The influence, promise, and potential perils of artificial intelligence in veterinary medicine: a call for improved awareness and literacy

Abstract

Artificial intelligence (AI) is rapidly becoming integrated into daily lives, including tasks unique to our professional domains. With technology outpacing the general knowledge base regarding veterinary AI tools, we are at a critical inflection point in our American College of Veterinary Internal Medicine (ACVIM) and European College of Veterinary Internal Medicine - Companion Animals (ECVIM-CA) community. This manuscript presents a perspective aimed at igniting a broader and deeper discussion of, and engagement with, AI tools among members and enhancing members’ AI literacy where necessary. The ACVIM AI Task Force encourages members to become actively involved in the processes that determine where, when, and how AI technology is adopted in our fields of expertise. Collectively, the principles outlined here promote thoughtful, transparent innovations while upholding standards of modern healthcare. However, it behooves us as potential users to support the need for critical oversight to evaluate and verify the safety and efficacy of AI tools in the routine patient care. To aid with an initial review of AI tools before use, a novel ACVIM AI Validation Factor checklist is introduced.

Introduction

Technology has encompassed almost every facet of life, including daily communication, personal and remote healthcare monitoring, and our work. Once a concept reserved for science fiction, artificial intelligence (AI) is currently transforming industries across the globe—and veterinary medicine is no exception. From voice recognition to tools that assist in diagnosis and treatment, AI is redefining veterinary medicine. Technology incorporating AI now aids in communication, teaching, administrative tasks, disease surveillance, diagnostics, client/patient care, error prevention, and research initiatives.

This is a pivotal time to explore and shape how AI integrates within veterinary medicine. This rapidly evolving, powerful technology has the potential to enhance diagnostic capabilities, optimize the quality of veterinary care, and reduce nonclinical burdens on teams. However, despite the optimism surrounding AI’s potential, critical clinical and ethical considerations must guide its implementation. Collectively, the ACVIM AI Task Force aims to improve AI literacy among ACVIM members, ensuring that AI supports and does not hinder veterinarians, but also truly enhances the well-being of the animals and communities we serve. The challenges associated with comprehension, validation, interpretation, and adoption of AI can be opportunities in disguise for the veterinary community dedicated to avoiding unnecessary harm to animal patients. Through the principles and considerations highlighted throughout this manuscript, members of the ACVIM AI Task Force seek to equip individuals with essential knowledge and tools necessary for practical assessment and possible implementation of current and emerging AI tools. An in-depth review of AI is outside the scope of this article, and readers are encouraged to consult recent and comprehensive reviews on AI.^1–9

The rise of AI in human healthcare

Artificial intelligence refers to the development of computer systems that perform tasks normally requiring human intelligence.^10,11 Artificial intelligence is an umbrella term that describes several underlying facets that drive these intelligent platforms (Table 1, Figure 1). Recognizing the potential power of implementing AI tools in clinical settings, a new field—healthcare data science or healthcare AI—has emerged. Early efforts within this field focused on developing computational tools to aid therapeutic decision-making and advance healthcare research, including computer-aided diagnostic (CAD) systems. There is tremendous potential for AI to be transformative in the personalization of medicine, both diagnostically and therapeutically.

Table 1.

Common terminology and definitions pertaining to artificial intelligence in healthcare.

Term	Definition
Artificial intelligence (AI)	The capability of a computer to perform functions that mimic human intelligence, such as learning, reasoning, problem solving, and decision-making.
Machine learning (ML)	Computer systems that analyze, learn, and adapt through the use of algorithms and statistical modeling of data.
Neural network (NN)	Within machine learning, a mathematical network of connected units/nodes termed artificial neurons that loosely resemble neurons within the brain.
Convolutional neural network (CNN)	Class of deep learning algorithms that recognizes, analyzes, and learns from images or visual data.
Deep learning (DL)	A subset of ML that uses multilayered networks of weights and decisions (NNs, CNNs) to process data and recognize and model complex patterns.
Supervised learning	Machine learning approach that incorporates human instruction to label data based on known outcomes.
Unsupervised learning	Machine learning approach that aims to identify clusters of data by identifying underlying patterns in unlabeled data using input feature variables without the observation or instruction by a human.
Large language model (LLM)	An AI algorithm that uses deep learning techniques to analyze and summarize extremely large, tokenized text datasets to generate and predict new text or data.
Generative AI	Machine learning tool that is trained to generate new data rather than making predictions about existing data.
Agentic AI	An autonomous AI system that can mimic human decision-making and tasks with limited supervision. These systems can respond to conditions, such as a spike in patient demand, and reallocate resources as needed without human input.
Computer vision (CV)	A field of AI that enables computers to interpret and process visual data from the world, such as images and videos. It involves techniques like image recognition, object detection, segmentation, and classification to extract meaningful information from visual inputs. In healthcare, computer vision is used for medical imaging analysis, automated diagnostics, and image-guided interventions.
Pathomics	Computational analysis of histopathological images to extract high-dimensional quantitative features that characterize tissue morphology and pathology.
Radiomics	An advanced imaging analysis technique that extracts many quantitative features from medical images, such as shape, texture, and intensity to characterize gray-level tissue heterogeneity. These features can be used to discover imaging biomarkers that correlate with clinical outcomes.
AI-enabled tool or AI-enhanced tool	A product integrating generic or off-the-shelf AI technology to superficially enhance existing health workflows without specialized clinical training or validation.
AI platform, Health AI platform, or clinical AI system	A solution specifically developed, trained, and validated using health-specific datasets and clinical expertise, directly impacting clinical decision-making, diagnostics, or patient outcomes.

