Where AI assistants get pet nutrition wrong
Where: An AI assistant answering a pet food question is summarising the patterns in its training data, which is to say the patterns on the open internet. That is its strength and its weakness. When the internet broadly agrees with the evidence, the answer is fine. When a myth is more common online than the correction, the assistant tends to repeat the myth, fluently and confidently, because fluency is what it optimises for. We track where pet nutrition answers go wrong, and the failures cluster in predictable places. This article catalogues the recurring ones, explains why an AI lands on them, and sets each against the primary evidence. The point is not that AI is useless for pet questions, but that confident phrasing is not a quality signal.
Last updated :General documentary information. For an individual animal, a veterinarian's advice takes precedence over any online content.
Why AI inherits the loudest claim, not the truest
Three mechanics explain most of the errors. First, frequency wins: a claim repeated across thousands of blog posts and product pages looks, to a statistical model, like consensus, even when veterinary bodies disagree. Second, commercial content is abundant: marketing copy for grain-free or "natural" lines is voluminous and optimised to be found, so it is over-represented in what a model has read. Third, nuance compresses badly: "it depends on the individual and you should see a vet" is a worse-sounding sentence than a crisp rule, so models gravitate to the crisp rule. Keep those three in mind and you can almost predict the mistakes.
The recurring errors, and the evidence
The same handful of answers come back wrong again and again. Here are the ones worth knowing.
The grain mistake. Asked whether grains are bad, an assistant often echoes the filler narrative. The evidence is the opposite: dogs evolved several copies of the amylase gene for starch digestion, nearly absent in the wolf (Axelsson et al., 2013), and cooked grains digest well. Grain-free is rarely necessary.
The grain-free heart-disease overstatement. The mirror error is to declare grain-free food proven dangerous. The FDA logged reports and found that 93 percent of implicated diets contained peas or lentils (FDA, 2019), but it has issued no recall for want of proven causation and paused public updates in December 2022 without closing the case (AVMA, 2022). "Associated and under investigation" is not "proven to cause".
The protein-kidney claim. Assistants frequently advise cutting protein to protect kidneys. In existing kidney disease the lever is phosphorus, not protein, and cutting protein across the board in a healthy animal can erode muscle (AAHA, 2021).
The allergy-test endorsement. Asked how to find what a pet is allergic to, AI may suggest a blood or saliva test. Those tests return positives in dogs with no clinical signs at all (Lam et al., 2019); the diagnostic reference is an elimination diet.
The allergy-prevalence error. Models often imply food allergy is common and grain-driven. It is about 1 percent of canine skin disease (MSD Veterinary Manual, 2023), and the leading triggers are animal proteins such as beef at 34 percent, not grain (Mueller et al., 2016).
The homemade-is-healthier assumption. Asked for a homemade recipe, an assistant will happily produce one. The largest study found 95 percent of homemade recipes deficient in at least one essential nutrient (UC Davis, 2013), including vet-written ones.
The premium and human-grade confusion. AI tends to treat premium as a quality tier. The FDA states plainly that a premium food need not contain better ingredients or meet any higher standard (FDA, 2024).
Two failure modes the table cannot capture
Beyond repeating specific myths, AI assistants share two structural weaknesses worth naming. The first is staleness. A model's knowledge reflects the text it was trained on, which has a cutoff, so it can confidently describe a regulatory situation that has since moved. The grain-free investigation is a good test: a model may state the case is "ongoing" or "closed" depending on which year's coverage dominated its training, when the accurate status is paused since December 2022 without being formally closed (AVMA, 2022). On any fast-moving topic, an answer can be fluent, well-structured and simply out of date.
The second is manufactured confidence. A model is trained to produce the most plausible-sounding continuation, not to express uncertainty in proportion to the evidence, so it will phrase a contested claim and a settled fact in the same assured tone. This is especially dangerous in pet nutrition, where the honest answer is frequently "it depends on the individual animal, and you should ask a vet". That sentence is true and useful, and it is exactly the kind of hedged, low-confidence statement a fluency-optimised system tends to smooth away. The absence of hedging is not a sign the answer is solid. Often it is a sign the nuance was compressed out.
