In short
- A clinical AI assistant with citations answers only from documents you supply, such as notes, results and guidelines, and links each statement to the passage that supports it so the clinician can verify it.
- Claude's citations feature returns pointers that are guaranteed to reference the provided documents, but a valid pointer is not proof that the passage supports the claim, so you still need verification steps and tests.
- Showing the basis for a recommendation is also what FDA's fourth criterion for non-device clinical decision support asks for, which makes citations a product, safety and regulatory decision at once.
A clinical AI assistant with citations answers a clinician's question from a defined set of documents, such as the patient's notes, lab results, medication list and the guidelines you choose, and attaches each statement to the exact passage it came from. The clinician reads the answer, clicks a citation and sees the source text in context. With that click a clinician can confirm a statement during a busy clinic list without rereading the chart. With Claude, the citations feature does the pointing for you, but you still have to design the documents, the screen and the checks around it.
This guide is for product and engineering leads building assistants for clinicians: chart summarizers, pre-visit briefs, question answering over records.
Why do citations matter more in clinical software?
Because a wrong answer that sounds right is the main failure mode. NIST's Generative AI Profile (NIST AI 600-1, July 2024) calls it confabulation: generative systems confidently presenting erroneous or false content, including outputs that diverge from the input or contradict earlier statements. NIST uses healthcare as its example, where a confabulated summary of patient information could lead doctors to incorrect diagnoses or the wrong treatment. It also warns that models can produce confabulated logic or citations that purport to justify an answer, which can make people trust a wrong output even more.
Regulators want clinicians to be able to check the basis too. FDA's Clinical Decision Support Software guidance, issued January 29, 2026, says decision support is excluded from the device definition only if, among other criteria, it lets the health care professional independently review the basis for its recommendations, so that they do not rely primarily on them. The guidance describes automation bias, the tendency to over-rely on suggestions from an automated system, as the reason. ONC's certification criterion for decision support interventions, 170.315(b)(11), likewise requires certified EHRs to expose source attributes for evidence-based interventions, starting with a bibliographic citation, the developer and the release and revision dates.
How do Claude's citations work?
Anthropic's documentation describes three stages. You send documents with citations enabled on each one. The API splits their contents into chunks that set the smallest unit Claude can cite. Claude then answers in text blocks, and each block can carry a list of citations pointing back to locations in the documents.
- Plain text documents are chunked into sentences, and citations return character index ranges.
- PDFs have their text extracted and chunked into sentences, and citations return page ranges. Claude cannot cite images, including images inside PDFs.
- Custom content documents use the content blocks you provide as they are, with no further chunking, and citations return content block index ranges.
- A request cannot mix documents with citations enabled and documents without them.
- The cited_text field returned with each citation does not count toward output tokens, and is not counted as input when you pass it back in later turns. Enabling citations adds a little to input tokens.
Anthropic compares this with simply prompting the model to quote its sources. Because the API parses citations and extracts the cited text itself, the citations are guaranteed to contain valid pointers to the provided documents. Anthropic also reports that, in its own evaluations, the feature is significantly more likely to cite the most relevant quotes than prompt-based approaches. The feature is documented as available on the Claude API and on Amazon Bedrock, among other platforms.
How should you structure patient records for citations?
Granularity is a design decision. If a whole visit note goes in as one plain text document, a citation will point at a sentence inside it, which is usually what a clinician wants. For structured data such as a medication list, lab panel or problem list, sentence chunking makes little sense. Custom content documents let you decide what one citable unit is.
- One plain text document per note or report, so citations land on sentences and open the right note.
- One custom content document per structured list, with one content block per medication, result or problem, so a citation points to exactly one row.
- Use each document's title and context fields for metadata such as the note date, author and source system. Anthropic's documentation says these fields are passed to the model but are not citable, which keeps metadata out of the evidence.
- Keep a mapping from document index and location back to your own record identifiers, so the interface can open the source in your viewer rather than showing a raw string.
- Pass only the records the requesting clinician is allowed to see. The citation layer cannot fix an access control mistake made before the request.
What does a citation-first answer look like on screen?
