AI tools are becoming part of everyday veterinary practice. You may already be using AI or software that has AI built into it without you being fully aware. A 2024 survey found that more than half of veterinary professionals who said they did not use AI were, in fact, using AI-embedded tools in their daily work.¹

This guidance is designed to help you identify where AI is relevant to your clinical practice, understand the risks it creates, and apply practical steps to manage those risks in a way that protects patients, clients, and your registration.

For information about the governance responsibilities of practice owners and managers, see the companion Vet Council guidance for veterinary businesses. ²

The Vet Council’s position on AI governance is about managing unvetted use rather than specifying an approved list of tools.

Your professional responsibilities as a veterinarian apply regardless of the tools you use. The Code of Professional Conduct requires you to exercise sound professional judgment in all your professional endeavours.³ Competency Standards require you to practise within your competence and to refer where necessary.⁴ These obligations are not altered by using AI, but they can be undermined if you are not alert to its risks.

International veterinary and medical regulators, including the Royal College of Veterinary Surgeons,¹⁵ state that clinical decision-making must not be wholly delegated to AI. Accepting an AI output without applying your own independent clinical judgment is a failure of professional responsibility, regardless of whether the output proves to be correct. The question is not whether AI was right; it is whether you exercised your own judgment.

This applies equally where a practice or employer has introduced an AI tool or expects its use. Practice policies and technology choices do not override your individual professional accountability. If you are required to use a tool you do not understand, or one that you believe poses a risk to patient care, you have both the right and the obligation to raise this.

Making a reasonable decision to decline to use an AI tool in a particular clinical situation is a legitimate professional judgment, not non-compliance. If you view that a tool is not appropriate for a particular patient, case, or clinical context you are entitled to proceed without it. For example, if a patient falls outside the tool’s validated population, you lack sufficient understanding of the tool’s outputs to evaluate them, or if the situation requires a speed of response that does not allow for adequate review. The obligation is to exercise clinical judgment competently. AI is a resource available in support of that, not a mandatory step in every clinical encounter.

You may not always know when you are using AI. It is embedded in many of the tools that have become routine in veterinary practice, including:

• Practice management and EMR systems: some include AI features for appointment triage, summarising records, or clinical reminders

• Diagnostic imaging platforms: some radiology viewers include AI-generated abnormality flags or preliminary reports

• Laboratory analysers: some point-of-care haematology, urinalysis, and cytology tools use AI classification algorithms

• Clinical scribes and transcription tools: these automatically transcribe and structure consultation notes

• Client communication tools: AI-drafted discharge instructions, reminder emails, or triage chatbots.

Before relying on any output from a digital tool in your clinical decision-making, or from a remote source where you can’t see the tools employed by others, ask: is there an AI component to this, and do I understand what it is doing? And, do I need to review the content?

AI tools can be wrong. Their reliability depends heavily on the quality and representativeness of the data they were trained on. A tool developed mainly using data from referral centres in the UK or North America may not perform well when applied to your patient population in New Zealand, or to breeds and species underrepresented in its training data.⁵

One published study found sensitivity for detecting pulmonary masses as low as 55%, despite high specificity.⁶ This means that while the tool was good at identifying normal findings, it missed nearly half of confirmed abnormal cases. Relying on a negative AI finding without independent clinical assessment could lead to missed pathology.

Automation bias is the tendency to accept AI-generated outputs without adequately questioning them — particularly when you are busy, when the output looks authoritative, or when it aligns with your initial impression.

Research has found that clinicians can lose diagnostic acuity in areas where they become heavily reliant on AI assistance.⁷

Automation bias is not a character flaw. It is a predictable cognitive response to working with apparently reliable technology. To mitigate this, it is important to build habits and processes that require active clinical engagement with AI outputs rather than passive acceptance. Useful habits include:

• Always forming your own clinical impression before looking at an AI output

• Actively asking “what might be wrong or missing from this?” rather than “does this look right?”

• Being more sceptical, not less, in ambiguous cases — those are precisely where AI is most likely to be unreliable.