Figure 1.

Artificial intelligence (AI) represents a broad category of computational intelligence. There are multiple facets of learning and intelligence within AI that have contributed to significant advancements and subspecialties of computational learning. Machine learning contains a few layers of algorithmic analysis, whereas deep learning exponentially increases the number of layers utilized in a learning algorithm. Generative AI takes this a step further by generating new data from previously learned data, using a similar network architecture.

Omics techniques are comprehensive assessments of biological molecules (genomics, transcriptomics, proteomics, metabolomics, etc.) or images (radiomics, pathomics), and their interactions. Omics differs from traditional approaches that focus on specific proteins, genes, or pathways by allowing a comprehensive study of biosignals and biomarkers present within patient data. Artificial intelligence approaches (Table 1) now allow for multi-omics integration, enabling the analysis of complex, multi-modal biological interplays to improve disease detection, therapeutic development, prognosis, and outcome prediction.^12–20 Many multi-omics platforms that leverage machine learning (ML) and deep learning have been developed, but few have been translated into routine human clinical practice, as high-level clinical validation remains pending.¹⁴

A landmark study in human breast cancer illustrated the promise of AI, with improvements in early detection rates, decreasing both false positive and false negative results, when radiologists use CAD systems compared to radiologist-alone reads.²¹ In addition to its potential to improve disease detection and precision approaches (treatment directed at features unique to the patient), the promise of reduced human workload, enhanced real or perceived productivity, increased clinician satisfaction, decreased error, and automated administrative tasks propels interest in the adoption of AI tools in medicine.²² Multiple studies support that CAD systems are likely to augment human expertise and improve performance, although there may be unique scenarios in which AI autonomy is acceptable.^22–25

In the realm of large language modeling in clinical medicine, a randomized and single-blinded study was widely publicized, in which a commercially available large language model (LLM) outperformed physicians in diagnostic reasoning specific to 6 clinical vignettes.²² This finding was most likely related to the detailed prompt formulation for the LLM and clinician experience.²² However, the results achieved when AI platforms and humans work together—their synergy—differ significantly depending on the approach used. This variability highlights the need to understand potential biases and differences in study designs that test AI performance in clinical settings.

Deep learning requires high-throughput and high-performance computer hardware and extensive datasets to overcome the unique challenges associated with heterogeneous data, inherent bias, feature selection, dimensionality (using an insufficient number of cases for the number of features you are trying to evaluate), and the needs of modeling algorithms.^14,26,27 The term “black box” refers to the opaque nature of how deep learning and other AI approaches arrive at their conclusions through complex algorithms. There is a black box aspect to many deep learning models; thus, methodology, quality assessment standards, and results must be transparent and reproducible to ensure scientific integrity, safety, and rigor before widespread adoption.^28–30 Moreover, the increasing recognition of AI misinformation, frequently called hallucinations in LLM settings, created through an ML model’s imaginative interpretation of data (eg, correlating radiographic findings with nonclinical meaning or the creation of journal titles that do not exist), requires that healthcare users are aware of AI’s limitations to avoid causing harm.³¹ Good ML Practice principles that promote transparent, safe, and effective development, testing, and adoption of medical devices that rely on AI and ML have been widely adopted by agencies such as the US Food and Drug Administration, Health Canada, and the UK Medicines and Healthcare products Regulatory Agency.^32,33

The emergence of AI in veterinary medicine

Veterinary medicine is experiencing a parallel AI tool transformation. The integration of operational AI-assisted tools that improve daily workflows has already occurred across many practices with the use of software with embedded AI, such as dictation, radiology image analysis, client-facing chatbots, integrated scribes, and medical record summarization tools, among others. Veterinary professionals and trainees across the globe increasingly recognize the need for AI literacy, particularly as pet owners and farmers might already use AI tools to analyze pet or flock and herd health.^34–39 Although readily available tools exist within both closed-source models—where the underlying code is proprietary and not publicly accessible—and open-source models—where the code is freely available for modification and distribution—it’s crucial that solutions are developed and rigorously tested specifically for the unique needs of veterinary medicine.⁴⁰ In the veterinary field, much work remains nascent, focusing on initial steps to validate tools such as pleural effusion detection on thoracic radiographs or correlating histology with CT images.^40,41 Caution is warranted before the adoption of tools not robustly evaluated in the relevant setting.

Unlike human healthcare, veterinary medicine lacks uniform regulatory oversight, and independent state and provincial regulatory boards in North America have unique regulations within their jurisdictions. Importantly, there is no regulatory oversight for the use of AI in veterinary medicine, allowing for faster innovation and integration of AI-driven solutions in research, education, clinical practice, and hospital management. However, this flexibility also presents significant challenges, including small and fragmented datasets, poorly structured data repositories, unintentional release of information without client consent, and the absence of standardized practices or requirements for transparency and good machine learning practices. The American Association of Veterinary State Boards (AAVSB) recently published its position statement on AI to educate veterinary license-holders on the risks and limitations of AI, as well as the regulatory considerations, highlighting that the lack of oversight places the weight of responsibility for using AI tools on the user.⁴² Addressing these considerations thoughtfully will ensure AI enhances, rather than undermines, the integrity and high standards upheld by ACVIM and ECVIM-CA clinicians.

