AI is giving pet nutrition advice — who’s checking the nutritionist?

Is having access to basically unlimited information the same thing as having expertise? Simply put, no.

Good nutritionists don’t simply look up nutritional requirements or pull information from peer-reviewed studies and textbooks. They evaluate and reconcile conflicting research, understand how ingredients interact, account for processing and sourcing, and apply that information to a specific food, animal and set of circumstances. That’s where critical thinking, experience and judgment come in. And when I say nutritionist, I’m talking about someone who is properly trained and has actual expertise — not a self-proclaimed nutrition expert.

To be clear, this isn’t an article about nutritionists versus AI. AI has enormous potential to help nutritionists work faster, more effectively and more efficiently. I welcome the ways AI can help us review literature, find relevant research, analyze large amounts of data and identify patterns that might otherwise take hours to uncover.

But access to more information doesn’t eliminate the need for someone who knows what to question, what may be missing and when an answer simply doesn’t make sense.

Google with a Ph.D.? Not quite

Can anyone formulate dog or cat food? Sure.

But formulation isn’t simply plugging ingredients into a software program until the numbers turn green. Can the person — or AI — look at conflicting peer-reviewed research and determine what actually applies? Do they understand the difference between a nutrient requirement, a recommendation and an optimal intake? Do they understand the difference between body weight and metabolic body weight? Do they know the potential anti-nutritional factors that can derail an otherwise complete and balanced food?

Can they formulate a food to be 100% complete and balanced regardless of the ingredients being used? Do they understand shelf life and what can happen to nutrients over time?

Then there are the practical questions. Is the recipe designed to feed one dog or 1 million dogs? How will manufacturing and processing affect it? Where are the ingredients coming from? Lamb liver in the U.S., for example, will be higher in selenium because of supplementation compared with lamb liver from New Zealand. And do they know when a nutrition question becomes a medical question?

Those aren’t simply questions of information access. They’re questions of expertise.

AI doesn't know what it doesn't know

One of the most important lessons I learned over my career came from my graduate-school advisor. It was three simple words: I don't know.

When you don’t know something, you go out and learn it or surround yourself with people who have the expertise you don’t. Knowing when you don’t know is part of expertise. And that’s where I think we need to be especially careful with AI.

The risk isn’t simply that AI will give the wrong answer. Wrong answers can sometimes be easy to spot. The harder problem is an answer that sounds right, is supported by seemingly credible information and is delivered with enough confidence that no one thinks to question it.

An experienced nutritionist has the benefit of knowing when something doesn’t pass the smell test. Experience creates that moment of hesitation: “That doesn’t sound right. What am I missing?”

But what happens when the person relying on the answer doesn’t have enough experience to have that reaction? Does AI know when the right answer is, “I don't know”? Does it recognize when there isn’t enough information to reach a conclusion? Or does it simply produce the best answer it can from the information available?

And if the person receiving the answer doesn’t know enough to question it, who catches what is missing?

The data says what? Challenge it

Knowing enough to question the answer also means knowing enough to question the information behind it. AI is already very good at finding information, and that capability will only improve. But finding a peer-reviewed study isn’t the same thing as understanding what the study means — or whether the data actually support the conclusion being drawn from it.

I've always believed you must challenge the data, whether the conclusion comes from a person or AI. Just because something is published, widely accepted or presented with confidence doesn’t mean we should take it at face value. It reminds me of the saying, “In God we trust; all others bring data.”

Nutritionists don’t simply count studies supporting one conclusion versus another. We have to ask questions. How was the research conducted? What population was studied? What was actually measured? Are the results relevant to the species, ingredient or formulation at issue? Are there limitations in the data? And does the conclusion make sense when considered against the broader body of evidence?

Two studies can look at similar questions and reach different conclusions without either one necessarily being “wrong.” The job isn’t simply to find the study that gives you an answer you may want. It’s to understand the data, challenge it and determine whether you would reach the same conclusion.

