How to Use Healthcare Public Data in ChatGPT and Codex: Complete Setup and Evidence-Safe Research Tutorial

What OpenAI launched on September 1, 2026

On September 1, 2026, OpenAI announced general availability of healthcare plugins for eligible U.S. users and workspaces, including a Healthcare Public Data capability for ChatGPT and Codex. OpenAI describes Healthcare Public Data as a read-only way to search nine public healthcare sources for research, clinical trials, medication information, Medicare data, and provider records. The operationally important boundary is simple: this is public-source retrieval and analysis, not patient-chart access, not an EHR connection, and not a route for sending protected health information to public providers.

This tutorial is written for teams that want to use public healthcare evidence without crossing data-governance lines. A clinical researcher might compare trial eligibility language across ClinicalTrials.gov records, a pharmacist might verify medication-label language against DailyMed and RxNorm, an operator might combine Medicare coverage evidence with public literature for a population-health planning memo, and a developer might use Codex to draft a reproducible evidence-extraction script or documentation workflow. In every case, the source material remains public and the output must be verified against the underlying public record before it influences clinical, regulatory, reimbursement, operational, or product decisions.

Operational warning: Do not send patient names, dates of birth, medical record numbers, member numbers, encounter details, chart excerpts, free-text notes, or other protected health information to Healthcare Public Data. OpenAI’s public-data healthcare apps are read-only public-source tools; a workspace Business Associate Agreement does not convert public data providers into approved recipients for PHI.

OpenAI’s September 1 healthcare update also introduced a separate Epic plugin path for eligible ChatGPT for Healthcare and HIPAA-enabled Enterprise workspaces. Epic access is not the same product boundary as Healthcare Public Data. According to OpenAI, the Epic plugin can provide read-only access to authorized patient information only after administrator configuration, organization-specific setup, individual Epic sign-in, and enforcement of the clinician’s existing patient-chart permissions. It cannot update records, place orders, message patients, or expand a user’s chart access beyond what the EHR already permits.

For ChatGPT Healthcare Workflows, ChatGPT Health Launches: How OpenAI’s Medical Records Integration Changes Personal Healthcare in 2026 is the most relevant adjacent resource. The ChatGPT Health launch analysis covers medical-record integration and patient-facing workflow implications, which helps distinguish those private-record use cases from this tutorial’s public, non-patient evidence sources.

Availability, eligibility, and the nine read-only public-source apps

OpenAI’s release notes describe Healthcare Public Data as available to eligible U.S. users and workspaces, including eligible ChatGPT for Clinicians users in the United States. OpenAI also states that eligible ChatGPT for Healthcare and HIPAA-enabled Enterprise workspaces can use healthcare plugins under approved configurations, with Epic governed separately from public-source access. Administrators should not assume that a user’s plan, workspace status, BAA status, plugin policy, and individual app access are the same control; OpenAI explicitly treats plugin availability and app access as separately governed.

The Healthcare Public Data capability connects to nine official public healthcare sources. OpenAI’s help materials specifically identify examples that include ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed, and PubMed. The same materials describe use cases involving clinical-trial comparison, medication-label and warning verification, and combining research, trials, and Medicare coverage evidence for population-health planning. Because source availability and app access can be governed independently, teams should verify the current in-product list and enabled status before documenting a workflow, training users, or relying on a particular public source for a regulated process.

Healthcare Public Data does not require separate accounts with the public source organizations, but that does not mean every app is automatically enabled for every user. Installing or allowing the healthcare public-data plugin is not the same as confirming that each of the nine public-source apps is accessible, approved for a role, and appropriate for a given task. In practice, an administrator should treat each source selector as a governed evidence channel: identify the allowed roles, permitted use cases, prohibited inputs, citation requirements, and escalation path when the model output conflicts with the underlying source.

What public healthcare data can and cannot do

Public healthcare data is useful when the question can be answered from public evidence rather than patient-specific records. A safe example is: “Compare the publicly posted inclusion and exclusion criteria for these trial identifiers and return a table with direct citations.” Another safe example is: “Check DailyMed and RxNorm for the current public label terminology and ingredient normalization for this medication name.” These tasks are evidence-retrieval and synthesis tasks; they do not require patient identifiers, chart notes, or insurance-member details.

Public healthcare data becomes unsafe or insufficient when the task depends on a real patient’s identity, current chart, coverage file, or clinical context. Do not ask a public-data app whether a named patient qualifies for a trial, whether a specific member is covered, whether a patient’s medication list has a contraindication, or whether a documented diagnosis supports a reimbursement decision. Those questions require approved internal systems, authorized patient access, human clinical review, and organization-specific policy. Public sources can help define the public evidence standard, but they cannot validate a patient-specific decision without the authorized private record.

Public sources can also be incomplete, outdated, or ambiguous relative to the operational question being asked. ClinicalTrials.gov records may contain eligibility text that requires sponsor clarification, medication labels may change over time, Medicare coverage material may require jurisdiction-specific interpretation, and PubMed abstracts may not contain enough detail to support a protocol decision. Treat ChatGPT or Codex output as an evidence-navigation aid: require links or source identifiers, compare the answer with the original public record, document the retrieval date, and assign a qualified reviewer before the result is used in clinical care, compliance, reimbursement, or product claims.