Attitudes of veterinarians toward AI

In a recent survey carried out by the American Animal Hospital Association (AAHA), approximately 40% of veterinary professionals reported using AI tools, with only 15% expressing resistance to the use of AI in veterinary medicine.⁴³ Notably, the survey noted a direct correlation between enthusiasm for AI and the use of AI tools.

An ACVIM and ECVIM-CA survey was conducted by this task force to understand the current perceptions, expectations, and concerns of members, and the results are presented in a companion manuscript DOI: 10.1093/jvimsj/aalaf036. Of 301 respondents from ACVIM and 155 respondents from ECVIM-CA, 53% of ACVIM and 63% of ECVIM-CA respondents considered themselves “slightly knowledgeable” or “not knowledgeable at all” regarding AI. Approximately 42% and 32%, respectively, considered themselves “moderately knowledgeable,” with the remaining minority of individuals either “very” or “extremely knowledgeable.” Without relevant training and education, the applications of AI tools might inadvertently lead to medical errors and patient harm. The ACVIM AI Taskforce calls for continued education surrounding these topics by the veterinary community.^8,44

When asked if individuals use AI tools in practice, 58% of the ACVIM respondents and 67% of ECVIM-CA respondents indicated they are not using AI; yet, when further queried about the use of specific AI-embedded software, of 294 ACVIM and ECVIM-CA respondents that answered “no” to have you used AI tools in your practice, 164 (56%) indicated they have, perhaps unwittingly, used tools that incorporate AI in the platform. This discrepancy highlights how subtly and quietly some AI tools have proliferated across the ACVIM and ECVIM-CA disciplines without overt recognition.

Examples of the incorporation of AI into clinical veterinary practice

An increasing number of commercially available AI tools are available for clinical use in veterinary medicine. The examples mentioned in this section are for illustration; the ACVIM AI Task Force and ACVIM explicitly do not endorse or refute their use, nor the use of products that are not specifically mentioned. Artificial intelligence tools that use convolutional neural networks have shown excellent accuracy in interpreting diagnostic imaging. Artificial intelligence algorithms can assist in detecting abnormalities in radiographs, CT, and MRI scans, such as pulmonary nodules, left atrial enlargement, fractures, or spinal cord compressions.^45–49 Several companies currently provide real-time AI analysis of images, while others are actively engaged in opportunities to test AI-assisted or AI-driven technologies for workflow incorporation (Artificial Intelligence in Veterinary Medicine. Whitepaper. June 2023. Accessed March 12, 2025, https://resources.vet-ct.com/vetct-whitepapers/artificial-intelligence-in-veterinary-medicine). Clinics can upload radiographs and receive AI-assisted analyses within minutes. However, the accuracy and reliability have not been robustly evaluated, and likely heavily depend on the quality of the input data, including imaging clarity and proper data annotation. A recent study compared 50 anonymized thoracic and abdominal radiographs from dogs and cats and found that a commercially available radiography platform performed best in the setting of normal radiographs. Still, it was less effective than veterinary radiologists at detecting abnormal findings.⁴⁷ However, limitations of this study have been explored with a call for larger studies with adherence to reporting standards to better evaluate AI tool performance in the clinical setting.⁵⁰

Artificial intelligence tools can also process electronic medical records (EMRs) to predict outcomes and diagnoses, flag patients at risk of adverse events, and suggest evidence-based treatment plans.^51,52 For example, predictive analytics in small animal medicine can forecast the likelihood of chronic kidney disease progression in feline patients or assess mortality risks in general practice^53–55 Some software integrates AI within its analysis of EMR data, aiming to provide insights on treatment outcomes and enhance client communication strategies. However, published data on improvements in outcome measures after integrating AI into EMR systems are currently limited.

In small animal pathology and dermatology, AI tools have been developed that detect cell types in hematology and differentiate between benign lesions, infections, or malignancies.^56–58 Some AI-assisted tasks might improve efficiency and reduce subjectivity in identifying histopathologic indices like mitotic count or nuclear pleomorphism.^59–61 As with other platforms, the performance of these tools is directly linked to the quality and diversity of the data used for training and input. A current limitation is a scarcity of large datasets available in veterinary medicine.⁶²

Ruminant and large animal medicine have also seen AI tools evolve that seek to improve efficient disease detection. Herd health monitoring systems, integrated with wearable sensors, track temperature, heart rate, and activity levels to detect early signs of illness like mastitis, lameness, or disease outbreaks.^63–65 For example, a specific software uses AI to monitor cow activity, including ruminal health and milk production, to reduce manual labor and improve herd management.^66–68 In reproductive management, AI assists in monitoring estrus cycles and predicting ovulation through data collected from sensors and imaging, optimizing breeding programs. These systems have shown utility in improving general production performance metrics in livestock.^69–71

