A sentence nobody said, in a record nobody can check.

What went wrong with transcription

Researchers from Cornell, the University of Washington and other institutions examined a widely deployed speech-to-text model and found invented content in roughly one in seventy transcriptions. Not garbled words: whole sentences that were never spoken, inserted most often during silences, and disproportionately in recordings of people with aphasia — the patients least able to correct the record.1

The compounding detail is structural. At least one commercial product built on that model deleted the source audio after transcription.2 Once the audio is gone, an invented sentence in a clinical note is indistinguishable from a real one, and will be read years later as something the patient said.

The rules already assume you can trace it

The rules governing electronic records in this sector — 21 CFR Part 11 in the US, Annex 11 in the EU — already require that any record can be traced back to who produced it and what it rested on.3 The sector’s own shorthand for those properties is ALCOA+: attributable, legible, contemporaneous, original and accurate, then complete, consistent, enduring and available. Applied to an AI system, that trail has to reach the sources behind the output.

Drug safety reporting makes the demand sharper still. A safety narrative is a written account of what a case report said, and an inspector reading it will want to reach the underlying report behind each statement. A narrative that fluently summarizes a case it partly invented is worse than none, because it is believable.

Where it bites in practice

• Clinical documentation. A note is a legal record, read by clinicians who were not in the room and by lawyers years later.
• Prior authorization and coverage. A determination has to rest on the payer's actual policy wording for the plan year in force.
• Guideline lookup. The relevant answer is the recommendation as written, with its grade and edition.
• Medical information. Answers must stay inside the approved label and the referenced literature.
• Safety narratives. Every statement traces to a source document, and inspectors follow the trail.

Where CiteOnly fits

Where the answer already exists in a guideline, a label, a policy or a case file, CiteOnly reports it verbatim, with its location. A sentence nobody wrote cannot enter the record, so there is nothing to detect afterwards. It does not write narratives; the limits are published.