Many AI tools, particularly those using deep learning, cannot explain how they arrived at their output. You may receive a diagnostic recommendation or flagged finding without any insight into the reasoning behind it. If you cannot understand why an AI tool reached a conclusion, it is harder to evaluate whether it is correct, explain it to a client, or document your decision-making appropriately.⁸

When a credentialed specialist produces a report you cannot fully interrogate, you can at least verify their qualifications and know they are professionally accountable for what they have written. AI has none of these attributes. There is no qualification to verify, no professional body to which it answers, and no mechanism by which it can be held responsible if its output causes harm. This does not mean AI outputs are without value but it does mean the entire burden of evaluating those outputs rests with you.

Where a tool cannot explain its reasoning, you need to apply more of your own — not less.

Large Language Model (LLM) based tools, including clinical scribes, differential diagnosis generators, and communication drafting tools, can produce plausible but incorrect information. In a clinical context this may appear as:

• Drug names, dosages, or interactions that are wrong or do not exist

• Clinical facts that are inaccurate or specific to the wrong species

• Details in a clinical record that were not said or did not happen

You must read AI-generated clinical records, discharge instructions, and clinical summaries critically before accepting them into the permanent record or sending them to a client.

Using an AI tool does not expand your scope of practice. AI does not substitute competency if you are unable to interpret a finding without AI assistance. A general practitioner without training in advanced diagnostic imaging should not use an AI imaging tool as a substitute for specialist referral.⁹ If you cannot independently evaluate whether a flag is clinically significant, your obligation is to refer.

The risk profile of AI use is not the same for every veterinarian. The ability to critically evaluate AI outputs depends on having enough clinical experience to recognise when something is wrong, unusual, or inconsistent with the patient in front of you.

For recently graduated veterinarians, this presents a particular and significant risk. A new graduate who has not yet developed broad clinical acumen may lack the experiential foundation needed to identify errors, biases, or implausible outputs from an AI tool. Complex or unfamiliar cases where a junior clinician might be most tempted to rely on AI are precisely the situations in which AI is most likely to be unreliable, and where errors are most serious.

This applies equally to veterinary students. AI tools are not a substitute for developing knowledge through primary sources. Using AI in place of textbooks, peer-reviewed journals, and structured clinical learning undermines the development of the foundational understanding that competent clinical judgment depends upon. Students should treat AI as a supplement to appropriately published clinical and scientific literature.

More experienced veterinarians carry their own risks: long-established habits and high confidence can also contribute to automation bias, particularly when an AI output appears to confirm an existing impression. All veterinarians, regardless of experience level, benefit from applying structured habits around AI use.

Practices should be particularly attentive to the AI governance support they provide for junior staff, including:

• Clear guidance on when AI outputs should prompt escalation to a senior clinician

• Explicit expectations that AI tools are not a substitute for supervision or mentorship

• A workplace culture that makes it safe to raise uncertainty rather than deferring to an AI output.

One of the most significant, and least visible, risks of AI use in clinical practice arises not from a tool’s technical limitations but from an unconscious assumption on the part of the veterinarian. When you use AI to help work up a clinical case, it is natural to feel that AI is reasoning about your particular patient. It is not. AI knows only what you have typed. Nothing more.

This matters because the gap between what you know about a case and what you have told the AI is almost always substantial. In a physical consultation, you accumulate information continuously through observation, examination, the animal’s behaviour, your own prior knowledge of similar cases, and the owner’s account. Only a fraction of this typically makes it into a prompt. AI’s differential list, treatment plan, or research summary looks complete and considered but it was built from that fraction, not from the full clinical picture.

A useful way to think about this: it is like leaving a voicemail for an unknown consultant — describing the case in the words you chose at that moment, with no opportunity for the consultant to ask a follow-up question, probe an inconsistency, or notice something you did not think to mention. Unlike a colleague or specialist who could request additional information or query something that does not add up, AI will work only with what was provided and respond as though it has everything it needs. Its reply will not flag what was missing because it has no way of knowing.

This example illustrates a general principle that extends beyond any single case type: AI outputs are bounded by the information they receive, but they look complete and confident regardless of how bounded they are.

The following specific risks follow from this:

Examination findings, the animal’s demeanour and presentation, environmental history, the owner’s manner of describing the problem, and the things noticed but not yet articulated as significant all shape clinical reasoning but rarely make it fully into an AI prompt. AI will generate a response that appears to account for the full picture without having it.