Equine care has seen advancements in precision medicine through AI tools such as gait analysis for lameness evaluation. Motion sensors and AI models can identify subtle abnormalities in stride, assisting in early detection and management.^72–75 Additional AI tools are under evaluation globally to improve gait analysis in sport horses and during lameness exams.^76,77

Opportunities for AI integration in veterinary research

Artificial intelligence impacts and transforms veterinary research by offering tools to analyze vast datasets, enabling epidemiological studies, and fostering collaborations across institutions. In epidemiology, AI algorithms process large-scale data to track disease outbreaks and understand transmission dynamics.^78,79 For instance, AI models can predict the spread of lumpy skin disease within herds or map the risk factors for zoonotic diseases.^80–82 Collaboration platforms in the context of One Health provide a framework for integrating data from multiple sources, including veterinary hospitals and research institutions, to develop predictive models for disease control.^83–86 As with other AI models, the reliability of these models is dependent on the quality of the input data, which varies widely between regions and institutions.

Artificial intelligence technology is increasingly used in veterinary research to investigate patterns of antimicrobial resistance and the use of antimicrobials in companion animals and livestock.^87–92 These advancements improve animal health and understanding of resistance mechanisms across species and contribute to global food security and one health initiatives.

Artificial intelligence can also support various aspects of scholarly work, such as grant applications, manuscript writing, and literature reviews. Specialized software helps streamline these complex processes. For example, tools have been developed to facilitate literature curation, while other platforms assist in drafting and refining research proposals. Moreover, data analysis tools can manage extensive datasets with routine statistical analyses and enhancing libraries with AI-coding tools (termed “vibe-coding”), generating invaluable insights and proving essential for successful research funding applications if allowed by targeted agencies.

Integrating AI with teaching and training

Artificial intelligence is reshaping veterinary education, offering a suite of interactive learning tools and simulation platforms that have piqued the interest of many students.^35,36,93,94 Virtual reality systems, powered by AI, provide immersive training experiences for veterinary students, allowing them to refine surgical skills in a risk-free environment or improve communication skills.^95,96 The role of AI in improving learning opportunities for trainees, including post-graduate and resident training programs, is a compelling area of focus for many educators.⁹⁷

Students have shown a generally positive reception to AI integration in veterinary training.^98,99 Faculty members express more skepticism about its utility and effectiveness, supporting the need for further discussion, training, and understanding.¹⁰⁰ Artificial intelligence platforms offer personalized learning by assessing individual student performance and adapting educational content accordingly.^100–102 Artificial intelligence-generated, tailored feedback on case-based learning modules might further enhance the educational experience for students, interns, resident trainees, and practicing veterinarians if models are generated purposefully and carefully.

Leveraging AI in clinical management to improve productivity and balance

Integrating AI into veterinary workflows might improve efficiency and reduce burnout by automating routine tasks like data entry, diagnostic processes, and therapeutic planning.^103–105 For example, veterinary students using AI-assisted imaging tools reported increased confidence in their diagnostic accuracy, leading to better patient care.^35,93 Long-term health benefits might also be realized through reduced physical demands enabled by automated AI-driven herd health monitoring in large animal settings.¹⁰⁶

These benefits, however, should be balanced against the potential for tool or interpretation inaccuracies when improper validation and ongoing quality control have occurred. A false sense of security about the robustness of a tool might promote improper reliance and lead to potential adverse outcomes. Furthermore, there might be a temptation to leverage AI tools to improve efficiency without regard for improving professional well-being, which might be counterproductive.

Ethics and governance that ensures responsible use of AI in veterinary medicine

Ethical and regulatory concerns surrounding AI use persist.^9,107 A flexible governance framework can address these issues and adapt to emerging challenges, by aligning ethical standards, safeguarding animal welfare, preserving veterinary autonomy, and maintaining public trust. Diverse input from key stakeholders—developers, clinicians, animal owners, legal experts, and ethicists—must be incorporated into governance decisions.¹⁰⁸

Bioethical considerations are crucial in shaping a governance framework for AI in veterinary care. These principles help ensure that AI technologies are implemented responsibly, fairly, and transparently while protecting the well-being of animals, clients, and veterinary professionals. Key areas that contribute to an ethical, reliable, transparent, and secure product include the use of core ethical principles (Table 2) as they relate to patients, owners, and clinicians. Artificial intelligence algorithms must be explainable to both veterinary professionals and owners to foster trust and ensure ethical accountability. This lack of explainability presents a challenge as clinicians want to understand the reasoning behind specific AI recommendations, decisions, and conclusions to rationalize the clinical decision suggested.⁸

Table 2.

Core ethical principles that guide the framework for artificial intelligence in veterinary medicine.