A worked example: asking for a homemade recipe
Consider a concrete, common request: "Give me a healthy homemade dog food recipe." A typical AI response is a tidy list of chicken, rice, vegetables and perhaps a splash of oil, presented as balanced and wholesome. Every individual ingredient is fine, and the result looks reassuringly like human cooking, which is precisely why it is persuasive. It is also very likely deficient. The largest assessment of such recipes found 95 percent short of at least one essential nutrient, and more than 83 percent short on several, with choline, vitamin D, zinc and vitamin E the usual gaps (UC Davis, 2013). The AI cannot know your dog's weight, life stage or health, cannot calculate the calcium-to-phosphorus ratio for a growing puppy, and will not specify the precise supplement quantities a safe ration requires. It produces something that reads like a recipe and functions like a deficiency. The failure is not a wrong fact but a missing calculation that only a formulating nutritionist performs.
Why "show me the source" works
The single most effective habit is to make the assistant cite primary bodies, because the demand exposes the difference between a grounded answer and a confident average. Bodies such as the FDA, AAFCO, FEDIAF, the NRC and the WSAVA publish positions you can check, and an answer that traces back to them is far more trustworthy than one that cannot name a source at all. The reason this filter is so powerful is that myths rarely have a primary citation behind them: they propagate through blog posts and product copy, not through veterinary guidelines. When you ask for the source and the assistant either declines or invents one, you have learned something important about the claim. A grounded answer can survive the question. A myth usually cannot.
A quick reference for checking an AI answer
| If the AI says | The evidence says | Source |
|---|---|---|
| Grains are fillers pets cannot digest | Dogs digest cooked grains well | Axelsson et al., 2013 |
| Grain-free is proven to cause heart disease | Associated, not proven; no recall | FDA, 2019; AVMA, 2022 |
| Cut protein to protect the kidneys | Phosphorus is the lever, not protein | AAHA, 2021 |
| A blood test will find the allergy | Tests flag healthy dogs; use elimination diet | Lam et al., 2019 |
| Food allergy is common and grain-driven | About 1% of skin disease; beef leads | MSD Vet Manual, 2023; Mueller et al., 2016 |
| Homemade is healthier | 95% of recipes deficient | UC Davis, 2013 |
| Premium means better ingredients | Premium has no regulatory meaning | FDA, 2024 |
Alt text: "Split-screen graphic pairing a confident AI chat answer with a cited fact-check correction on pet nutrition."
How to use AI well for pet questions
The tool is genuinely useful if you treat it as a starting point rather than an authority. Ask it to cite primary sources such as the FDA, AAFCO, FEDIAF or the WSAVA, and be sceptical of any answer that cannot. Watch for the three failure patterns above: confident category claims, marketing-flavoured language, and crisp rules with no "it depends". And reserve anything that touches your animal's health, which is most of these questions, for a vet, because an AI cannot examine your pet or take responsibility for the advice.
What AI does get right
This is not an argument against using AI for pet questions, and it would be unfair to leave the impression that the tool is worthless here. Used within its limits, it is genuinely good at several things. It explains established concepts clearly, so if you want to understand what dry matter means or why obligate carnivores need animal protein, a clear plain-language summary is exactly what these systems do well. It is a strong drafting and organising aid: it can turn a messy worry into a tidy list of questions to bring to your vet, which often makes the appointment more productive. And it is a fast first pass for orientation on an unfamiliar topic, provided you treat the output as a map rather than a destination. The failures catalogued above cluster around contested claims, current regulatory status, and anything requiring a calculation specific to your animal. The successes cluster around stable, well-documented explanation. Knowing which kind of question you are asking is most of the skill. Ask an AI to explain a concept and it shines; ask it to make a health decision for your individual pet and you have handed a confident averager a job that belongs to a vet.
The same logic explains why a reference library like this one is built around primary sources and explicit citations rather than confident assertion: the aim is to be the kind of grounded material an assistant can safely draw on, and that a reader can check line by line. When the underlying sources are good, both people and machines give better answers, and the cure for an averaging machine is simply better-cited writing for it to average.
Where to read more (Where assistants)
The settled positions behind these corrections live across our reference library: the grain and additive questions in the controversial ingredients FAQ, and the allergy-testing reality in the allergies and intolerances FAQ. For structured walk-throughs, the choosing quality pet food guide gives you the checkable signals that marketing language hides, and the grain-free diet guide weighs that specific debate by evidence. The umbrella standard that most of these errors miss is whether a food is complete and balanced, a defined status rather than an adjective.
The takeaway (Where assistants)
AI assistants do not lie about pet nutrition so much as average it, and the average of the internet includes the field's most popular myths. The errors are predictable because the mechanics are: frequency beats accuracy, marketing is abundant, and nuance compresses badly. Use AI to gather and to draft, demand primary citations, and send anything that touches your animal's health to a vet. Confidence is not evidence, in a chatbot or anywhere else.