When you research a case independently, you make active judgments about which sources to consult, how to evaluate study design and population, whether the findings translate across species, and what the limits of the relevant literature are. When AI does the research, none of these judgments are visible. You see the conclusions, but not what was searched, what was excluded, how conflicting evidence was weighted, or where the gaps in the literature are. AI’s research process is invisible and therefore cannot be evaluated.

AI tools frequently produce references that appear authoritative. These citations may not exist, or may not say what the AI claims. The critical step is not just confirming that a paper can be found but reading it and confirming that it says what you have been told, that the study population is relevant to your case, and that the conclusions support the clinical recommendation being made. An AI-generated literature summary should never be accepted without verifying the primary sources.

How you describe a case to an AI platform shapes the response you receive. If your description already reflects a working hypothesis, the AI’s response is likely to be consistent with that hypothesis — not because it has independently evaluated the evidence, but because it is responding to the framing you provided. This can feel like confirmation from an independent source when it is not. The differential list may reflect your thinking back to you, rather than challenging it.

There are categories of information that AI is unlikely to have, even if your prompt is thorough. This includes:

• New Zealand-specific context: AI does not know you are practising in New Zealand unless you tell it, and even then, may not adjust its recommendations accordingly. Things like drug availability, relevant legislation, common breeds, farming practices, and New Zealand-specific disease prevalence patterns may not be adequately represented in tools trained predominantly on international datasets.

• Recency: AI training data has a cutoff date, and may not flag when a recommendation reflects evidence or guidelines that have since been superseded. If a clinical area is rapidly evolving, treat AI outputs as potentially outdated and verify against current sources.

• Client and practice context: financial constraints, the owner’s history with the practice, previous treatment attempts, or the practical limitations of the setting may be the most important factors shaping what is possible in a case. The AI does not have this information. A technically correct treatment plan may be wholly impractical without it.

• Evidential confidence: The confidence of the language AI uses does not reflect the strength of the evidence. Where a recommendation matters clinically, you should assess the quality of the evidence independently rather than relying on how certain the AI sounds.

AI tools are trained on historical data, and that data may not adequately represent the patients and cases you see in your practice. This is not simply a procurement question to be settled at the point of evaluating a new tool — it is an active concern at the point of clinical use. The World Health Organisation’s guidance on AI for health¹² identifies bias as one of two risks (alongside privacy) that must be addressed throughout the deployment of AI tools, not just at adoption.

At the point of using any clinical AI tool, ask: “Does this tool’s validated performance apply to this patient?” A tool trained predominantly on certain breeds, species, geographic populations, or disease presentations may produce outputs that are less reliable or systematically biased, when applied to patients that fall outside those parameters.

In a New Zealand context, there are specific dimensions that are not always visible in internationally developed tools, such as:

• Breeds and species with low representation in international training datasets — including many livestock breeds, working dogs, and native wildlife, which may receive less reliable AI outputs than breeds common in the UK, North America, or Europe

• Disease ecology in New Zealand — including conditions linked to specific farming systems, geography, or biosecurity context, which may not be adequately represented in tools developed from overseas clinical data

• Presentations that are common in New Zealand practice but uncommon internationally may be under-weighted in AI’s differential reasoning, leading to outputs that are reasonable for a generic international population but less reliable for your patients

Where you notice that an AI tool consistently produces outputs that seem less applicable to particular animal types or clinical presentations in your practice, document this and raise it through your practice’s incident process. These patterns are important quality information, both for your own practice and, where systemic, for the wider profession.

The sections above focus on the risks of using AI without adequate scrutiny. There is a separate and increasingly relevant question that this guidance addresses directly: could a veterinarian be criticised for not using AI in a complex case, particularly if an appropriate tool was available?

This question is real and the profession will need to engage with it as AI tools become more validated and widely adopted. The American Association of Veterinary State Boards has noted: “Using AI to aid clinical decisions may one day become a catch-22 for Veterinarians. As AI advances, it is possible that consulting with AI may become common practice. Failure to do so may be considered below the standard of care.”⁹ The Vet Council is not raising this to create an obligation where none currently exists, but because veterinarians deserve a clear understanding of the current context of AI use by veterinarians.