Preflight governance checklist before you enable or use it

For Healthcare AI Governance, The Enterprise Guide to GPT-5.5 Instant for Healthcare: Clinical Applications, Safety Protocols, and Implementation Strategies is the most relevant adjacent resource. The enterprise healthcare implementation guide focuses on safety protocols, workspace controls, and deployment strategy, supplying the governance layer needed before teams enable public healthcare apps.

1. Confirm eligibility and geography. Verify that the user or workspace is eligible under OpenAI’s current availability rules, including the U.S. eligibility requirement for ChatGPT for Clinicians users described by OpenAI.
2. Separate public-data access from EHR access. Document whether the workflow uses Healthcare Public Data, Epic, or both. Do not treat a public-source workflow as HIPAA-enabled patient-chart access.
3. Verify workspace configuration. For ChatGPT for Healthcare or HIPAA-enabled Enterprise environments, confirm the approved workspace configuration and applicable Business Associate Agreement before any PHI is used in approved healthcare workflows. Do not send PHI to public-source apps.
4. Check plugin and app permissions separately. Confirm that the healthcare plugin is available, that the relevant public-source apps are enabled for the user role, and that administrators have approved the specific use case.
5. Define prohibited inputs. Block patient names, dates of birth, addresses, medical record numbers, member IDs, claim numbers, encounter dates tied to a person, chart excerpts, screenshots, and free-text notes from public-data prompts.
6. Require source-grounded outputs. Instruct users to request source names, identifiers, retrieval dates where available, and direct evidence snippets that can be checked against the public record.
7. Assign human review. Name the role responsible for verifying outputs before use, such as a clinician, pharmacist, regulatory specialist, researcher, reimbursement analyst, or data-governance lead.
8. Control Codex context. When using Codex, keep repositories, sample files, logs, and prompts free of PHI unless the workspace and workflow are explicitly approved for that data class. Public-data automation should use public identifiers, synthetic examples, or de-identified operational templates.
9. Preserve audit artifacts. Save the prompt, source list, generated table or code, reviewer notes, and final disposition according to your organization’s retention policy.
10. Define escalation rules. Escalate when sources conflict, when the model cannot cite the underlying record, when a public result is being used for a patient-specific decision, or when a user attempts to include PHI in a public-source workflow.

The rest of this tutorial will build on that boundary: use Healthcare Public Data as a governed public-evidence layer in ChatGPT and Codex, keep PHI out of public-source prompts, verify every consequential answer against the underlying source, and reserve patient-chart workflows for separately approved Epic-connected environments with inherited EHR permissions.

Setup sequence for administrators and end users

Set up Healthcare Public Data as a governed research capability, not as a general-purpose clinical note workspace. OpenAI describes the Healthcare Public Data plugin as a read-only way for eligible United States users and workspaces to search nine official public healthcare sources for research, clinical trials, medication information, Medicare data, and provider records. That means the setup goal is narrow: allow authorized users to query public evidence sources while preventing protected health information, local patient identifiers, or chart excerpts from being routed to public-data apps.

The safest implementation pattern is to treat plugin policy, role access, app enablement, and user connection as separate gates. A workspace administrator decides whether the organization permits the plugin class; a role or group decision determines which users can see and use it; individual source apps must be enabled or made available; and each user still needs to install or connect the relevant public-data apps before using them in ChatGPT or Codex. OpenAI’s healthcare release notes also state that plugin availability and app access are governed separately, so administrators should not assume that installing a plugin automatically makes every underlying source usable.

Administrator setup: approve the plugin before users install it

1. Confirm workspace eligibility and intended use. Verify that the workspace or account type is eligible for Healthcare Public Data and that users are in the United States where OpenAI has made the feature generally available for eligible users and workspaces. If the workspace is not eligible, do not work around the restriction by copying public-data outputs from another account into regulated workflows.
2. Document the permitted use cases. Write a short internal policy that limits the plugin to public evidence retrieval, such as PubMed literature checks, ClinicalTrials.gov trial discovery, DailyMed or RxNorm medication-reference lookup, CMS coverage research, and Medicare or provider-record review where applicable. The policy should explicitly prohibit patient names, dates of birth, medical record numbers, member numbers, addresses, chart excerpts, appointment histories, and any other PHI.
3. Set plugin policy at the workspace level. In the administrator controls for apps and plugins, allow Healthcare Public Data only if the organization has approved public-source searching inside ChatGPT and Codex. Keep unrelated plugins disabled unless they have gone through the same review. For a broader administrative pattern, see

   For ChatGPT Plugin Administration, How to Import and Govern ChatGPT Plugin Marketplaces from GitHub: Enterprise Workspace Playbook is the most relevant adjacent resource. The GitHub plugin-marketplace governance playbook explains how workspace owners control plugin availability, app permissions, synchronization, and rollout, reinforcing why installation and app access remain separate decisions.

   .