Core principles
Autonomy	Respect the autonomy of animal owners in decision-making through informed consent
Beneficence	Ensure AI improves animal welfare and patient outcomes
Justice	Ensure fairness in access to AI tools and avoid automation of bias
Non-maleficence	Minimize risks and prevent harm caused by AI errors or biases
Operational guidelines
Transparency	Transparent in the operations with avoidance of the “black box” problem129
Insider	Transparency between the model and developers
Internal	Transparency between developers and veterinary users
External	Transparency between veterinary users and clients
Explainability	Establishment of mechanisms to explain decisions to clinical users and veterinary clients
Animal owner consent	Details must be provided to owners regarding how and why AI technologies are used in the provision of veterinary care to acquire written consent
Bias mitigation	Design should mitigate bias across AI datasets and algorithms
Privacy and data protection	Animal and owner data must adhere to strict governing guidelines for collecting, storing, processing, and sharing data to protect privacy data
Stakeholder engagement	Incorporation of veterinary clinicians, researchers, end-users, ethicists, specialists, and policy-makers when gathering perspectives for development and use
Implementation and review
Education and training	Incorporation of ethical training as it pertains to use of AI extensions and ML/AI tools
Ethical review	Ensure ethical review processes are tailored to AI projects where indicated, and that evolving ethical standards and veterinary care goals are considered
Monitoring and re-evaluation	Monitor outcomes of AI tools in veterinary medicine to identify and relay unexpected ethical challenges, clinical accuracy, or adverse outcomes
Accountability and oversight	Clarify responsible party accountable when AI tools make errors or cause harm, and determine how liability applies in accordance with local laws

The lack of explainability and transparency in AI systems can hinder clinicians’ understanding, undermining their autonomy and ability to obtain informed consent from pet owners. This might result in overreliance on algorithms rather than clinician reasoning, especially in complex cases, and lead to skill erosion.¹⁰⁹ Understanding how algorithmic results are derived can enhance consultations, akin to seeking a second opinion.¹¹⁰ However, the issue of AI autonomy in clinical decision-making is multifaceted. If an AI model consistently outperforms clinicians, its use might still be ethically justified, even without being fully explainable, as in human stroke diagnosis and the prediction of hypotension.^111,112

Transparency and explainability of AI tools are vital to foster trust within the veterinary community and the public.¹¹⁰ The lack of a regulatory framework for AI veterinary tools places a burden of responsibility on veterinary professionals for AI-driven decisions. Regular ethical reviews of AI tools can ensure that they align with ethical standards and veterinary care goals. In larger organizations, governance teams that include bioethics advisors could oversee AI deployment and address ethical dilemmas. The establishment of an industry association to create and maintain AI standards, promoting quality and public trust, could benefit individual practices, similar to existing organizations like the Association of American Feed Control Officials (AAFCO) and the National Animal Supplement Council (NASC). Without such an organization, a governance framework can define guidelines to identify reliable AI products, recognize questionable ones, evaluate their performance, report errors, and ensure their ethical development over time. Clinicians should use ethical principles to evaluate AI systems and maintain practice standards until such time as standards are adopted. With their recent guidance document, the AAVSB examined key regulatory considerations that pertain to the review and adoption of AI technologies.⁴² As AI tools become widespread, it will be crucial for practice acts to clarify how the use of AI tools fits within existing laws and regulations, and how disciplinary actions might evolve to address AI in clinical decision making.¹¹⁰

Fundamental guidelines for AI evaluation in veterinary medicine

To help clinicians interpret results, guide clients, and ensure results are actionable, an AI literacy framework is required. Several detailed, comprehensive reviews have described the various methodologies used to train AI tools.^113–115 When using any diagnostic tool, 2 fundamental questions must be asked: (1) What is the tool’s purpose? (2) What is the ground truth to which the results are compared?

Understanding the purpose of a test is vital for accurate interpretation. For instance, a serologic test for antibodies to an infectious agent asks, “Does this animal have circulating antibodies to this organism?” It does not ask, “Does this animal have disease X?” Instead, it is up to the clinician to interpret a positive antibody response as indicative of disease X. The same principles apply to AI tools. While an AI tool might identify patterns in thoracic radiographs associated with cancer, the clinician must understand the test’s development, performance metrics, disease biology, and limitations before concluding, “this animal has cancer.”

Model development

Newly developed diagnostic tests are compared to a ground truth, a reliable established methodology, to determine performance statistics such as sensitivity and specificity (Table 3). Similarly, when evaluating AI tools, it is crucial to consider the ground truth methodology and objectivity used in the evaluation. The more objective and concrete the ground truth is, the greater the confidence in the resulting performance statistics. Some veterinary data are inherently subjective, such as radiographic interpretation of probable neoplasia in the absence of histopathology. To circumvent this challenge, expert panels or consensus labeling might improve input when objective measurements are impractical.

Table 3.

Common performance statistics or model metrics applicable to the evaluation of machine learning and artificial intelligence tools.