AI tools in veterinary medicine currently lack the standardisation, validated benchmarking, and regulatory oversight that would make AI use mandatory in most contexts. Most commercially available tools have limited peer-reviewed validation for veterinary settings, and no professional consensus exists on when the use of any particular tool is expected. In this environment, choosing not to use an AI tool, or not having access to one, does not represent a departure from the standard of care.

Where there is genuine clinical doubt in a complex case, the appropriate response is a specialist referral. This is established in the Code of Professional Conduct and in the Competency Standards. AI is not a substitute for this and should not be positioned as equivalent. A specialist brings verifiable qualifications, professional accountability, the capacity to ask follow-up questions, and direct clinical judgment formed from the full clinical picture. No current AI tool replicates this.

A veterinarian who uses AI in preference to referring a case that warrants specialist review has not discharged their professional obligation. AI may be a useful part of the workup that leads to referral but is not an alternative to it.

Where referral is not pursued (e.g. cost, owner decision, geographic constraints, or clinical urgency) and where a validated AI tool is available and directly relevant to the clinical question at hand, using it as part of a thorough clinical approach is both reasonable and professionally defensible. It serves the same function as any other form of decision support: considered input to the veterinarian’s clinical reasoning, weighed alongside examination findings, clinical experience, and published literature.

The AAVSB draws a useful parallel with specialist consultation: “If the Veterinarian that holds the Veterinary-Client-Patient Relationship disagrees with a consulted specialist, that rationale is documented in the medical record and discussed with the Client as part of the decision-making process.”⁹ The same framework applies to AI: vets can use it as an input, exercise their own judgment, and document their reasoning, including where they departed from what the tool suggested.

The standard of care in any profession evolves as technology is validated and adopted. As specific AI tools accumulate peer-reviewed evidence for specific clinical applications and become more widely used, the expectation that veterinarians are aware of and appropriately consider relevant tools will grow. It is possible to envisage a point where, for certain well-validated diagnostic applications, a failure to use an available and relevant tool could be questioned in the same way that a failure to consult a specialist or run a standard diagnostic test might be questioned today.

This is not the current position. But veterinarians should be aware of this direction and engage actively with AI developments in their clinical area so they are not caught off-guard by an expectation that develops without their awareness.

Whether you decide to use an AI tool or not, your protection is a clear clinical record of the reasoning behind it. Where a validated AI tool was available and you chose not to use it in a case where it was directly relevant, a brief note of why you chose not to refer, or why you departed from a published guideline, should be included. As AI tools become more established in specific clinical contexts, this practice will become more important.

Before using an AI tool in clinical decision-making, you should be able to answer the following. If you cannot, find out before using the tool, or use it only in a way that does not require you to rely on its outputs without independent verification.

• Purpose and fit: Does this tool answer the clinical question I have? Has it been validated in a patient population similar to mine?

• Evidence: Is there peer-reviewed evidence of performance? What are the published accuracy, sensitivity, and specificity figures?

• Validation: Was the tool trained and validated using actual data (e.g. specialist-confirmed diagnoses)?

• Transparency: Can the tool explain why it reached a particular output? Are its limitations clearly documented?

• Data governance: What data does the tool collect and transmit? Will client or patient data be used to train the model? Is data stored in New Zealand or overseas? The latter is regarded as a difficult area in the context of NZ and veterinarians are encouraged to check their AI tool's data storage and retention policies.

• Data privacy: Where you are using a cloud-based AI scribe, is the data used for training? Is there a setting that must be enabled to avoid inadvertent data sharing?

The ACVIM AI Validation Aid (AIVAID) 12-point checklist provides a practical framework for this assessment and is reproduced in the Appendix to this guidance.¹

A prompt that asks AI to “research this case and suggest a treatment plan” will generate exactly that: a plan, presented confidently, without any challenge to your working hypothesis. The AI will not spontaneously critique your reasoning, identify what you may have missed, or ask whether you have considered an alternative differential. It will do what it was asked to do.

If you want the AI to act as a genuine check on your thinking rather than a tool for building the case you already have in mind, you need to instruct it to do so explicitly. Useful prompts for critique include:

• “What important differentials might I have overlooked given this presentation?”

• “What additional clinical information would most change this differential list?”

• “What are the weaknesses in the diagnostic approach I have described?”

• “What assumptions am I making about this case that I should question?”