4. Assign role access using least privilege. Grant access to clinicians, researchers, formulary reviewers, clinical operations analysts, population-health staff, or trial-navigation teams only when their job function requires public healthcare evidence. Do not enable access simply because the plugin is read-only; read-only tools can still receive sensitive input if users paste the wrong content.
5. Enable only the source apps that match the workflow. If a team only needs publication and trial research, enable the public-source apps relevant to those tasks rather than assuming all nine sources should be active. If a billing-policy group needs Medicare coverage evidence, enable the appropriate coverage source for that group. Source-level enablement makes it easier to audit why a source was available and to disable it later without removing the entire plugin capability.
6. Review the user-facing warning text before rollout. Require administrators, compliance leads, and clinical champions to read the warnings shown during installation or connection. OpenAI explicitly warns not to send protected health information to public sources. Put that warning into local training materials rather than relying on users to remember it from the product flow.
7. Separate Healthcare Public Data from Epic deployment decisions. Healthcare Public Data does not access patient charts and does not require separate accounts with the public source organizations. Epic is a separate read-only plugin for approved ChatGPT for Healthcare and HIPAA-enabled Enterprise workspaces, requires organization-specific FHIR/OAuth setup, requires individual Epic sign-in, and inherits existing chart permissions. A Business Associate Agreement for the workspace does not authorize users to send PHI to public providers.

Operational warning: Do not treat public-source access as a HIPAA workflow just because it is available inside a healthcare workspace. Healthcare Public Data is for public evidence. If a prompt requires a patient’s chart, patient-specific history, or identifiers, stop and use only an approved patient-data integration that the organization has configured for that purpose.

Why installation, app enablement, and connection are different controls

This separation is useful because different risks appear at different stages. Plugin policy addresses enterprise exposure; role access addresses job necessity; app enablement addresses source fit; and user connection addresses the last-mile act of using a source in a conversation or Codex task. If a medication-safety team needs DailyMed and RxNorm but not provider-record lookups, source-level enablement avoids unnecessary exposure. If a research team’s project ends, the administrator can remove access for that role without changing the whole organization’s configuration.

End-user setup in ChatGPT: install, connect, and test without PHI

1. Start in the approved workspace. Confirm that you are using the organization-approved ChatGPT workspace or eligible ChatGPT for Clinicians account, not a personal account used for unrelated work. If your task involves patient-specific facts, do not proceed with Healthcare Public Data.
2. Find the Healthcare Public Data plugin or app entry. Use the product’s plugin or app discovery area to locate Healthcare Public Data and the available public-source apps that your administrator has enabled. If you cannot see the plugin, ask the workspace administrator to verify eligibility, policy, role access, and app enablement rather than attempting to use another user’s output.
3. Read the warning before installation. The warning is not a formality. It is the point where you confirm that you will not send PHI to public sources. Translate the rule into practical behavior: never paste a patient’s name, date of birth, appointment note, lab narrative, insurance member number, medical record number, phone number, full address, or free-text chart excerpt.
4. Install the plugin if your workspace permits it. Installation makes the plugin available to your account or session according to the workspace policy. It does not automatically connect every source app and does not mean the model will always choose the correct public source without instruction.
5. Connect only the individual apps you need. If your task is trial discovery, connect the trial source app your administrator has enabled. If your task is medication-reference verification, connect the medication-reference source apps that match the question. OpenAI’s help materials describe Healthcare Public Data as not requiring separate accounts with the source organizations, so this connection step is a product-level connection, not a separate login to each public source.
6. Use “Try in chat” for a non-sensitive test. After installation and connection, open the app through the product’s “Try in chat” path if available. Test with a generic, non-patient prompt, such as a public disease term, a medication name, or a coverage topic. Do not use a real patient scenario as the first test.
7. Target the source explicitly with @ when appropriate. In ChatGPT, use the supported source-targeting syntax, such as @, to call the relevant public-source app rather than relying on a broad web-style answer. Explicit targeting is helpful when you need a verifiable output from ClinicalTrials.gov, PubMed, DailyMed, RxNorm, CMS Coverage, or another enabled public source.

Sample ChatGPT setup test — non-PHI only

@Healthcare Public Data
Find public evidence about current ClinicalTrials.gov studies for adults with migraine prevention.
Return:
1. Search terms used
2. Trial identifiers or public records found
3. Eligibility criteria that are visible in the public record
4. A warning if the source does not answer part of the question
5. Links or citations I can verify manually

The test prompt above is intentionally generic. It does not include a patient’s age, name, location, appointment date, diagnosis timeline, prior authorizations, medication fill history, or clinical note. If your actual workflow needs those details, Healthcare Public Data is the wrong input channel. You can still use public sources to understand evidence, labels, trials, or coverage policies, but the patient-specific matching and clinical decision process must remain inside approved systems and professional review.

Codex setup: create a bounded task that can use public healthcare sources

Codex setup should be stricter than ordinary chat setup because Codex tasks often produce files, scripts, data transformations, or pull-request-ready artifacts. Before starting, define the task as a public-evidence engineering or research-support task, not as a patient-care automation. Examples include building a public-source evidence table, generating a reproducible search log, creating a script that formats ClinicalTrials.gov public records, or preparing a non-PHI review packet for human researchers.

