Treating OpenAI's nine-source Healthcare Public Data plugin as nine versions of the same thing, medical evidence, generally, is the single most common misreading available of this announcement. Each source answers a genuinely different question, and OpenAI's own implementation guidance explicitly warns against several specific misuses worth naming directly rather than glossing over.
The nine sources, and the question each actually answers
PubMed: what biomedical research has been indexed on this topic, a bibliographic and abstract database rather than a curated evidence-quality filter. The common misuse: treating every citation surfaced as conclusive evidence, when indexing says nothing about a study's design quality, sample size or applicability to a specific patient.
ClinicalTrials.gov: which trials are registered or currently recruiting, a registry rather than an efficacy database. The common misuse: assuming a listing itself proves efficacy or confirms current, actual eligibility, when registry status reflects only that a trial exists and was registered, not that it has produced results or that a specific patient genuinely qualifies.
DailyMed: what the current US product label actually says, official regulated labelling information. The common misuse: treating the label as the complete treatment pathway, when a label describes an authorised product's approved use, not where that product sits within a broader clinical guideline or treatment sequence.
RxNorm: which standard medicine concept or identifier applies, a terminology-normalisation resource connecting ingredients, strengths, dose forms and branded products for interoperability purposes. The common misuse: inferring appropriate dose or interchangeability from a terminology match, an inference OpenAI's own guidance explicitly warns against, since RxNorm standardises naming, it does not itself determine clinical appropriateness.
openFDA: whether public recalls, safety records or adverse-event reports exist, a regulatory and pharmacovigilance dataset. The common misuse: treating adverse-event reports as proof of causation or as a reliable measure of incidence, when spontaneous reporting systems capture association without establishing that a medicine caused a specific reported event or how common that event genuinely is.
CMS Coverage: what national or local Medicare coverage policy exists, a US payer-policy resource. The common misuse: assuming a coverage policy establishes one specific patient's actual benefit, when coverage determinations involve additional patient-specific and plan-specific factors a general policy lookup does not capture.
CMS Open Data: what selected Medicare payment, prescribing or utilisation data exists, an operational and claims-adjacent dataset. The common misuse: treating claims data as a complete clinical dataset, when billing and utilisation records capture what was claimed and paid, not the full clinical picture a chart review would provide.
Medicare Care Compare: what public quality information is reported about facilities, a comparative quality-reporting tool. The common misuse: comparing incompatible measures or time periods across facilities without accounting for how those specific metrics were defined and collected.
And the NPI Registry: which provider or organisation is associated with a specific identifier, an identification and lookup resource. The common misuse: assuming an NPI entry proves current licensure or quality, when registry presence confirms identification, not an active, unrestricted, currently verified professional status.
Why this matters beyond academic precision
A tool that can search all nine sources fluently risks presenting genuinely different categories of information with a uniform confidence and tone, precisely the failure mode that makes distinguishing a bibliographic index from a regulatory dataset from a payer-policy resource matter practically rather than only academically. A clinician or administrator reading a ChatGPT-generated answer that draws on several of these sources at once should be asking, for each specific claim, which source supported it and whether that source is actually built to answer the question being asked, exactly the discipline this cluster applies to every synthesis layer across every clinical AI category it covers.
The UK parallel worth holding in mind
None of these nine sources maps directly onto UK practice, and the equivalent UK stack, NICE, CKS, SIGN, MHRA safety communications, UK SmPCs on emc, the NHS Specialist Pharmacy Service, local formularies, answers a structurally similar but substantively different set of questions, the parallel this cluster's dedicated UK-connector analysis works through source by source.
Frequently asked questions
Are all nine sources equally authoritative?
Each is authoritative for the specific question it is built to answer and not for questions outside that scope, which is the whole point this article makes: authority is source-and-question specific, not a general property any of these nine sources carries uniformly across every possible use.
Does OpenAI itself warn against these specific misuses?
Yes: OpenAI's own implementation guidance explicitly cautions against inferring dose or interchangeability from RxNorm and against treating openFDA adverse-event reports as causal incidence data, warnings worth taking as seriously as the connector capability itself.
How should a clinician verify which source actually supports a specific claim in a ChatGPT-generated answer?
By checking the citation or source attribution directly against the specific database named, and asking whether that database's actual scope, bibliographic index, product label, registry, claims dataset, genuinely supports the claim being made, rather than accepting a confident-sounding synthesis at face value.