Performance statistic or model metric	Definition
Sensitivity	The true positive rate. The ability of a diagnostic test to correctly identify all affected individuals.
Specificity	The true negative rate. The ability of a diagnostic test to correctly identify unaffected individuals.
Accuracy	The proportion of correct classifications, both true negatives and true positives, out of all predictions made.
Area under the receiver operating characteristic (ROC) curve	An ROC curve is a plot of the true positive rate against the false positive rate. The best performing predictors have an area under the curve (AUC) of 1 and a random guess would have an AUC of 0.5.
Positive predictive value (PPV)	The proportion of positive test results (false positives and true positives) that are truly positive. This value is affected by disease prevalence.
Negative predictive value (NPV)	The proportion of negative test results (true negatives and false negatives) that are truly negative. This value is affected by disease prevalence.
Precision	Measurement of how accurately a model identifies true positives among all instances it predicts as positive (synonym for PPV).
Recall	The proportion of true positive cases that the model correctly predicts as positive (synonym for sensitivity).
F1 Score	The harmonic mean of precision (PPV) and recall (sensitivity).
Variance	The variability or spread of the model predictions when using different subsets of the training dataset. High variance may imply a model is overfit.
Bias	Systematic errors within the model that leads to incorrect assumptions about the dataset. Bias can occur in simplified models that cannot accurately capture data complexity, or with small datasets.
Overfitting	Overfit models have high variance and provide accurate results for the training dataset, but not for the test dataset.
Underfitting	Underfit models have high bias and provide inaccurate results for both the training dataset and the test dataset. Underfitting can occur if the model is too simple to capture data patterns and variability.

All AI tools require substantial data for training, with sample sizes ranging from hundreds to tens of thousands, depending on the complexity of the problem. The sample size also depends on the effect size. Unlike traditional scientific studies, there are no standards to determine necessary sample sizes a priori. A proposed guideline suggests starting with 10 data points for each feature in the training model, with more complex estimates available through post-hoc analysis.^116,117

When training a model, the data should reflect the target population for deployment. A common pitfall with veterinary AI tools is using training data from academic hospitals with complex or unusual cases, which might not represent the average cases in first-opinion clinics. Consequently, the model’s output might not be applicable to first opinion practice. In addition, if a model is trained on animals with a high pretest probability of having disease X, it should only be used for pets with a similar high pretest probability. An example is the ML model to predict leptospirosis in dogs.¹¹⁸ All dogs in the dataset had a high pretest probability of leptospirosis, even when the disease was ruled out, limiting the applicability of the model in dogs with low pretest probabilities.¹¹⁸

It is also vital to recognize that some AI models might continuously learn from end-user data that is entered into AI applications, and it is critical that end-users evaluate the terms and conditions and privacy policies within any AI application. Although AI use in veterinary medicine is not well regulated, local laws might govern the use of veterinary data for ongoing model development. For example, California passed the “Generative Artificial Intelligence: Training Data Transparency Act” (Assembly Bill 2013) in 2024, which will take effect on January 1, 2026, and requires developers to disclose detailed information about datasets used to train their models. Other local jurisdictions might introduce data use and privacy laws applicable to veterinary use that must be considered when adopting tools, particularly if client information is included.

Considerations when validating veterinary applications using AI

Veterinarians must determine whether any test, traditional or novel, is accurate and precise. Over the years, good laboratory practices have been the foundation for assessing test reliability by assessing test intra-assay and inter-assay coefficients of variation, along with appropriate reference intervals and detection limits (both upper and lower).¹¹⁹ For instance, when measuring thyroid hormone concentration, veterinarians ensure the assay accurately measures the targeted analyte without interference from other substances, referring to analytical sensitivity and specificity. Since testing conditions can change over time, laboratories should participate in ongoing internal and external quality assurance (QA) programs.

The responsibility of the test provider includes the validation process, and the test user must ensure it is being performed. As previously noted, the test should only be used when there is an appropriate pre-test probability. If a well-validated free T4 assay is applied to cats with concurrent diseases with a low total T4 and lacking signs of hyperthyroidism, there is a risk of a false positive despite the assay’s good analytical performance.

With the increased creation and use of AI tools for diagnosis and monitoring, it is essential to realize that uniformly agreed-upon minimum standards do not exist for validation and QA processes. As a clinician or researcher, it is prudent to ensure that proper validation and QA procedures have been completed and are actively maintained before employing such tools for clinical or research purposes, especially when they might impact patient welfare.

Artificial intelligence tools are particularly vulnerable to producing unreliable results when faced with uncommon scenarios or testing populations that differ significantly from the group whose data was used to train the AI model. Factors including geographical variations, breed differences, and the use of differing clinical terminology and treatment practices can interfere with the tool’s performance. For example, a Cushing’s Prediction tool was developed using a large database from primary care practices in the United Kingdom.⁵² The model showed reasonable transportability when the same predictors for a Cushing’s diagnosis in dogs were applied in a different country, the Netherlands, and a different population, a referred population. However, distinct differences emerged that could lead to suboptimal performance of the tool.¹²⁰ A robust description of the patient population used for model training is essential to understanding when it would be appropriate to apply the tool.

When evaluating literature describing an AI tool, one must ensure the dataset is sufficiently large to answer the posed question and that it is randomly divided into training (70%-80%) and testing (20%-30%) subsets (Figure 2). The model’s performance on validation and test sets should be described and compared to the ground truth and outlined in standardized reporting guidelines.^52,119 Common performance statistics can be applied to evaluate AI tool performance (Table 3). However, note that accuracy can be misleading, with low disease prevalence. For non-binary outputs, a receiver operator characteristic curve can assess sensitivity and specificity across various thresholds.

Figure 2.