1. Create a task with a narrow objective. State exactly what Codex should produce: a table, JSON schema, search log, draft analysis, or code change. Avoid open-ended instructions such as “find everything relevant” because broad tasks encourage unsupported synthesis.
2. Declare the PHI exclusion rule in the task instructions. Tell Codex that inputs and outputs must contain no patient identifiers, chart text, member data, appointment details, or local case descriptions. If the task file contains sample data, use synthetic examples clearly labeled as synthetic.
3. Connect the necessary Healthcare Public Data apps before running the task. Confirm that the required source apps are available in the Codex environment according to administrator policy. Installation alone is not enough if the task needs a specific public source that has not been enabled or connected.
4. Select a source with $ when using Codex source targeting. Where Codex supports source selection with $, reference the specific public-data source needed for the task. Use targeted source calls for evidence retrieval and keep local files limited to non-PHI project artifacts.
5. Require a verification artifact. Ask Codex to include source names, record identifiers where available, search terms, retrieval date, assumptions, and unanswered questions. Do not accept a clinical or policy conclusion that cannot be traced back to the underlying public record.
6. Require human review before reuse. If Codex creates code, a table, or a narrative summary, a qualified human must review the evidence, source fit, and limitations before the artifact is used in clinical, operational, research, or compliance work.

Sample Codex task — public evidence extraction only

Objective:
Create a CSV schema and extraction script for public ClinicalTrials.gov records about hypertension device trials.

Data boundary:
Use public records only. Do not use patient names, MRNs, dates of birth, chart notes, claims files, member IDs, addresses, or any PHI.

Source targeting:
Use $Healthcare Public Data or the enabled ClinicalTrials.gov public-source app if available.

Required output:
- trial_id
- public_title
- condition
- intervention
- recruitment_status
- eligibility_summary
- source_url
- retrieval_date
- fields_not_found

Verification:
Include a README that lists search terms, source names, and limitations. Do not infer eligibility for any patient.

The Codex pattern works best when the output is auditable. A script that extracts public trial metadata can be reviewed, tested, and rerun. A free-form conclusion such as “this patient qualifies” is not appropriate for Healthcare Public Data because it would require patient information and clinical judgment. Keep Codex focused on source retrieval, formatting, comparison, and evidence-log construction.

Final setup checks before the first real workflow

• Policy check: The workspace administrator has approved Healthcare Public Data and documented that public-source apps must not receive PHI.
• Role check: Only users with a legitimate research, clinical-reference, operations, or administrative evidence need have access.
• Source check: The specific public-source apps needed for the workflow are enabled and connected; unused sources remain disabled where possible.
• Prompt check: The first production prompt contains only public concepts, generic clinical terms, public identifiers, medication names, trial topics, coverage questions, or provider-record queries that do not identify a patient.
• Evidence check: Outputs will be verified against the underlying public records before being used in a memo, workflow, patient conversation, research packet, policy analysis, or code artifact.

If any check fails, pause the rollout rather than asking users to self-police after launch. The feature is valuable precisely because it can bring public healthcare evidence into ChatGPT and Codex, but the operational boundary is non-negotiable: public data sources are for public evidence, and patient-specific data belongs only in approved, configured patient-data systems with the required organizational safeguards.

Evidence-safe workflows for public healthcare research in ChatGPT and Codex

OpenAI describes Healthcare Public Data as a read-only way for eligible users to search nine public healthcare sources from ChatGPT and Codex for research, clinical trials, medication information, Medicare data, and provider records. Treat that capability as an evidence-retrieval layer, not as an authority that decides care, coverage, licensure, enrollment, or causality. The safe operating pattern is to ask for a source-specific answer, require citations or source identifiers, verify the result against the originating public record, and document any unresolved ambiguity before the output is used in a clinical, operational, research, or product decision.

Operational boundary: do not enter protected health information into Healthcare Public Data workflows. Do not include patient names, dates of birth, medical record numbers, member IDs, contact details, appointment details, free-text chart excerpts, or rare combinations of facts that could identify a person. A HIPAA-enabled workspace or Business Associate Agreement does not make public-source apps appropriate for PHI; OpenAI distinguishes these public sources from separate Epic access, which is read-only, permission-bound, and organization-configured.

Source-by-source workflow table

Workflow 1: literature-to-trial evidence scan without PHI

Use this workflow when a researcher, clinician, founder, or product team needs a rapid map of public evidence for a condition, intervention, or medication class. Start with a de-identified question such as “adult migraine prevention with CGRP-targeted therapies” rather than a patient story. Ask ChatGPT to separate PubMed literature, ClinicalTrials.gov records, and medication-label evidence into different sections so that trial status, published evidence, and labeling are not blended into one unsupported conclusion.

1. Define the research frame. State the condition, intervention, comparator, outcomes, population boundaries, and date range if known. Avoid patient-specific facts and use broad eligibility concepts such as age category or diagnosis category only when necessary.
2. Retrieve PubMed evidence first. Ask for study type, PMID, population, endpoints, and a one-sentence limitation for each record. Require the answer to distinguish randomized trials, observational studies, reviews, editorials, and case reports.
3. Retrieve ClinicalTrials.gov records second. Ask for NCT identifiers, recruitment status as shown in the source, inclusion and exclusion themes, interventions, and whether results appear in the public record if available.
4. Retrieve DailyMed and RxNorm evidence third. Use DailyMed for label language and RxNorm for normalization, then ask ChatGPT to identify where product names, formulations, strengths, or routes might change the interpretation.
5. Build an evidence ledger. Store each claim as a row with source, identifier, exact supporting text, verification status, and reviewer initials. Do not move the claim into a report until the source record has been opened and checked.

Example research prompt:

Use Healthcare Public Data only. Do not use or request patient-identifying information.

Research question: For adults with [condition], what public evidence exists for [intervention or medication class] compared with [comparator]?

