived help page says it is not intended for clinical or HIPAA covered-entity use and does not offer a Business Associate Agreement. OpenAI directs covered entities and business associates to separate products, including ChatGPT for Healthcare, which offers organizational controls and a BAA to support HIPAA-compliant use.^[2][9]

That distinction matches general U.S. Department of Health and Human Services guidance. HHS says that when a person directs a covered entity to transmit health information to an app that is neither a covered entity nor a business associate, the information is no longer protected by the HIPAA Rules after the app receives it. This does not mean consumer health data has no legal protection. The Federal Trade Commission says its Health Breach Notification Rule applies to covered personal-health-record vendors and related entities that are not governed by HIPAA, and state consumer health laws may also apply.^[4][7][8]

OpenAI's product materials describe Health in ChatGPT as a consumer feature that is not intended for diagnosis or treatment.^[1][2] FDA guidance describes intended use as part of device determination and separately asks whether an appropriate product classification exists.^[13] OpenAI's stated-use boundary therefore does not, by itself, establish the feature's classification under U.S. medical-device law.

Safety work and evaluation

OpenAI says it developed the initial Health experience with more than 260 physicians who had practiced in 60 countries and supplied over 600,000 pieces of feedback across 30 focus areas. The company also reports physician testing of the July product with connected health data before release.^[1][3]

One evaluation framework used in this work is HealthBench. OpenAI introduced the benchmark in 2025 with 5,000 simulated health conversations and custom rubrics created with 262 physicians. The rubrics cover qualities such as factual accuracy, communication, context seeking, completeness, and appropriate escalation. At the July launch, OpenAI said Free users received GPT-5.5 Instant for health conversations and paid users could use GPT-5.6 Sol, both in the broader GPT-5 family.^[1][6]

A product-specific Nature Medicine study tested the January 2026 ChatGPT Health web interface, which the researchers identified as using a gpt-5-mini thinking backbone. The study collected 960 responses from 60 clinician-authored vignettes across 21 clinical domains, with each vignette tested under 16 conditions. In the 480 responses to clear cases, ChatGPT Health undertriaged 33 of 64 emergency presentations (51.6 percent) and overtriaged 83 of 128 nonurgent presentations (64.8 percent). Accuracy was higher for semi-urgent and urgent presentations, at 93.0 percent and 76.9 percent, respectively.^[12]

The study was a structured stress test, not a trial with patients. It used synthetic vignettes, required one response in a fixed explanation-and-four-level format, did not test prompt sensitivity, and evaluated one model at one point in time. The queries ran from January 9 to January 11, before OpenAI's July redesign and model changes. The results therefore do not measure the July 2026 product or its connected-record workflow, but they directly document safety failures in the early product and show why evaluations must name the tested version and interaction format.^[1][12]

HealthBench scores are model evaluations, not clinical trials, regulatory review, or proof that users will make better decisions. OpenAI itself says ChatGPT can still make mistakes. A separate 2026 Nature Medicine study illustrates the difference between benchmark performance and real interaction: 1,298 U.K. participants using GPT-4o, Llama 3, or Command R+ did not choose more accurate levels of care than a control group and were less likely to identify relevant conditions. The study tested general-purpose models in 2024, not Health in ChatGPT or its connected-data workflow, so it cannot be treated as a product evaluation. It does support caution about applying benchmark results directly to consumer outcomes.^[10]

Limits and open questions

The most immediate limits are practical. Provider coverage varies, Apple Health requires an iPhone, some wearable metrics do not pass through Apple Health, and the integration cannot update source records. Synced records may also be incomplete or outdated. OpenAI gives the example of a medication that remains listed after a person has stopped taking it and tells users to verify important details against the original source.^[2]

The larger limitation is that personalized context does not eliminate language-model error. A response can misunderstand an incomplete record, combine reliable and unreliable advice, or fail to recognize that information is missing. Independent reporting by the Associated Press emphasized that testing of consumer health chatbots was still immature and that major medical decisions should not rely on a large language model alone. The report also noted that the Oxford study did not test the newer health products.^[10][11]

Health in ChatGPT is therefore best understood as a consumer-facing layer within AI in healthcare: it organizes and explains information the user chooses to connect, while leaving verification, diagnosis, treatment, and clinical judgment to healthcare professionals. Its usefulness depends on record completeness, permission choices, model behavior, and a user's ability to recognize when professional care is needed.

References

1. ^OpenAI. "Launching Health in ChatGPT." July 23, 2026. openai.com/...health-in-chatgpt
2. ^OpenAI Help Center. "Health in ChatGPT." Archived July 25, 2026. web.archive.org/...20001036-health-in-chatgpt
3. ^OpenAI. "Introducing ChatGPT Health." January 7, 2026, with July 23 update. openai.com/...introducing-chatgpt-health
4. ^OpenAI. "Health Privacy Notice." Updated June 29, 2026. openai.com/...health-privacy-policy
5. ^b.well Connected Health. "Connecting Your Health Records Into ChatGPT." January 8, 2026. resources.icanbwell.com/...th-records-into-chatgpt
6. ^OpenAI. "Introducing HealthBench." May 12, 2025. openai.com/...healthbench
7. ^U.S. Department of Health and Human Services. "The access right, health apps, and APIs." Reviewed May 30, 2025. hhs.gov/...access-right-health-apps-apis
8. ^U.S. Federal Trade Commission. "Complying with FTC's Health Breach Notification Rule." ftc.gov/...-ftcs-health-breach-notification-rule-0
9. ^OpenAI. "Introducing OpenAI for Healthcare." January 8, 2026. openai.com/...openai-for-healthcare
10. ^Bean, Andrew M., et al. "Reliability of LLMs as medical assistants for the general public: a randomized preregistered study." Nature Medicine 32, 609-615. Published February 9, 2026. doi.org/...s41591-025-04074-y
11. ^Matthew Perrone. "What to know before asking an AI chatbot for health advice." Associated Press, March 2, 2026. apnews.com/...llm-1008892e0eb8ef4dbab4818beb15daef
12. ^Ramaswamy, Ashwin, et al. "ChatGPT Health performance in a structured test of triage recommendations." Nature Medicine 32, 1671-1675. Published February 23, 2026; version of record March 12, 2026. doi.org/...s41591-026-04297-7
13. ^U.S. Food and Drug Administration. "How to Determine if Your Product is a Medical Device." Archived July 13, 2026. web.archive.org/...-if-your-product-medical-device