Why hallucinated citations are a serious problem in regulatory affairs AI

A citation is not decoration in RA

In most writing a citation supports an argument. In regulatory affairs it is the argument. A classification note, a justification in a cover letter or an internal impact assessment is judged on whether the cited provision says what the author claims, in the version that applies, with the legal weight implied.

That is why a fabricated or misapplied citation is more dangerous here than an obviously wrong answer. An error without a source invites checking. An error with a confident, correctly formatted reference to a real-sounding guideline entry tends to pass straight into a draft.

Five ways a citation goes wrong

'Hallucinated citation' usually means an invented source, but that is the least subtle failure. In regulatory text the more common problems involve real documents.

• Invented: the document, article or guideline entry does not exist.
• Misplaced: the document is real but the article, section or code cited is not where the claim is.
• Superseded: the passage existed, in a version that no longer applies.
• Mis-weighted: a Q&A answer or draft is presented as if it were legislation.
• Unsupported: the source is real and current, but does not say what the sentence claims.

A live example: the 2026 variation codes

The EU variations framework shows how easily this happens. The Commission's 2013 variations guidelines were replaced by C/2025/5045, which applies from 15 January 2026, and the classification annex was restructured into chapters E, Q, C and M. The finished-product batch release entry that RA teams knew as B.II.b.2 is now Q.II.b.2.

A language model trained mostly on text from before that change has seen 'B.II.b.2' thousands of times and 'Q.II.b.2' rarely. Asked about a batch release site today, it can produce an answer that names the current guideline correctly and quotes the old code beside it. Every part looks right. The combination is wrong, and only a reader who opens the current annex will notice.

The same pattern repeats wherever guidance is revised: eSubmission validation criteria, QRD templates, CMDh best-practice guides and procedural Q&As all change on their own schedules.

Why language models do this

A language model generates the most plausible continuation of text. Citation formats are highly regular — a regulation number, an article, a paragraph — so producing a well-formed reference is easy for the model even when no retrieved passage supports it. Nothing in plain generation checks that the reference points anywhere.

More training data or a larger model does not remove the problem, because the failure is structural: the model's knowledge has no version dates and no notion of legal status. The controls have to sit around the model.

Controls that actually reduce the risk

The reliable approach is to make citations a product of retrieval rather than of generation. The system searches a defined set of documents first, the answer may only cite passages that were retrieved, and every citation resolves to the source document and location the reader can open.

Version control belongs in the source layer, not the prompt. Each document should carry its authority, legal status and whether it is current or superseded, and superseded material should be kept out of normal retrieval unless the question is explicitly historical.

The system also needs permission to fail. When the retrieved evidence does not support an answer, the output should say so. A tool that always answers will eventually cite something to fill the gap.

Finally, fabrication has to be measured, not assumed away. An evaluation set with known expected sources can count invented citations on every release, alongside whether the right sources were retrieved at all.

• Citations restricted to retrieved passages from a defined corpus.
• Version and legal status stored as source metadata.
• Superseded documents excluded from normal retrieval.
• A not-found result instead of an unsupported answer.
• Fabricated citations counted in every evaluation run.

What a reviewer should still check

Controls reduce invented references. They do not remove the fifth failure: a real, current source that does not support the sentence attached to it. That check stays with the reviewer, and it is quick if the output makes it easy.

• Open the cited source rather than trusting the label.
• Confirm the version and date against the authority's current page.
• Find the passage and read the conditions around it, not only the matched sentence.
• Check the legal weight: legislation, guideline, Q&A or draft.
• Ask whether the passage supports the specific claim, for this product and route.

How RafiHive handles it

RafiHive answers from a reviewed knowledge base of EU/EEA official sources with version and status metadata, cites the sources it used, and returns a not-found result when support is missing. Its golden-set evaluation counts fabricated citations on every run and publishes the result, together with its limits: the set is small, and not all cases yet carry named RA sign-off. A golden set is a regression guard, not proof that a particular answer is right, which is why every output still requires qualified review.