Return three separate sections:
1. PubMed: up to 10 relevant records with PMID if available, study type, population, endpoints, and one limitation per record.
2. ClinicalTrials.gov: relevant trials with NCT ID, recruitment status as shown, intervention, inclusion/exclusion themes, and whether the record appears active, completed, or otherwise listed.
3. DailyMed/RxNorm: label or normalization points that affect interpretation, including formulation, route, warnings, or concept ambiguity.

For every claim, include the source record I must verify. Do not infer causality from non-randomized evidence. Do not make patient-specific recommendations.

For ChatGPT Clinical Research Prompts, GPT-5.5 Prompts for Healthcare: Clinical Decision Support and Medical Research is the most relevant adjacent resource. The healthcare prompting guide provides clinical-research and evidence-synthesis patterns that can be adapted to PubMed and ClinicalTrials.gov once source targeting and verification rules are in place.

Workflow 2: medication safety and terminology reconciliation

Use DailyMed, RxNorm, PubMed, and openFDA together when the task is to reconcile medication terminology or prepare a medication-safety backgrounder. DailyMed should anchor label language, RxNorm should normalize drug concepts, PubMed should provide published context, and openFDA should be treated as a public safety-data exploration source rather than proof of incidence or causality. This separation prevents a common failure mode: mixing label warnings, adverse-event reports, and published studies into a single overconfident safety claim.

1. Start with the exact product question. Specify ingredient, brand or generic level, route, dosage form, and strength if those details are already public and non-identifying.
2. Ask DailyMed for label-backed statements. Require section names and exact supporting language, especially for boxed warnings, contraindications, warnings and precautions, dosage, and use in specific populations.
3. Ask RxNorm to normalize terminology. Require the model to flag concept-level uncertainty, such as whether a term refers to an ingredient, clinical drug, branded drug, or pack.
4. Ask openFDA for hypothesis-generating signals only. Require a written caveat that report counts do not equal event rates and cannot establish causation.
5. Ask PubMed for peer-reviewed context. Require study type and endpoint details so that case reports are not weighted like controlled studies.

Example research prompt:

Use Healthcare Public Data to create a medication evidence brief for [public medication name].

Separate the answer into:
- DailyMed label evidence: exact label sections and language I should verify.
- RxNorm normalization: ingredient, branded or generic concepts, and ambiguity to resolve.
- PubMed context: relevant studies or reviews with PMID if available and study type.
- openFDA exploration: public safety-report themes only, with an explicit warning that reports do not prove causality or incidence.

Do not provide patient-specific advice. Do not ask for patient medication lists, dates, identifiers, allergies, or chart excerpts.

Workflow 3: coverage, utilization, and provider-market research

Use CMS Coverage, CMS Open Data, Medicare Care Compare, and the NPI Registry when the job is operational research rather than clinical evidence grading. These sources can help a team understand public Medicare policy language, public program datasets, quality-measure records, and provider identity data. They should not be used as a shortcut for individual benefits decisions, licensure determinations, contracting status, credentialing, or claims-payment predictions.

1. Coverage policy review: Ask CMS Coverage for documents related to a service, device, diagnostic, procedure category, or condition. Require the response to quote the specific public policy language that supports any coverage statement.
2. Dataset exploration: Ask CMS Open Data for relevant dataset names, periods, field definitions, and aggregation assumptions. Require the model to list filters and denominators before calculating or comparing figures.
3. Quality-profile review: Ask Medicare Care Compare for public measures relevant to a provider or facility category. Require measure names, reporting periods if shown, and comparability warnings.
4. Provider matching: Ask the NPI Registry for identity and taxonomy fields, then verify the NPI record directly before importing it into a directory or CRM.
5. Decision separation: Label outputs as “public research,” “requires payer verification,” “requires licensure verification,” or “requires measure-methodology review” before they enter a workflow queue.

Example operations prompt:

Use Healthcare Public Data to prepare a public Medicare research memo about [service, provider category, or market].

Return four sections:
1. CMS Coverage: relevant public coverage documents and the exact policy language I must verify.
2. CMS Open Data: datasets that could support utilization or market analysis, with fields, time periods, filters, and denominator caveats.
3. Medicare Care Compare: relevant public quality or profile measures, including reporting-period and comparability limitations.
4. NPI Registry: provider or organization identity fields and taxonomy details to verify.

Do not state that a patient, member, or claim is covered. Do not state that a provider is licensed, credentialed, in network, or accepting patients unless that fact is verified through the proper authoritative process outside this public-data workflow.

Codex workflow: create a reproducible evidence ledger, not an autonomous conclusion

Codex is useful when the output should become a reproducible artifact: a CSV evidence ledger, a validation script, a notebook template, or a pull request that adds source-checking logic to an internal tool. Keep the task bounded: Codex may gather public-source context, structure data, propose tests, and prepare code for human review, but it should not decide clinical guidance, publish external conclusions, approve benefits, or deploy production changes without review.

Recommended Codex task specification:

Task: Create a reproducible evidence-ledger template for a Healthcare Public Data research project.

Scope:
- Sources may include PubMed, ClinicalTrials.gov, DailyMed, RxNorm, openFDA, CMS Coverage, CMS Open Data, Medicare Care Compare, and NPI Registry.
- No PHI, patient-level facts, member IDs, chart excerpts, or private records.
- Output a CSV schema, a README, and validation checks.

