Cited or abstain: why a veterinary AI assistant should never guess
General-purpose chatbots will confidently invent a dose. A clinical assistant should do the opposite - quote a source, or say it does not know. Here is why that principle is the whole point.
A clinical AI assistant should follow one rule above all others: answer with a citation, or do not answer. If it cannot ground a claim in a source, it should say so plainly rather than fill the gap with a confident guess. We call this cited or abstain, and it is the difference between a research aid and a liability.
The problem with a confident guess
Large language models are built to produce fluent, plausible text. That is exactly what makes them dangerous in a clinical setting: a fabricated drug dose reads just as smoothly as a correct one. A general chatbot asked for a feline analgesic dose will happily produce a number. It has no obligation to be right, and no way to tell you when it is uncertain.
In veterinary medicine that is not a minor flaw. A wrong dose, an invented contraindication or a made-up study can change what happens to an animal. Fluency is not evidence.
What "cited" actually means
Cited is stronger than "sounds authoritative." A grounded answer:
- Points to the specific source the claim came from, so you can open it and check.
- Quotes doses and figures from that source rather than calculating or estimating them.
- Separates what the literature says from what is general background, and labels the difference.
The citation is not decoration. It is the mechanism that lets a clinician verify in seconds instead of taking the tool's word for it.
What "abstain" protects
Abstaining is the half people forget. An assistant that always answers has quietly decided that a plausible guess is better than an honest "I do not have a grounded source for that." For clinical work the opposite is true.
A cited-or-abstain assistant will:
- Tell you when the available sources do not cover your question, rather than inventing coverage.
- Fall back in clear steps - grounded sources first, then a labelled wider search, and only then clearly-marked general knowledge - so you always know which footing you are on.
- Make it obvious when you have left evidence behind, so the responsibility for the next step sits with you, informed.
Why this is a design choice, not a disclaimer
Plenty of tools bolt a "may be inaccurate" warning onto an assistant that still guesses freely. That is not the same thing. Cited or abstain has to be built into how the answer is produced: retrieve real sources, ground every claim, and refuse to manufacture the ones that are missing. A disclaimer shifts the risk to you; a cited-or-abstain design removes it at the source.
The clinician stays in charge
None of this replaces judgement. The point of grounding and abstaining is to give a clinician a faster, safer way to reach the evidence and decide - with the sources in front of them and no hidden fabrication in the answer. The assistant does the reading; the vet makes the call.
If you want to see cited-or-abstain answers on real clinical questions, book a demo.