Generative AI can produce an answer that reads like a polished clinical explanation even when part of it is unsupported. In healthcare, that problem is often described as an AI “hallucination.” But not every incorrect output looks the same, and one especially dangerous form deserves separate attention: fabrication—the invention of a specific patient fact, test result, guideline, quotation, or citation that does not exist.

The distinction matters because different errors require different safeguards. A vague but unsupported interpretation may be caught by clinical review. A fabricated source can appear verifiable, survive a quick scan, and become embedded in documentation or decision-making. Both can cause harm, yet fabrication creates a false evidence trail that may be unusually persuasive.

The video below provides a useful starting point for understanding why fluent AI output should not be mistaken for verified clinical knowledge. As you watch, consider not only whether an answer is correct, but also whether every material claim can be traced back to the patient record or a legitimate evidence source.

What do “hallucination” and “fabrication” mean?

There is no universally enforced boundary between these terms. Researchers, vendors, clinicians, and regulators sometimes use “hallucination” as an umbrella label for any plausible but unsupported model output. Within that umbrella, “fabrication” is a useful practical term for invented particulars presented as though they were real.

A hallucination might be an unsupported explanation of why a symptom occurred, an incorrect summary of a guideline, or an inference that is not justified by the available chart. A fabrication is more concrete: the model may create a laboratory value that was never measured, attribute a medication to the wrong patient, invent an article title, or quote a recommendation that does not appear in the named source.

Neither term means the model has human intent. A language model is not necessarily trying to deceive anyone. It generates sequences that fit learned patterns and the supplied context. The safety problem is that its unsupported content can be delivered with the same confident tone and formatting as verified information.

Why generative AI can produce unsupported clinical claims

Generative models are optimized to produce a likely and useful continuation, not to function as a guaranteed database of clinical truth. They can combine patterns from training data with information supplied in a prompt or retrieved from external sources. When the evidence is incomplete, conflicting, or outside the model’s reliable knowledge, the system may still generate a coherent response.

This risk increases when the prompt demands certainty. A user who asks for “the diagnosis,” “the exact dose,” or “three supporting studies” may unintentionally encourage the model to fill gaps rather than clearly state that information is unavailable. The result can look complete because completeness is part of the requested format.

Context limitations also matter. A long patient record may be summarized, truncated, or separated across systems. The model may receive the latest consultation note but not the corrected medication list, a radiology report but not the pathology result, or a copied history containing an old error. An answer can be internally consistent while contradicting the authoritative record.

Retrieval does not eliminate the problem. A retrieval-augmented system may search approved documents before answering, but it can retrieve the wrong passage, misunderstand the source, merge separate recommendations, or generate a statement that goes beyond what the evidence supports. Retrieval improves grounding only when provenance and claim-level verification remain visible.

Tool integration introduces another layer. If AI can query laboratory systems, schedule data, formularies, or clinical guidelines, a failed tool call or misread field may be transformed into fluent prose. Without clear error states, users may not know whether a statement came from a verified database, a model inference, or a fallback response.

The practical lesson is simple: fluency is not provenance. Clinical teams need a way to distinguish source facts, calculated results, model inferences, and unresolved uncertainty. Those categories should not be blended into one polished paragraph.

Where hallucinations appear in healthcare workflows

Hallucinations can arise during documentation, patient education, chart summarization, coding support, clinical search, referral drafting, and decision support. The harm depends on how the output is used. An incorrect sentence in a brainstorming draft differs from an incorrect statement copied into a permanent health record or used to support treatment.

Clinical summarization is especially sensitive. If a model overstates symptom duration, changes laterality, confuses family history with patient history, or turns a “rule out” diagnosis into a confirmed condition, downstream users may treat the summary as authoritative. Repetition can make the error progressively harder to detect.

Patient-facing systems add urgency. A chatbot might provide false reassurance, misstate a medication interaction, or omit a reason to seek immediate care. Even when a disclaimer is present, users may act on a confident answer. High-risk symptoms require clear escalation pathways rather than an unrestricted generative response.

Fabrication: when false information looks traceable

Fabrication is dangerous because precision can create credibility. A model may generate a realistic author list, journal name, publication year, drug dose, laboratory value, or quotation. The item may follow a familiar pattern so closely that it survives a quick review.

Invented citations are a well-known example. A fabricated paper can contain a convincing title and plausible bibliographic details, while a real citation may be paired with a claim the paper never made. Checking only that a paper exists is therefore insufficient; the reviewer must confirm that the source actually supports the statement.