Definition of done:
- Each row can store source name, source identifier, claim, quoted support, retrieval date, verification status, reviewer, and limitation.
- Validation fails if source identifier, verification status, or limitation is blank.
- README explains that public sources are read-only and must be verified against originating records.
- No clinical, benefits, licensure, enrollment, or causality decisions are automated.

A minimal evidence-ledger schema can make audits easier because every downstream claim remains attached to a source and a reviewer. The following example is a proposed structure, not an OpenAI product requirement:

source_name,source_identifier,topic,claim,quoted_support,retrieval_date,verification_status,reviewer,limitation,decision_boundary
PubMed,PMID-or-record-id,example topic,example claim,exact text to verify,YYYY-MM-DD,pending,,study design limits causal inference,not medical advice
ClinicalTrials.gov,NCT-number,example topic,example trial status,exact status text to verify,YYYY-MM-DD,pending,,status may change before enrollment,not enrollment confirmation
CMS Coverage,document-id-or-title,example topic,example policy statement,exact policy text to verify,YYYY-MM-DD,pending,,not an individual benefits decision,requires payer verification

Verification steps before a public-data answer leaves the workspace

Use the same verification gate for clinician education, research memos, founder diligence, product requirements, and enterprise operations. The reviewer should open the originating public record, compare the model’s summary with the source text, and mark each claim as verified, corrected, excluded, or unresolved. If the answer combines sources, the reviewer should verify that the synthesis does not convert “associated with,” “reported in,” “listed as recruiting,” or “covered under conditions” into a stronger claim than the source supports.

1. Confirm source identity. Check that every material claim names the public source and provides a usable identifier, such as a PMID, NCT ID, label reference, NPI, dataset name, or coverage-document title where available.
2. Confirm exactness. Compare the summary against the source text for population, intervention, comparator, outcome, status, label section, measure definition, date, and product or provider identity.
3. Confirm timeliness. Recheck ClinicalTrials.gov recruitment status, DailyMed label text, CMS policy language, Care Compare measures, and NPI records immediately before operational use because public records can change.
4. Confirm decision boundary. Add a note when the output cannot answer causality, current enrollment, individual coverage, licensure, credentialing, network status, or quality comparability.
5. Confirm PHI exclusion. Inspect the prompt and output for names, dates, identifiers, chart excerpts, or private facts before saving, sharing, exporting, or using the conversation as an internal example.

For AI Evidence Verification, 50 GPT-5.5 Prompts for Healthcare Professionals: Clinical Decision Support, Medical Documentation, Patient Communication, and Research Analysis is the most relevant adjacent resource. The healthcare-professional prompt collection includes research-analysis and documentation workflows, showing where evidence extraction ends and clinician verification and interpretation must begin.

Limitations to state explicitly in every deliverable

• Causality: PubMed abstracts, openFDA reports, and observational public records can generate hypotheses, but they do not automatically prove that an exposure caused an outcome or that one intervention is superior to another.
• Current trial enrollment: ClinicalTrials.gov records can change, and a recruitment status shown in a public record is not a guarantee that a site is accepting a specific person or that eligibility will be met.
• Benefits decisions: CMS Coverage and public Medicare data can inform policy research, but they are not individual coverage determinations, prior authorization approvals, coding advice, or payment guarantees.
• Licensure and credentialing: NPI Registry records identify providers and organizations for administrative purposes, but they do not by themselves establish active licensure, privileges, board certification, sanctions status, employment, or network participation.
• Measure comparability: CMS Open Data and Medicare Care Compare measures may differ by reporting period, denominator, exclusions, provider type, case mix, geography, and methodology, so cross-provider or year-over-year comparisons require measure-level review.
• Product and formulation specificity: DailyMed and RxNorm workflows must preserve ingredient, brand, dose form, route, and labeler distinctions, because medication evidence can change when the product being analyzed changes.

The safest final deliverable is not a single confident paragraph; it is a source-indexed brief that shows what was asked, which public sources were searched, what each source supports, what remains unverified, and which decisions are out of scope. That structure lets clinicians, researchers, administrators, and product teams use Healthcare Public Data for faster evidence retrieval while preserving the human review and source verification that public healthcare information requires.

Safety boundary: treat public healthcare sources as public providers, not patient systems

The safest operating rule is simple: Healthcare Public Data is for public evidence retrieval, not patient-specific care coordination. OpenAI describes the Healthcare Public Data plugin as a read-only way to search nine official public healthcare sources for research, clinical trials, medication information, Medicare data, and provider records. It does not access patient charts, and it should not be used as a bridge between a patient record and public databases unless every patient detail has been removed before the query is sent.

Do not paste, upload, summarize, encode, hash, or “lightly anonymize” protected health information into a Healthcare Public Data prompt. Names, dates of birth, medical record numbers, member IDs, accession numbers, encounter dates, precise appointment details, addresses, phone numbers, email addresses, full-face images, rare-disease narratives, and free-text chart excerpts can identify a patient alone or in combination. A query such as “find trials for a 43-year-old woman with metastatic condition X treated at Hospital Y on May 12” may still be identifiable even without a name because the combination of demographics, diagnosis, institution, and dates can narrow to one person.

