Concern about becoming accustomed to ready-made answers is understandable. A confident explanation can feel compelling long before it has been shown to fit the patient in front of us.
But the question “will AI make clinicians worse at thinking?” actually combines two different concerns.
The first is short-term over-reliance: accepting an incorrect automated suggestion during a particular decision.
The second is long-term skill retention: whether repeatedly delegating parts of reasoning to a tool changes a clinician’s ability to perform those tasks independently over months or years.
There is research on the first problem in human factors and human healthcare. Evidence for the second is much more limited. And there is currently no veterinary longitudinal evidence showing that AI assistants cause deterioration of veterinarians’ clinical reasoning skills.
Keeping those claims separate matters.
Short-term automation bias
Automation bias describes a tendency to rely too heavily on automated advice.
It can appear in different ways. A user may accept an incorrect recommendation, or fail to notice a problem because the system did not flag it.
A systematic review by Goddard, Roudsari, and Wyatt found evidence of automation bias across several research domains, including clinical decision-support systems. Importantly, automated systems often improved overall task performance in the studies reviewed. The risk appeared when users continued to rely on the system when its advice was wrong.
Lyell and Coiera later reviewed automation bias in healthcare and human-factors research. Their review highlights that the effect is not simply a matter of “people trusting computers too much.” Task complexity, working conditions, and the effort required to verify a recommendation can all influence how automated advice is used.
These findings are relevant to clinical technology.
They are not, however, veterinary evidence. They do not establish how often the same behaviour occurs with veterinary AI assistants, under which conditions, or with what clinical consequences.
Long-term skill atrophy is a different question
A separate concern is whether a clinician who routinely receives differentials, evidence summaries, or suggested next questions will gradually become less capable of producing them independently.
That is a hypothesis about long-term skill retention, not about agreement with one incorrect recommendation.
Testing it properly would require longitudinal research: a defined tool, repeated use, meaningful measures of independent clinical reasoning, an appropriate comparison group, and follow-up long enough to distinguish convenience from actual skill change.
Evidence of short-term automation bias does not meet that standard.
The broader automation literature raises related concerns about over-reliance, reduced monitoring, and difficulty detecting system errors when automation becomes highly trusted. These ideas are useful as conceptual warnings, but they do not prove that veterinary AI causes diagnostic deskilling.
At present, that question remains open.
Why evidence does not transfer automatically to veterinary medicine
Most direct research on automation bias comes from human healthcare and other high-stakes technical settings.
That literature can help veterinary teams ask better questions about how people interact with automated systems. It cannot settle the veterinary question by itself.
Veterinary practice differs in important ways:
- patients span species with different physiology and disease prevalence;
- histories and follow-up may be incomplete;
- evidence can be sparse for uncommon diseases, breeds, or species;
- owner goals, finances, welfare considerations, and diagnostic feasibility can substantially affect decisions;
- workflows differ between general practice, emergency medicine, referral hospitals, pathology, production medicine, and equine practice.
So the most defensible statement today is narrow:
Humans can over-rely on automated recommendations. We do not yet know what sustained use of AI assistants does to independent veterinary clinical reasoning over time.
Neither reassurance nor alarm should fill that evidence gap.
Why a plausible answer is easy to accept
A coherent answer is especially attractive when the case is complicated, information is fragmented, and time is limited.
But fluency does not tell us whether an inference fits the individual patient.
Particular caution is warranted when:
- the assistant has incomplete or ambiguous context;
- a conclusion sounds confident without showing which data drive it;
- the recommendation confirms the clinician’s first hypothesis;
- the cited evidence looks relevant by title but differs in species, population, or clinical task.
This is not an argument against using AI.
It is an argument against treating “tell me what this patient has” as both the beginning and the end of the interaction.
Use AI to challenge a working hypothesis
One possible way to keep the clinician’s own reasoning visible is to formulate a working hypothesis before asking for an AI response.
For example:
“My leading hypothesis is X. Findings A and B support it, but C does not fit.”
The tool can then be used to challenge that structure rather than simply generate a conclusion.
Questions might include:
- Which findings support this hypothesis, and which argue against it?
- Which alternative differentials remain plausible?
- What missing information would most change their ranking?
- What evidence supports this claim, and how closely does the study population match this patient?
- What is the strongest argument against my current leading diagnosis?
This approach has not been shown to prevent loss of clinical skills. It is simply one way to structure the interaction so that the clinician’s own hypothesis remains available for comparison and challenge.
That distinction matters. We should not replace one unsupported claim — “AI causes deskilling” — with another — “the right way of prompting prevents it.”
Both require evidence.
Test tools on cases you can judge
When evaluating an AI assistant, familiar de-identified cases can be useful.
Cases may be especially informative when:
- the initial impression later changed;
- two differentials were genuinely plausible;
- a laboratory result appeared straightforward but proved context-dependent;
- a small detail in the history changed the interpretation.
The goal is not only to see whether the system reaches the same final label.
Look at where it identifies contradictions, where it misses them, how it handles uncertainty, whether its alternative explanations are relevant, and whether the evidence it cites actually fits the patient.
This kind of exercise cannot establish that a tool will preserve clinical skills over time.
It can show how that tool behaves on material whose clinical logic the veterinarian is already able to evaluate independently.
Where Terav fits
Terav is designed as an AI assistant for veterinary professionals working with clinical cases, laboratory results, and clinical documents.
It can participate in discussion of a working hypothesis, surface alternatives, help organize available case context, and support evidence review.
Terav has not been shown to prevent automation bias, preserve clinical skills, improve clinical outcomes, or guarantee better decisions.
Clinical interpretation, verification of important information, and decisions about the patient remain the veterinarian’s responsibility.
The long-term question remains open
There is good reason to take immediate over-reliance on automated recommendations seriously. Automation bias has been studied for years in human factors and human healthcare.
That is not the same as evidence that AI inevitably erodes professional skill over time.
And it is not veterinary evidence.
For now, a more useful question than “Will AI make veterinarians stop thinking?” may be:
How can we use these tools while keeping our own clinical reasoning visible, testable, and open to contradiction — and then study what actually happens over time?
Sources and further reading
- Parasuraman R, Manzey DH. Complacency and bias in human use of automation: an attentional integration. Human Factors. 2010;52(3):381–410.
https://pubmed.ncbi.nlm.nih.gov/21077562/
DOI: 10.1177/0018720810376055 - 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/ - 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: 10.1093/jamia/ocw105 - Cabitza F, Rasoini R, Gensini GF. Unintended Consequences of Machine Learning in Medicine. JAMA. 2017;318(6):517–518.
https://pubmed.ncbi.nlm.nih.gov/28727867/
DOI: 10.1001/jama.2017.7797