A chatbot instead of a vet? When AI will help and when it can harm the animal

The cover illustration was created using generative artificial intelligence. It does not depict a real veterinary practice or identifiable people.

The dog starts vomiting. The cat stops eating. A lump appears on the abdomen. The mobile phone is at hand, the answer comes in a few seconds and sounds calm, professional and specific. This is precisely the strength and danger of today's chatbots: the fluency of the answer is not a certificate of truthfulness.

My main message on TV was simple:

AI can help the owner to better navigate and ask better. However, it must not become a reason for postponing the examination of a sick animal.

One in ten uses AI. It does not mean that one in ten have left the vet

The RSPCA's Animal Kindness Index 2026is based on a representative weighted UK survey: YouGov polled 5,797 adults and 1,189 children. Ten percent of pet owners cited AI as a source of advice, while 64% cited a veterinarian. At the same time, 75% of owners experienced an increase in maintenance costs.

This is an important correction for dramatic subtitles. The survey does not say that a tenth of people have replaced the vet with a chatbot, nor does it measure the number of animals harmed. But it does show that AI has already entered the routine decision-making of breeders, and that financial pressures can increase the temptation to get a quick, free tranquilizer.

What does AI "hallucinate" mean?

The word hallucination is experienced, the more accurate technical term is confabulation. NIST describes itas a situation where a generative system confidently presents erroneous or false content.

A chatbot usually does not sit on top of a transparent database of verified veterinary diagnoses. From the learned patterns, the context and, if necessary, found sources, he composes an answer that should be useful and linguistically plausible. When it has quality search, expert background and well-designed security rules, the risk can be significantly reduced. However, it does not disappear just because the answer contains technical names, a table, or a citation.

Therefore, there is not one fair percentage for the question "how often AI hallucinates in veterinary medicine". The result depends on the model, version, animal species, task type, input completeness, resource access, and metric. It is one thing to explain a laboratory abbreviation and another to use a photo, a few sentences and decide whether it is an acute condition without an examination.

Veterinary medicine is not a quiz with one correct answer

The vet doesn't just ask "what's the diagnosis?". He finds out what he doesn't yet know, which dangerous possibilities he must rule out, what examination he needs and how quickly it is necessary to act. He sees the animal's movement, breath, mucous membranes and soreness; can listen, feel, measure and take a sample. A chatbot only has what the user tells it.

A study of twenty neurological cases of dogsshowed this vividly. The ChatGPT of the time did not correctly identify even one of the five cases of paroxysmal dyskinesia that were initially formulated. After slightly adjusting the grammar and words, the results varied, although the medical meaning remained the same. The authors cautioned that some of the wrong answers sounded reasonable and were difficult for a layperson to recognize.

It's fair to add that this was a little work from 2023 with the older GPT-3.5 model, not a benchmark of today's systems. Its value is not in the universal percentage of error, but in the demonstration of instability, dependence on formulation, and convincingness of error.

A more recent study in human medicine showed a similar gap. JAMA Network Open in 2026tested 21 state-of-the-art models on 29 clinical vignettes. The models handled the final diagnosis significantly better than the creation of a complete differential diagnosis, i.e. a list of possibilities that must be considered and excluded. These are not veterinary data, but they illustrate well the difference between hitting the mark and getting through the clinical process safely.

The second risk: the chatbot can be guided

When one asks "it's just going to be stress, isn't it?", one is not giving the model a neutral description. It offers him a hypothesis as well as a desired emotional outcome. Some models tend to cater too much to the user and confirm his frame. This behavior is called sycophancy or condescension.

In veterinary medicine, it is dangerous in two ways. A chatbot can falsely reassure a person and delay care, or, on the contrary, develop a catastrophic scenario and cause panic. A better specification therefore separates observation from conjecture:

However, even a good prompt will not turn a chatbot into a veterinarian. It just reduces one part of the risk.

Where AI can really help

A review in Frontiers in Veterinary Sciencedescribes a number of useful roles for generative AI in the clinic, education and research. For the common breeder, it makes sense in particular:

• translate the veterinary report into comprehensible Czech and prepare questions for inspection,
• sort out a timeline of symptoms, diet, medication and behavior changes,
• create a list of information that the owner should communicate to the veterinarian,
• explain technical terms, routine follow-up care or the principle of examination,
• compare several searchable professional sources,
• help with administration, summary of documentation and education under the supervision of an expert.

Therefore, I do not see the most interesting future in the battle of AI against veterinarian, but in the cooperation of veterinarian with well-validated AI. More recent scholarly debate recommends keeping AI in the role of a clearly delineated, critically reviewable support tool, rather than giving it authority over clinical judgment (Frontiers, 2026).

A practical traffic light for owners

When to close the chatbot and call emergency

According to the Cornell University College of Veterinary Medicineand MSD Veterinary Manualreviews, immediate veterinary contact is required, among other things:

• difficulty breathing, blue, very pale or purple mucous membranes,
• unconsciousness, collapse, seizures or sudden inability to stand,
• heavy or unstoppable bleeding and serious injury,
• suspicion of poisoning or ingestion of medicine or a foreign object,
• inability to urinate, especially in a cat,
• severe pain, significantly swollen abdomen or rapid deterioration.

Don't continue a long chat at such a time. Call the vet, describe the condition and follow their instructions for first aid and transport.

How to ask AI more safely

If it is not an acute condition, the following procedure may help:

1. Write an observation, not a diagnosis. Include species, breed, age, weight, length of difficulty, behavioral changes, water and food intake, medications, and development over time.
2. Ask for uncertainty. Ask what can't be determined from the text or photo and what information is missing.
3. Request sources and actually open them. A fabricated quote can look just as professional as a real one.
4. Don't let the AI decide to wait. If an acute condition is suspected, verify the urgency by phone with the vet.
5. Do not give the chatbot sensitive personal information without thinking. Veterinary documentation may contain the owner's name, address, contacts and payment information.

My conclusion

Artificial intelligence expands access to knowledge and can improve the conversation between owner and veterinarian. He can explain a term in a moment, organize chaotic notes and draw attention to a question that did not occur to you. But the same speed and eloquence can create a false sense of security.

AI's biggest problem isn't absurd error. It's a reasonable sounding mistake when it comes to health and time.

So let's use AI to lead us to better information and better human care, not as a convenient excuse to postpone professional care.