Use synthetic or generalized descriptors when public evidence is enough. Instead of “our patient John Smith, DOB 4/4/1974, failed drug A and has creatinine 1.9,” use a non-patient prompt such as “summarize public trial eligibility criteria for adults with condition X who have prior exposure to drug A; list renal-function thresholds stated in the trial records and cite the source records.” The second prompt asks for public criteria and requires the human user to compare those criteria against the patient chart inside an approved clinical workflow.

Operational rule: if the answer requires the model or a public-source provider to see patient identity, chart text, appointment history, insurer identifiers, or uniquely identifying clinical details, do not use Healthcare Public Data for that step. Use an approved EHR-connected workflow only if your organization has configured it, authorized the user, confirmed the applicable Business Associate Agreement, and approved the specific use case.

For HIPAA Safe AI Use, OpenAI’s Healthcare Gambit: Privacy, Trust, and the Future of AI-Powered Personal Medicine is the most relevant adjacent resource. The healthcare privacy and trust analysis examines the consequences of mixing sensitive records with AI services, giving practical context to the strict rule that public-data searches must exclude PHI.

What data may be sent outside your workspace during public-source retrieval

When a user asks ChatGPT or Codex to query public healthcare sources, the query terms, selected source targets, and retrieval context may be sent to the public data providers or their associated interfaces so the requested search can run. That is a different data path from asking the model to reason over text already inside the chat. Administrators should assume that source-directed search terms can leave the workspace boundary for the purpose of retrieval unless OpenAI’s current product documentation and the customer’s contract say otherwise.

This distinction matters for compliance review. A broad literature query such as “PubMed evidence for guideline-relevant endpoints in condition X” is typically a public-information task. A query that includes “MRN 12345,” a beneficiary number, a full adverse-event narrative copied from a chart, or a rare combination of dates and locations is not a public-information task. The public provider does not become part of your HIPAA workflow simply because the request was initiated from an approved ChatGPT workspace.

OpenAI’s Healthcare Public Data help materials state that the public-data plugin does not require separate accounts with the source organizations. That convenience should not be confused with authorization to transmit regulated data. No separate account means users can reach the public sources through the plugin without logging into each source provider; it does not mean those public providers are covered by your organization’s BAA with OpenAI, and it does not mean the public sources can receive PHI.

BAA boundaries: Healthcare Public Data is not the Epic plugin

OpenAI distinguishes Healthcare Public Data from the Epic plugin. Healthcare Public Data searches public sources and must not receive PHI. The Epic plugin is a separate read-only integration for eligible ChatGPT for Healthcare and HIPAA-enabled Enterprise workspaces, subject to organization-specific setup, plugin policy, individual Epic sign-in, and the clinician’s existing chart permissions. OpenAI states that organizations must confirm an applicable Business Associate Agreement and approved workspace configuration before using Epic with protected health information.

A Business Associate Agreement for a qualifying workspace is not a universal pass to send PHI anywhere in the interface. It does not convert public sources into business associates, does not authorize the Healthcare Public Data plugin to receive patient identifiers, and does not bypass the need for approved configuration. Treat the BAA as one required component of a specific regulated workflow, not as a blanket permission across every plugin, app, model mode, or connected provider.

Epic access, where available and approved, is also read-only under OpenAI’s description. It does not update records, place orders, message patients, or expand a clinician’s chart access. If a user cannot access a patient chart in Epic, the plugin should not provide a workaround. If an answer suggests a clinical action, the clinician must verify the underlying EHR facts and public evidence in the source systems before acting.

Audit logs, Compliance API workflows, and evidence records

OpenAI states that ChatGPT for Healthcare includes audit logs. Administrators should use those logs to monitor plugin use, investigate incidents, and confirm that users are not submitting patient identifiers to public-source workflows. If your plan exposes a Compliance API, log export, SIEM connector, or eDiscovery integration, treat it as part of the governance system, but verify exact availability, fields, retention, export cadence, and permissions in OpenAI’s current documentation and your contract before relying on it for regulatory controls.

Logging should capture enough operational detail to reconstruct a questionable workflow: who initiated the task, which workspace and plugin policy applied, which source apps were targeted, whether Codex or ChatGPT was used, what files were attached, and what final output was shared. Do not assume that logs preserve every retrieved record or every source-page version. For research-grade work, maintain an evidence ledger in the deliverable itself with source names, record identifiers where available, retrieval dates, quoted passages, and uncertainty notes.

Administrators should define escalation thresholds before rollout. Examples include any prompt containing obvious identifiers, repeated no-result searches with increasingly specific demographic combinations, file uploads from EHR exports, or attempts to combine Epic-derived patient context with public-data source calls in an unapproved workflow. The escalation path should identify the workspace owner, privacy officer, security contact, and clinical lead who can decide whether containment, user retraining, or formal incident review is required.

Retention, residency, and model-improvement caveats to verify before production use

Retention and data-residency settings depend on the customer’s plan, workspace configuration, contractual terms, and the specific feature path being used. Do not copy retention assumptions from a different OpenAI product tier or from an unrelated temporary-chat workflow. Before enabling Healthcare Public Data in a regulated organization, confirm where conversation content, plugin interaction metadata, audit logs, uploaded files, and Codex task artifacts are stored, how long they are retained, who can export them, and what deletion mechanisms apply.

Temporary chats require separate attention. OpenAI’s release notes state that temporary chats can optionally use existing memory, plugins, and custom instructions when personalization is selected at the start; that choice cannot be changed later. Personalized temporary chats do not create new memories and remain out of history unless saved, but saving converts the conversation into a regular chat governed by account-level personalization and model-improvement settings. A temporary chat is therefore not a substitute for a compliant data-handling policy.