Patient-specific fabrication can be even more consequential. If a system inserts an allergy, prior procedure, diagnostic result, or medication that is absent from the chart, the false detail can alter care. Once copied into a note, it may be repeated in future encounters and acquire the appearance of corroboration.

Fabricated quotations and policy statements can also affect administration. An AI-generated summary might attribute a requirement to Health Canada, a professional college, or an institutional policy when the source uses different language. Regulatory and compliance claims should always be checked against the current official document.

It is important not to describe fabrication as intentional lying by the model. The model does not need deceptive intent for the output to be hazardous. What matters is the combination of specificity, unsupported content, and a presentation that invites reliance.

The strongest response is traceability. Every patient fact should be linked to the authoritative record, every calculation should expose its inputs, and every cited claim should link to a source that a reviewer can open. If provenance is unavailable, the information should be labelled as unverified rather than silently incorporated.

Why clinicians can miss these errors

Healthcare professionals work under time pressure. A clear summary can reduce cognitive load, but it can also encourage automation bias—the tendency to accept a system’s recommendation or fail to act when the system remains silent. The more polished and specific an answer appears, the more effort it takes to challenge it.

Domain expertise helps but does not eliminate risk. Specialists may detect obvious medical errors while overlooking a plausible citation outside their subfield. Junior staff may hesitate to contradict a system presented as advanced. Patients may assume the output has been medically reviewed because it uses clinical terminology.

Verification fatigue is another concern. If most AI-generated text is accurate, reviewers can become less attentive over time. A single fabricated detail may hide inside several paragraphs of correct content. Safety cannot depend only on telling users to “check everything” without giving them efficient tools to do so.

Designing verification into the workflow

Safer systems separate extraction from generation. Information copied directly from a verified record should be visually distinguishable from model-generated interpretation. Users should be able to open the source location instead of searching an entire chart.

High-impact fields should be constrained. Allergies, medication doses, diagnostic results, and patient identifiers should come from validated structured data whenever possible. The model may help format or explain those facts, but it should not invent missing values to complete a template.

Citations require claim-level grounding. A system should return stable links or identifiers and quote only the relevant supporting passage. The application can automatically check whether a source exists, but a qualified reviewer must still determine whether it is current, applicable, and accurately represented.

Uncertainty should change behaviour. When evidence is missing or conflicting, the system should abstain, ask for clarification, or escalate rather than produce a definitive answer. An honest “insufficient information” is safer than a complete but unsupported clinical narrative.

Human review must occur before the output influences care or becomes part of the permanent record. The reviewer needs authority to edit or reject the result, enough time to inspect critical claims, and a clear understanding of what the tool can and cannot do.

Organizations also need feedback and monitoring. Overrides, corrected fabrications, near misses, subgroup performance, and user complaints should be captured as safety signals. A model that performs well overall may still fail systematically in a particular specialty, population, language, or document type.

A practical verification checklist

What healthcare organizations should require

Governance should begin with the intended use. A low-risk drafting assistant and a tool influencing diagnosis require different evidence, controls, and monitoring. Organizations should evaluate the entire human–AI workflow rather than relying only on model benchmarks.

Procurement reviews should ask how the system grounds answers, handles missing data, displays uncertainty, protects personal health information, logs source access, and responds when integrations fail. Validation should use representative local scenarios, including deliberately incomplete, contradictory, uncommon, and high-risk cases.

Training should include realistic examples of hallucinations and fabrications. Users need to know which facts demand source verification, how to report an error, and when to stop using the output. Policies should prohibit entering identifiable health information into unapproved consumer AI services.

Finally, organizations should measure more than accuracy. Useful indicators include correction frequency, fabricated citation rate, unsupported patient facts, clinician overrides, time required for verification, near misses, and whether performance differs across patient groups or clinical settings.

Conclusion: trustworthy AI must make verification easier

AI hallucinations and fabrications are closely related, but distinguishing them improves safety. Hallucination describes the wider problem of unsupported generated content. Fabrication highlights a particularly concrete failure: invented evidence presented in a form that appears real and traceable.

The answer is not to assume that every AI output is wrong, nor to accept fluent output as reliable. The safer position is calibrated trust. Use AI where it adds value, constrain it where errors could cause harm, and verify every material clinical claim against an authoritative source.

In healthcare, a trustworthy system should not merely generate an answer. It should show where the information came from, distinguish fact from inference, communicate uncertainty, and support accountable human review. When the evidence is missing, the system’s most responsible response may be to say so.

Further reading