An AI assistant can sound remarkably settled when the case itself does not feel settled at all.
Its answer may be fluent, logically organised and supported by citations, yet still conflict with the examination, chronology or details that do not fit your working picture.
In that situation, the useful question is not simply:
Who is right?
It is:
What is this assessment hanging on, and what would change it?
Disagreement does not automatically mean the system is wrong. Agreement does not make it right either.
The more useful approach is to turn the disagreement into something clinically inspectable.
Disagreement is a request for inspectable reasoning
“Check again” is rarely the strongest follow-up.
An assistant may simply restate the same conclusion with slightly different wording, or sound more confident without resolving the uncertainty.
More useful questions are:
- Which data from this patient are carrying the assessment?
- What argues against the leading hypothesis?
- Which finding may be receiving too much weight?
- What is easiest to miss in this case?
- What new information would materially change the assessment?
These questions do not ask the system to “admit that it was wrong.”
They ask it to separate patient-specific evidence, general disease knowledge, assumptions and uncertainty.
Challenge in Terav
In Terav, this second pass is built directly into the conversation as Challenge.
Instead of simply asking the assistant to reconsider the case, the clinician can request a structured critique of the existing assessment.
Challenge examines four areas:
- Against the lead — what weakens the leading hypothesis;
- Overweighted finding — what may have been given disproportionate importance;
- Easiest to miss — what deserves another look;
- What would change this — what new information could materially alter the interpretation.
Example of the Challenge interface in Terav. Challenge provides a structured critique of an existing assessment. The screenshot demonstrates the interaction format; it is not evidence of the clinical accuracy of the specific output shown.
Challenge does not turn an AI answer into a verified diagnosis.
It also cannot guarantee that a system will identify its own mistake.
Its purpose is narrower: to make counter-evidence, alternatives and conditions for revising the conclusion easier to inspect.
What to inspect in the answer
Is the conclusion actually grounded in this patient?
An assistant may correctly describe a typical disease pattern while quietly replacing an individual case with the general pattern.
One particularly important failure mode is treating information that was not provided as though the finding were absent.
A useful follow-up is:
Which parts of your conclusion are based directly on the patient data provided, and which depend on general disease patterns or assumptions?
Are alternatives represented fairly?
A long differential list can look comprehensive while offering very little decision support.
What matters is:
- which findings support each alternative;
- what argues against it;
- what observations would distinguish competing hypotheses.
You might ask:
Give the strongest case for the nearest competing hypothesis. What findings would make it more likely?
Does the evidence actually apply?
A real citation may still concern another species, population, disease stage or clinical setting.
The existence of a source does not establish that the source applies to the patient in front of you.
When a paper matters to the decision, check:
- species;
- study population;
- study design;
- sample size;
- outcomes;
- whether the study actually addresses the clinical distinction you are trying to make.
What would change the assessment?
A working hypothesis should be revisable.
Asking:
What finding would make you change your current assessment?
helps expose whether the conclusion has meaningful conditions under which it would be reconsidered.
If no plausible new information can change it, the response is behaving more like a fixed assertion than a clinical hypothesis.
An example: instead of “Are you sure?”
Imagine a dog with lethargy, reduced appetite, intermittent vomiting, mild azotaemia and urine that appears insufficiently concentrated for the clinical context.
An AI assistant places intrinsic renal injury at the top of the differential list.
Instead of asking:
“Are you sure?”
try:
“Which findings in this case are carrying the leading hypothesis? What argues against it? Which findings could also fit pre-renal or post-renal causes, and what new information would most change the current ranking?”
That prompt does not tell the assistant which diagnosis to choose.
It asks it to make the distinctions between hypotheses explicit enough for the clinician to examine.
Automation bias: what the evidence does and does not show
Human-factors and healthcare research has documented automation bias: people can become overly reliant on automated recommendations or fail to notice information that contradicts them.
Systematic reviews have identified this phenomenon in decision-support settings, including clinical tasks. Most of that evidence comes from human healthcare and other professional domains rather than veterinary clinical practice. [1,2]
We have previously looked in more detail at what is and is not known about AI and veterinary clinical reasoning.
Short-term agreement with an incorrect automated recommendation is not the same question as long-term loss of clinical skill.
There is currently no longitudinal veterinary evidence showing that sustained use of AI assistants causes deterioration of veterinarians’ clinical reasoning skills.
The existing literature supports taking over-reliance on automated advice seriously. It does not establish that AI inevitably causes clinical deskilling. [1,2]
So it is better not to claim that one interaction style has been proven to “preserve clinical reasoning,” just as it would be too strong to claim that AI inevitably erodes it.
A more immediate question is whether a clinical assessment remains open to contradiction, alternatives and new evidence.
Try it with a familiar case
Start with a familiar, de-identified case.
Before asking an assistant, write down:
- your own working hypothesis;
- the findings supporting it;
- the parts of the case that make you hesitate.
After the AI response, do not stop at “agree” or “disagree.”
Use Challenge, or ask the same questions manually:
- What argues against the lead?
- What may have been overweighted?
- What is easiest to miss?
- What would change the assessment?
Then compare the two versions.
Note:
- what supported your original view;
- what made you refine it;
- what remains unresolved.
The goal is not to win an argument with the tool, or to let the tool win one with you.
It is to make the clinical assessment visible, testable and revisable when stronger evidence appears.
Sources and further reading
- Goddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association. 2012;19(1):121–127.
https://pubmed.ncbi.nlm.nih.gov/21685142/
DOI: https://doi.org/10.1136/amiajnl-2011-000089 - Lyell D, Coiera E. Automation bias and verification complexity: a systematic review. Journal of the American Medical Informatics Association. 2017;24(2):423–431.
https://pubmed.ncbi.nlm.nih.gov/27516495/
DOI: https://doi.org/10.1093/jamia/ocw105 - Amann J, Blasimme A, Vayena E, Frey D, Madai VI. Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC Medical Informatics and Decision Making. 2020;20:310.
https://doi.org/10.1186/s12911-020-01332-6 - World Health Organization. Ethics and governance of artificial intelligence for health. 2021.
https://www.who.int/publications/i/item/9789240029200