Deliberately prompting for critique is not optional if you intend to use AI as a genuine second opinion. Without it, the default mode of any case workup prompt is confirmatory. AI builds on what you have provided rather than testing it.

Before consulting any AI output, whether a radiograph analysis, a differential list, or a predicted diagnosis, assess the patient yourself and form your own clinical impression. This protects against automation bias and ensures you engage actively with the case.

AI-generated findings, diagnoses, or recommendations should be treated the same as a colleague’s opinion: useful input for your decision-making, but not a substitute for it. If you disagree with an AI output, document your reasoning.

Any clinical record generated or significantly drafted by an AI tool must be reviewed by you before it is finalised. This means reading it critically, not scanning it. While AI can improve the efficiency of documentation, accuracy must not be sacrificed in the name of efficiency. You are responsible for the accuracy of the record regardless of how it was created. Errors introduced by AI that enter the permanent record under your name are your professional responsibility.

Clients have a right to know when AI is used in their animal’s care. The level of disclosure should be proportionate to the use:

• Administrative AI: General practice-level notice is sufficient.

• Consultation recording and transcription: Clients must be informed and given a genuine opportunity to decline. See the Vet Council’s guidance on AI transcription and client privacy.¹⁰

• Clinical AI: Clients should be informed. Disclosing AI use in practice can be appropriately handled through terms and conditions (referenced at consent or available on the practice website) i.e. it is not considered necessary to require verbal consent for each consultation.

• Of note - Voice recording consent is a separate privacy matter and should be clearly distinguished from clinical consent for AI-assisted decision-making.

Beyond prospective disclosure, clients have a right to understand how AI was used in their animal’s care and to raise concerns about AI-influenced decisions after the fact. International governance frameworks, including the World Health Organisation’s guidance on AI for health¹² and the principles adopted by NHS England,¹³ treat the client’s right to understand and contest AI-influenced clinical decisions as a baseline expectation — not a courtesy.

In practice, this means you should be able to explain to a client not just that AI was used, but what the AI suggested, whether you followed or departed from that suggestion, and your reasons. Where a client raises a concern about a decision that AI contributed to, treat it with the same seriousness as any other clinical concern. The fact that AI was involved does not reduce the practice’s accountability for the outcome.

Where AI has contributed to a clinical decision, the record should reflect this — including the name of the tool used, any points at which your own assessment differed from the AI output, and the basis for the decision ultimately made. Noting the specific tool matters: in the event of a complaint or adverse outcome, understanding which system was involved is relevant to any review.

A complete AI-related record entry should capture enough information to allow the clinical reasoning to be understood and assessed independently: what information was provided to the AI, what the AI produced, and how that output was used, accepted, modified, or overridden, in reaching the clinical decision. This is not additional bureaucracy; it is the record of the professional judgment you exercised. This approach is consistent with the Medical Council of New Zealand’s guidance, which requires doctors to document AI use where it influences clinical decision-making.¹¹

Do not allow AI tools to compromise your ability to work without them. This is a long-term risk that can develop gradually. Ensure you continue to practise diagnostic reasoning and other clinical skills in ways that do not depend on AI assistance. Ways that this may be managed include:

• Draft comprehensive prompts for your chosen tool that include your clinical assessment rather than just inputting clinical symptoms

• Avoid using differential diagnosis prediction tools until you have listed your own

• Develop your own treatment plans before AI review.

As AI tools in veterinary practice rapidly evolve, maintaining awareness of developments in this area is now a professional responsibility. CPD should include AI literacy, informing people how to use and critically evaluate the tools, and how to recognise their failure modes. They should remain current as changes will arise in tools that are regularly used. Tools you have found reliable in the past may behave differently after an update, so always check the release notes when software is updated and re-evaluate whether your assessment of the tool’s reliability is valid.

Individual practitioners are also part of the quality improvement system for AI tools in veterinary practice, not merely passive users of them. When you notice patterns, for example that you are consistently overriding a tool’s recommendations for certain breeds or species, that a clinical scribe routinely removes types of findings, or that a tool’s outputs seem less reliable after an update, these observations have real clinical and professional value. Document them through your practice’s incident or quality reporting process, and raise them with your practice’s AI governance lead. The Federation of State Medical Boards has noted that AI systems must be continually monitored and refined, and that this requires collaboration between practitioners, organisations, and regulators.¹⁴ Your observations are a contribution to that process, not a sign that something has gone wrong.