Schematic illustration highlighting the breakdown of datasets comprising healthcare data, and the division of data into training and testing sets. Clinical data might include signalment, clinicopathologic data including images, clinician notes from the electronic medical record, radiographic images, and many other sources. Data can be combined and subsequently randomly divided into training datasets for model development, while a smaller subset is reserved for testing the trained model. The model’s performance on training and validation datasets should be compared to the ground truths or modern reference standards. Created in BioRender. Reagan (2025). https://BioRender.com/n5h37rn.

Throughout development and during clinical monitoring and post-market surveillance, clinical benchmarks and validated metrics will be crucial performance measures. In addition to performance metrics like accuracy, sensitivity, and specificity, the F1 score is a critical AI model evaluation metric that provides a more nuanced assessment of a model’s capabilities by combining precision and recall (Table 1). It therefore handles imbalanced datasets, can assess consequences of errors (such as false positives or false negatives), and provides a reliable benchmark for future comparisons. In a recent study, LLMs trained on human health data performed poorly on veterinary medical records, with a low F1 score.¹²¹ Reporting multiple performance metrics will improve the overall assessment of an AI tool, and provide an objective means by which to compare models within the same clinical scenarios.

Evaluating clinical performance in the setting of research and clinical trials

It is vital to evaluate clinical performance before incorporating AI tools into routine practice. In medicine, there are strong drivers to accelerate the implementation of AI interventions, even in the absence of evidence.¹²² Established, standardized study design, conduct, and reporting guidelines (SPIRIT-AI, CONSORT-AI, CLAIM, TRIPOD-AI, and DECIDE-AI) provide minimum guidelines for reporting to aid in the interpretation and critical appraisal of the clinical trial design when there is an AI component.^123–127 Key elements of these AI extensions include the instructions for or skills required by the user, the setting in which the AI intervention is integrated, input and output data handling, the relationship between the human and AI tool, and error analysis.^123,124 Analytical performance is also a significant component of clinical performance; however, clinical performance might encompass other factors, such as the ease of clinical integration, the accessibility and interpretability of the output, and performance in the target population.

Proposed veterinary AI-validation scheme

It is important to establish a systematic approach for assessing AI tools, as evidence supporting most tools is unavailable. To assist in this effort, a proposed checklist for clinicians and researchers, outlined in Table 4, offers a guide for a semi-objective evaluation of the trustworthiness of AI-assisted tools. This checklist is not exhaustive and will require modifications as the field evolves quickly. The ACVIM AI Task Force believes that this checklist will help individuals unfamiliar with AI technology to assess the quality of the validation of AI-based tools.

Table 4.

Proposed ACVIM AI validation factors.

12-item ACVIM AI validation factors
Questions to ask about an AI-driven or AI-assisted application:	0 = no/not sure 1 = yes	Example or explanatory comments
Does the application answer the precise question asked?	0/1	AI tools exist to predict the likelihood of azotemia in cats in the future. They do not answer the question: should I start a renal diet now? Should the AIVAID score be zero for this item, strongly re-consider using the tool.
Does my patient resemble the population that was used to create this test?	0/1	A Cushing’s prediction tool exists to predict the likelihood of a diagnosis of Cushing’s in dogs. Data were used from UK GP practices. There may be differences in performance when using this tool in different settings, such as referral practices or in distinct geographic regions. Should the AIVAID score be zero for this item, one must interpret the tool’s results with the greatest of care.
Am I able to provide the correct and comprehensive input (eg, patient data, biological sample, image) needed for the AI tool to generate reliable output?	0/1	AI tools have been developed to answer questions when specific input is available to evaluate. For example, if an AI tool heavily relies on accuracy of age of the patient and the age is unknown, its performance will be suboptimal.
Has functional testing taken place?	0/1	As a clinician/researcher, do I have access to peer-reviewed publications where the tool was tested on real-life patients to answer the same question I have, and were answers accurately generated?
Has cross-validation been performed?	0/1	As a clinician/researcher, do I have access to peer-reviewed publications where the dataset was split into subsets and the AI tool was trained on one subset, followed by testing on the other subset?
Has A/B testing taken place?	0/1	AI tool development usually involves creating not one, but several models, each with distinct performance characteristics. As a clinician/researcher, do I have access to peer-reviewed publications where different user groups are testing each model? The use of user groups can reduce bias during model development and verification.
Has ground truth validation taken place?	0/1	This type of validation compares the performance of the AI tool with that a known standard, eg, diagnoses made by a group of board-certified specialists. Do I, as clinician/researcher, have access to peer-reviewed publications, where ground truth validation is described for the tool in hand?130
Has performance, robustness, and usability testing taken place?	0/1	This testing assesses a model’s accuracy under diverse conditions impacting performance, robustness, and usability. Examples include: Performance: Evaluating an AI-driven urinalysis system using urine samples tested at various temperatures or under conditions where users input incorrect or unexpected data. Robustness: Determining whether acquisition variations, such as patient movement or machine differences during imaging procedures, affect diagnostic tool accuracy. Usability: Assessing whether voice recognition software maintains accuracy when used by clinicians with varying accents.
Have ethics and bias testing taken place?	0/1	AI-generated results may be prone to bias toward commonly seen trends in the training population. This concept is especially important in veterinary medicine, where unique breeds are predisposed to different disease predispositions, and different owner populations make treatment decisions based on a variety of circumstances, including socioeconomic factors. This might therefore render, for instance, a cancer diagnosis more quickly in a predisposed breed.
Have cybersecurity and privacy testing been addressed?	0/1	While rarely addressed in peer-reviewed publications, it is essential to have knowledge and legally valid assurance regarding an AI tool’s resilience against hacking and other digital intrusions. Clinics and research institutions in most countries bear the responsibility of ensuring the security of the patient data they possess, even when trusted to a third-party AI provider. The vendor’s intent to utilize input data for further model refinement and training must be transparent.
Has the tool been tested in the real world by one or multiple independent parties? Will it continue to be tested on an ongoing basis as part of a quality assurance program?	0/1	Real world testing is the ultimate way to know whether a tool accomplishes what it is intended to do. Testing should ideally be conducted by independent parties, using real-world cases and real-world circumstances. Since many factors influence the consistent performance of a tool and many of these factors (including disease behavior, population characteristics) change over time, repeated testing should be part of a quality assurance program.