Data residency should be verified at the workspace and contract level, not inferred from a user’s location or the fact that public sources are US healthcare sources. If your organization has residency obligations, require administrators to document the applicable OpenAI commitments, the scope of any regional processing controls, the treatment of connected public-source requests, and the handling of logs or exports. If the documentation is ambiguous, pause regulated use until procurement, legal, and security teams obtain a written answer.

Troubleshooting common setup and retrieval failures

How to handle no-result or low-confidence searches

A no-result answer is not evidence that a trial, label warning, coverage policy, publication, or provider record does not exist. It means the current query, source selection, plugin path, and retrieval session did not surface a result. Require users to report no-result searches as “not found in the queried sources using the documented search terms on the retrieval date,” not as “there is no evidence.”

Use a structured fallback sequence. First, broaden non-identifying terms, such as using the generic condition name instead of an uncommon local abbreviation. Second, search source-specific synonyms, such as drug ingredient names, brand names, RxNorm concepts, and MeSH-like terminology where appropriate. Third, query the official source directly in a browser if the result will support a clinical, operational, or business decision. Fourth, document remaining uncertainty and route the item to a qualified reviewer rather than asking the model to infer the missing fact.

For clinical-trial screening, do not let the model decide eligibility from public criteria alone. Public trial records can help identify inclusion and exclusion criteria, enrollment status, locations, interventions, and contacts, but the responsible clinician or research coordinator must verify the current record and compare it with the patient chart inside approved systems. If patient-specific matching is needed, perform the matching in an authorized clinical or research workflow, not by pasting chart facts into a public-data prompt.

Plan, workspace, and provider limits to communicate to users

OpenAI’s September 2026 materials describe Healthcare Public Data as available to eligible ChatGPT for Clinicians users in the United States and eligible workspaces. Availability can depend on plan, geography, administrator policy, app access, and product rollout status. Do not promise a feature to a department, client, or study team until the intended users can see and run a compliant non-PHI test in the actual workspace they will use.

OpenAI also describes ChatGPT for Healthcare and HIPAA-enabled Enterprise workspaces as supporting approved healthcare configurations, including separate Epic access where applicable. The Epic plugin is not described as available to individual ChatGPT for Clinicians accounts. Keep user training explicit: public-data access, healthcare workspace controls, and EHR access are separate capabilities with separate eligibility, setup, and compliance requirements.

Provider-side limits also matter. Public sources may have incomplete records, update delays, terminology inconsistencies, deprecations, or source-specific search behavior. PubMed abstracts do not replace full-text review; DailyMed labeling should be checked against the current label record; ClinicalTrials.gov entries should be verified for current status and eligibility details; CMS coverage materials should be read in context with dates and jurisdictional details where applicable. The plugin can accelerate retrieval, but it does not remove the need to inspect the underlying source.

Medical-judgment and operational-use disclaimer

This workflow is educational and operational guidance for using public healthcare sources with ChatGPT and Codex. It is not medical advice, legal advice, a diagnosis system, a treatment recommendation engine, or a substitute for institutional review, clinician judgment, pharmacist review, research-coordinator verification, privacy-office approval, or counsel. Any output that affects care, coverage, research enrollment, prescribing, patient communication, or operational policy must be checked against the original source records and the organization’s approved procedures.

Use explicit reviewer sign-off for high-impact deliverables. A medication-safety summary should identify who verified the current label and terminology source. A trial landscape should identify who confirmed trial status and eligibility criteria. A Medicare coverage analysis should identify who checked the relevant CMS record and date. A Codex-generated report or pull request should identify who reviewed the evidence ledger, tests, and assumptions before publication or deployment.

Closeout checklist before scaling the workflow

1. Confirm eligibility: verify that the intended users, plan, workspace, and geography support Healthcare Public Data under OpenAI’s current documentation.
2. Approve policy: document that Healthcare Public Data is for public-source research only and must not receive PHI or patient chart excerpts.
3. Separate integrations: train users that Epic, where approved, is a separate read-only plugin with its own BAA, configuration, sign-in, and chart-permission requirements.
4. Test safely: run known non-PHI queries against selected public sources before allowing real research workflows.
5. Log and review: use audit logs, available Compliance API or export tooling, and evidence ledgers to monitor use and reconstruct decisions.
6. Verify outputs: require source-level review before any answer informs care, coverage, enrollment, product strategy, publication, or code deployment.

The practical value of Healthcare Public Data in ChatGPT and Codex is speed with traceability: faster retrieval of public records, clearer comparison across sources, and reproducible evidence artifacts for reviewers. The safety condition is equally important: keep patient data out of public-source prompts, understand where data is sent, respect BAA boundaries, and make humans responsible for clinical, compliance, and operational decisions.

Useful Links

• OpenAI release notes for product availability and September 2026 healthcare plugin updates
• OpenAI announcement on connecting EHR context and healthcare public sources to ChatGPT
• OpenAI Help: Using Healthcare Public Data in ChatGPT and Codex
• OpenAI Help: Using the Epic plugin with ChatGPT and Codex
• OpenAI Help: ChatGPT for Healthcare
• ChatGPT release notes for temporary chat, personalization, and product behavior changes