Reporting a technical product failure (appropriate for the AI provider) should be distinguished from reporting a concern about a veterinarian's AI-dependent decision-making, which can be reported to the Vet Council.

Always perform your own image assessment before reviewing an AI report. Be aware that most commercial tools have been validated mostly on dogs and cats, so performance for other species may be substantially lower. A negative AI finding does not exclude pathology. Sensitivity is often lower than specificity in published studies.⁶ Where an AI imaging service is provided remotely, you remain responsible for the diagnostic conclusion you document.

Always review the output carefully before accepting it. Do not rely on the information just because it is formatted as a SOAP (Subjective, Objective, Assessment, Plan) note. Be alert to hallucinated details, species errors, incorrect drug names, and omitted or spurious findings. Clients must be informed and given the opportunity to decline before an AI transcription tool is used. See the Vet Council’s transcription and privacy guidance¹⁰ for more information.

Where a consultation has been recorded as part of an AI transcription workflow, there is no requirement to retain the audio or video recording in the clinical record once you have reviewed and finalised the written summary. What is required is an accurate clinical record of the consultation, not the recording itself. Keeping recordings beyond this time can introduce risks, such as:

• Privacy: Recordings contain raw, unprocessed personal information about clients. If the system storing them was accessed without authorisation, the privacy consequences would be more significant than breaching a structured clinical record.

• Professional: Recordings capture unguarded language and in-the-moment reasoning that, taken out of context, may not accurately represent the quality of care provided. Retaining recordings indefinitely creates exposure to misuse or misrepresentation.

• Storage and security obligations: Retained recordings create ongoing obligations around secure storage, access controls, and eventual deletion.

Once you have reviewed and finalised the clinical notes, recordings should be promptly deleted. If a practice uses a third-party transcription provider, ensure the provider’s terms of service reflect a clear data deletion policy after transcription is complete. This is addressed in the Vet Council’s guidance on AI transcription and client privacy.¹⁰

Use these as a second opinion, not a primary source. Apply the same critical discipline you would to any clinical consultation. For example, consider what information AI had; if the output is consistent with the full clinical picture you have observed; and if the recommendations apply to this patient in this setting.

Two specific cautions apply:

• Drug dosages: Do not use AI-generated dosages without verifying them against an authoritative source such as a current formulary or product label. AI tools can and do generate incorrect dosage information, including for species the tool was not trained for.

• Citations and references: If an AI tool provides references to support a recommendation, verify each source directly. Confirm that the paper exists, that it says what you have been told it says, and that the study population and methodology are relevant to your case. An authoritative-sounding citation does not guarantee that the underlying evidence supports the clinical recommendation being made.

Where AI suggestions refer to treatment approaches that may be unavailable, restricted, or differently regulated in Aotearoa New Zealand, confirm current local requirements before acting on them. AI tools trained on international datasets may not reflect New Zealand-specific drug availability, prescribing restrictions, or practices.

Wearable monitoring and AI-assisted herd health tools can provide useful early warnings of health issues but are subject to the same limitations that come with all AI technology. Always make sure you understand what species, breed, and production system data was used in the tool’s development before relying on its outputs. New Zealand farming systems, including pasture-based, seasonal, and extensive operations, may differ substantially from settings that the international tools were developed and validated on.

Many veterinarians use general-purpose AI tools, such as publicly-available large language models, to assist with clinical queries, draft communications, or differential diagnoses. This is increasingly common, and when used carefully, can be valuable. However, these tools carry specific risks that purpose-built veterinary AI does not always share. Understanding these risks requires understanding of AI tools.

General-purpose AI tools are trained on massive, largely un-curated datasets pulled from the open internet. These datasets include peer-reviewed scientific literature alongside news articles; social media posts; patient and pet owner forums; product marketing material; blog posts; and general web content of widely varying quality and accuracy. Clinical veterinary content forms a small fraction of this training data and that content has not been selected or validated for clinical accuracy. The model has no reliable mechanism for distinguishing a peer-reviewed randomised controlled trial from a discussion forum post — both contribute to the patterns the model has learned. When you receive a clinical recommendation from a general-purpose tool, there is no way of knowing whether it draws on validated veterinary evidence, outdated information, human medical guidance that does not translate to your species, or informal online content. A confident output will offer no indication of its evidential basis.