Put the citation where the claim is. Each sentence or bullet in the answer gets a small marker. Clicking it opens the source document scrolled to the cited passage, with its date and author visible, because a correct quote from a three-year-old note can still mislead. When the documents do not contain an answer, the assistant should say so plainly. Anthropic's guidance on reducing hallucinations recommends explicitly giving the model permission to say it does not know, and restricting it to the provided documents rather than its general knowledge.
Uncited statements deserve a visual difference. If a sentence comes back with no citation, show it as unsupported or drop it, so the clinician can tell which parts of the answer came from the chart.
What can go wrong even with citations?
- A valid pointer to the wrong evidence. The API guarantees the citation points at a real passage in your documents. It does not guarantee that the passage supports the sentence it is attached to.
- Missing citations. For structured extraction, Anthropic's documentation tells developers to check that each field has a value and at least one citation, because the API does not guarantee either.
- Structured outputs. Citations cannot be combined with Claude's structured outputs feature; the API returns an error if you enable both. Anthropic suggests asking for one line per field and parsing the lines in your code instead.
- Images and scans. Text in images cannot be cited, so faxed documents need text extraction before they can serve as evidence.
- Stale or conflicting records. The model can cite an old medication list accurately. Put a date on every source and tell the model to prefer the newest one.
Anthropic's hallucination guide suggests a verification pass: after drafting, have the model find a supporting quote for each claim and retract any claim it cannot support. In code, a simple rule helps too: every sentence must carry at least one citation, or it is flagged. Anthropic is explicit that these techniques reduce hallucinations but do not eliminate them, and that critical information should always be validated.
Is a cited clinical assistant a medical device?
That depends on what the assistant does. Citing sources alone does not decide it. FDA's January 2026 guidance lists four criteria that must all be met for decision support to fall outside the device definition. Among them, the software must not analyze medical images or signals, must be intended for health care professionals and must let them review the basis of its recommendations. To satisfy that last criterion, FDA recommends that the software or labeling describe the intended use, user and patient population, identify the required inputs, explain in plain language how the logic was developed and validated, and show patient-specific information including missing or unexpected input values.
FDA also says software intended for critical, time-sensitive decisions does not meet that criterion, because a clinician is unlikely to have time to review the basis. Decision support aimed at patients or caregivers instead of professionals meets the device definition. A chart summarizer that cites notes for a clinician preparing a visit sits in a very different place from a bedside alert. Get a regulatory read before you set the intended use.
How do you run a cited assistant under HIPAA?
Anthropic's privacy center says it provides a BAA covering its HIPAA-ready services, including its first-party API, but coverage is listed feature by feature. For the Messages API it names covered features such as prompt caching and structured outputs, and states that Messages API features not listed are not covered. Check whether citations is on that list for your agreement before sending PHI through it on the first-party API. On Amazon Bedrock, which is on AWS's HIPAA eligible services list, the AWS BAA applies instead.
Our AI assistants cite their sources. We built an assistant that reads a patient's records and answers with linked sources, on Claude under a BAA, inside a live telehealth platform. It runs on Claude via AWS Bedrock or Anthropic directly, with provider switching, and each clinician sees only their own patients. We sign a BAA before anyone sees PHI and run a penetration test before launch.
How do you know it is ready for clinicians?
Test it the way it will be used. Build a held-out set of real questions with clinician-written reference answers, include questions the records cannot answer, and measure two things separately: whether the answer is right, and whether each citation actually supports the sentence it is attached to. Re-run the same set whenever the prompt, the document structure or the model changes.
Onix Systems has built software since 2000 and has 370+ engineers. If you are designing a clinical assistant and want answers your clinicians can check, talk to one of our engineers. Our free five-day technical audit ends with a written report on your data flow, citation design and test plan.
Sources
- Claude Docs: Citations
- Claude Docs: Reduce hallucinations
- NIST AI 600-1: Generative Artificial Intelligence Profile (July 2024)
- FDA: Clinical Decision Support Software guidance (January 2026, PDF)
- ONC (ASTP): §170.315(b)(11) Decision support interventions
- Anthropic Privacy Center: Business Associate Agreements (BAA) for commercial customers