AI may dramatically improve our ability to find and analyze evidence. That’s a good thing. But it shouldn’t stop us from challenging the data. If anything, it gives us more data to challenge.

When the math says the food is fine

And this isn’t just an academic exercise. Case in point: There was a recent recall tied to a particular freeze-dried and raw cat food. Assuming there was no heat-kill step, and all ingredients were brought in raw, would AI have picked up on the fact that the food was deficient in thiamine?

Not likely.

On paper, AI may have concluded that thiamine was sufficient. But the formulation also included mussels, which contain thiaminase. Would AI have recognized that interaction and accounted for it?

Thiaminase is an anti-nutritional factor that needs to be heat inactivated. Otherwise, it can degrade the thiamine present in the food — particularly in foods that do not go through a heat-kill step. The same issue can arise with a potentially biotin-deficient food if raw eggs are used without a heat step to inactivate the avidin.

This is where nutritional expertise matters. The numbers may say the nutrient is there. The harder question is whether something else in the formulation or process changes what those numbers mean in the finished food — and, more importantly, what they mean for the animal consuming it.

A formula is not finished when the spreadsheet says it is.

A pet food doesn’t exist only at the moment someone formulates it. It must be manufactured, packaged, shipped, stored and ultimately fed to an animal — potentially months after it was produced. Ingredient variability matters. Processing matters. Nutrient degradation matters. Shelf life matters.

A formula can work perfectly on paper on day one, but someone still needs to understand what happens to that food through manufacturing and over the course of its intended shelf life.

AI can be incredibly useful in evaluating those variables. But someone needs to know which variables matter in the first place.

The superfood trap

And the same need for judgment applies when we’re evaluating potential benefits, not just nutritional adequacy. I would be remiss if I didn’t discuss ingredients and their potential benefits.

With the recent approval of pomegranate pomace at the American Association of Feed Control Officials’ (AAFCO) 2026 annual meeting, many companies may be rushing to incorporate it into pet foods and talk about its polyphenol and antioxidant benefits.

Run an AI search and pomegranate pomace may quickly start looking like the next go-to superfood. But is, “Does pomegranate pomace have benefits?” really the question we should be asking? Or should the question be, “Is pomegranate pomace the best ingredient for the benefit we’re trying to deliver?”

Those are vastly different questions.

Do people know that tomato pomace has similar — and in many cases greater — amounts of these compounds compared with pomegranate pomace? What about cost? Availability? Sourcing? Tomato pomace is more cost-effective and produced in the U.S., while pomegranate pomace may be sourced from Central Asia and Iran. Now that you have this information, is pomegranate pomace still the ingredient you choose?

Identifying a potential benefit is one step in ingredient selection. Determining whether that ingredient makes sense in a particular product is something else entirely.

The newest or most interesting ingredient isn’t necessarily the best ingredient from either a supply or nutrition standpoint.

AI can recommend it. Who owns the decision?

All of this leads to one more question: Who is ultimately responsible for the decision?

AI can generate an answer. It can summarize research. It may eventually become extraordinarily good at evaluating formulations.

But someone still must make the decision.

Someone must decide whether the evidence is sufficient, whether the formulation is appropriate, whether the risks have been adequately considered and whether the finished product is something they are willing to stand behind.

AI can recommend a formulation. It can’t take responsibility for one.

AI still needs a nutritionist

AI is going to change what nutritionists do. I think that’s a good thing.

Used well, it can help us find information we might otherwise miss and give us more time to focus on the questions that require judgment. It may become one of the most valuable tools available to nutritionists.

But AI doesn’t eliminate the need for nutritional expertise. If anything, it may make expertise more important.

The real risk isn’t that AI will always give the wrong answer. Sometimes it will give an answer that looks completely right. The question is whether someone with enough expertise will be there to recognize when it isn’t. AI can provide information. It can identify patterns. It can even recommend a formulation. But someone still must know what questions to ask, when to challenge the answer — and when to say, “I don't know.”