This is different from the risk profile of purpose-built veterinary AI tools that are intended to be trained and validated on relevant clinical datasets. General-purpose tools were not designed with veterinary practice in mind, were not trained to weigh veterinary evidence by quality, and have not been benchmarked against veterinary clinical standards.

Key risks specific to these tools include:

• No veterinary validation: Outputs have not been benchmarked against veterinary expertise or tested for species-specific accuracy. For example, the model may apply human medical reasoning to veterinary questions in ways that are not clinically appropriate.

• Privacy: Entering patient details, clinical histories, or other client information into a general-purpose AI tool raises significant privacy concerns. Many tools, particularly free versions, may use inputs to train their models. Do not enter identifiable client or patient information into tools that do not have contractual data protection in place. See the Vet Council’s guidance on AI transcription and client privacy for further details.¹⁰

• Hallucinations: General-purpose models operating in specialist domains where their training data is sparse, are particularly prone to generating plausible but incorrect information. Drug names, dosages, species-specific recommendations, and references to clinical guidelines may all be fabricated. Always verify any clinical information against a primary source.

Adapted from: Bertin F-R et al., ACVIM AI Task Force, Journal of Veterinary Internal Medicine, 2026.¹

Before using an AI tool in clinical practice, consider the 12 questions below. Where the answer to either of the first two questions is “no”, the ACVIM Task Force strongly recommends reconsidering the use of the tool.

The more questions answered “yes”, the greater confidence you can have in the tool. Any tool where questions 1 or 2 cannot be answered “yes” should not be used in clinical decision-making without significant independent verification.

¹ Bertin F-R et al. The influence, promise, and potential perils of AI in veterinary medicine. J Vet Intern Med. 2026;40(1):aalaf045. https://doi.org/10.1093/jvimsj/aalaf045

² VCNZ. Guidance for Veterinary Businesses: AI in Practice. Wellington: VCNZ; 2026.

³ VCNZ. Code of Professional Conduct. Wellington: VCNZ; July 2025. https://hub.vetcouncil.org.nz

⁴ VCNZ. Competency Standards and Performance Indicators for Veterinarians. Wellington: VCNZ; October 2015.

⁵ Heinlein M. Ethical considerations of AI in veterinary medicine decision-making. Front Vet Sci. 2026;13:1780868.

⁶ Pomerantz LK et al. Performance of commercial AI for detection of pulmonary nodules. Vet Radiol Ultrasound. 2023;64:881–889.

⁷ Budzyń K et al. Endoscopist deskilling risk after AI exposure in colonoscopy. Lancet Gastroenterol Hepatol. 2025;S2468-1253(25)00133-5.

⁸ Cascella LM. AI Risks: Black-Box Reasoning. MedPro Group. https://www.medpro.com

⁹ AAVSB. Regulatory Considerations of AI in Veterinary Medicine. AAVSB; March 2025. https://www.aavsb.org

¹⁰ VCNZ. Guidance on AI Transcription and Client Privacy. Wellington: VCNZ; 2025. https://hub.vetcouncil.org.nz

¹¹ Medical Council of New Zealand. Guidance on Using AI in Patient Care. Wellington: MCNZ; March 2026. https://www.mcnz.org.nz

¹² World Health Organization. Ethics and Governance of AI for Health: Guidance on Large Multimodal Models. Geneva: WHO; 2024. https://www.who.int/publications/b/70584

¹³ NHS England. AI Governance: Principles for Healthcare Organisations. London: NHS England; 2025. https://digital.nhs.uk/services/ai-knowledge-repository/ai-in-practice

¹⁴ Federation of State Medical Boards. Navigating the Responsible and Ethical Incorporation of Artificial Intelligence into Clinical Practice. FSMB; April 2024. https://www.fsmb.org/siteassets/advocacy/policies/incorporation-of-ai-into-practice.pdf

¹⁵ RCVS. Standards and Advice: Spring 2026 update. Royal College of Veterinary Surgeons; April 2026. https://www.rcvs.org.uk/about-us/news-and-views/features/standards-and-advice-spring-2026-update