(Continued)

Table 4.

Continued

12-item ACVIM AI validation factors
Has regulatory compliance testing taken place by the tool developers? Is the use of the tool by the clinician/researcher in compliance with local veterinary and non-veterinary rules?	0/1	Clinicians and researchers should always ensure that the use of an AI tool, including the submission of client and animal details, complies with local veterinary and non-veterinary rules and legislations. An example includes compliance with a recently implemented Artificial Intelligence Act of the European Union (https://data.consilium.europa.eu/doc/document/ST-5662-2024-INIT/en/pdf).

Total factors met (0-12).

As part of this ACVIM AI Validation Factor 12-point checklist, the user can determine how many criteria are met to help assess different AI applications. Although no specific cut-offs are suggested, the general principle is that the higher the number of factors met, the greater confidence that the user might have in the product. Importantly, if the first 2 items are answered as “no,” the use of the AI tool in question is strongly discouraged, regardless of other robust characteristics it might display. Future work should aim to develop a validation score in the clinical setting to determine the utility and help determine cut-offs that would equate with trustworthy AI tools.

Integrating AI-assisted or AI-driven tools into clinical practice

While not all AI tools will be integrated with the EMR, understanding barriers to successful integration is vital. Many AI tools are proprietary, limiting customization, local integration, and adaptation.¹²⁸ External AI tools might require manual data entry, which could be time-consuming and error-prone. Clinical integration should therefore be part of AI development.

Effective use of AI tools requires adequate training by end-users to facilitate critical assessment. Continuing education with updated content will be necessary for literacy maintenance. Incorporating AI tool analysis into veterinary education curricula will also be key, as many students recognize AI’s potential and influence but possess low literacy and lack formal training.⁹³

Conclusions and future directions

AI is poised to transform veterinary training, care, and research. The ACVIM AI Task Force encourages member engagement in developing and assessing AI applications, connecting innovation with care. Supporting AI education will empower our members to ensure that AI benefits our profession, pets, and their families. Collectively, we can shape AI to prioritize animal health, the communities we serve, and veterinarian job satisfaction. While challenges exist in understanding, validation, optimization, and adoption of AI, they also offer opportunities to leverage members’ knowledge to advance AI-driven technologies in veterinary medicine.

Abbreviations

ACVIM

American College of Veterinary Internal Medicine

AI

artificial intelligence

AIVAID

ACVIM AI validation aid score

CAD

computer aided diagnostic systems

ECVIM-CA

European College of Veterinary Internal Medicine - Companion Animals

EMR

electronic medical records

LLM

large language model

ML

machine learning

QA

quality assurance

Author contributions

Francois-Rene Bertin (Conceptualization, Writing—original draft, Writing—review & editing), Jessica Lawrence (Conceptualization, Writing—original draft, Writing—review & editing), Stijn Johannes Maria Niessen (Conceptualization, Writing—original draft, Writing—review & editing), Christopher J. Pinard (Conceptualization, Writing—original draft, Writing—review & editing), Krystle Reagan (Conceptualization, Writing—original draft, Writing—review & editing), and Virginia Rentko (Conceptualization, Writing—original draft, Writing—review & editing). All authors contributed equally to this work as members of the American College of Veterinary Internal Medicine Artificial Intelligence Task Force and share first authorship.

Conflicts of interest

All authors are members of the American College of Veterinary Internal Medicine Artificial Intelligence Task Force. In addition, C. Pinard is funded by the OVC Pet Trust and American Kennel Club for Artificial Intelligence-based research. C. Pinard is the CEO and co-founder of ANI.ML Health and their research subsidiary focused on artificial intelligence research and commercial applications. S. Niessen is a regular independent consultant for and recipient of grants from numerous pharmaceutical companies; currently working for the Veterinary Information Network, which is developing independent AI tools for use by veterinarians. K. Reagan has received research support from EveryCat Foundation, The Center for Companion Animal Health at University of California, Davis, and Mars Veterinary Health to conduct research related to artificial intelligence.

Funding

The authors received no specific funding for this work.

Off-label Antimicrobial Declaration

The authors declare no off-label use of antimicrobials.

Institutional animal care and use committee or other approval declaration

The authors declare no institutional animal care and use committee or other approval was needed.

Human ethics approval declaration

The authors declare that human ethics approval was not needed.