Can an AI Chatbot Write a Balanced Dog Food Recipe? What a 2026 Test Found

In December 2025, Brazilian researchers asked four widely used, free AI chatbots for a “complete and balanced” homemade diet for a healthy small dog. Of the 11 recipes they got back, 10 were short on calcium and 9 were short on energy (Rosa et al., Veterinary and Animal Science 2026).

While chatbots improve quickly and this was a paper exercise, no dog was fed, the results match what studies of human-written recipes have found for years. A recipe that says “complete and balanced” isn’t proof that it is.

→ Why Most Homemade Dog Food Isn’t Complete, And What the Studies Actually Found

---

1. The study

The question. Each chatbot got the same single prompt, in Portuguese, asking for one beef and one chicken diet for an 8 kg adult dog with little exercise. The researchers didn’t ask follow-up questions, to mimic how an owner might use them.

The recipes. Some chatbots offered alternative ingredients, so the four tools produced 11 recipes in total.

The analysis. The researchers entered each recipe, unchanged, into diet analysis software and compared it with FEDIAF, the European pet food industry’s nutrient guidelines.

Funding. A Brazilian government research agency funded the work. The authors declare no competing interests.

---

2. What they found

Problem	Recipes affected
Calcium below the minimum	10 of 11
Too few calories for the dog	9 of 11
Calcium-to-phosphorus ratio outside the range	7 of 11
Vitamin D below the minimum	7 of 11
Any vitamin-mineral supplement included	2 of 11, in amounts the authors call insufficient

What they got right. Every recipe met the minimums for protein and fat. Amino acids were mostly adequate.

Raw food. One chatbot was the only one to suggest raw ingredients.

→ The Calcium Problem: Why Meat-and-Rice Diets Break Bones

Our note: No dog was fed these recipes. The paper describes risks from long-term feeding.

---

3. Why the chatbots struggled

Words, not sums. The authors argue the chatbots know the principles, such as the calcium-to-phosphorus ratio, but can’t turn them into accurate quantities.

Confident formatting. The authors worry that owners may trust these plans because they look “well-structured and apparently authoritative”.

Follow-up prompts. Asking again might help. But the authors point out that this needs an owner who can already spot what’s missing.

Our note: The authors suggest a chatbot could produce drafts for a professional to check.

→ How to Get a Properly Formulated Recipe (and What It Costs)

---

4. What the study can’t tell you

Newer versions. The authors say their results “may not reflect the capabilities of future model versions.”

Software estimates. Nutrients were estimated from a food database, not measured in cooked food.

One dog profile. Only a healthy adult was tested. The authors expect the risks to be greater for dogs with medical conditions.

A puzzling zero. Every recipe scored zero for chloride. None of the recipes may have included salt, or the database may lack chloride values. The paper doesn’t say which.

Small slips. The software is listed as version 1.3 in the methods and version 8.0 in the references.

Our note: The zero chloride result may be a database gap rather than a recipe failure. We can’t tell.

---

5. What we could not verify

Commonly said	What we found
“AI can write a complete dog food recipe”	Not in this test. None of the 11 recipes met every guideline.
“Newer chatbots fix this”	Not tested. The study used free versions from December 2025.
“A recipe labelled balanced is balanced”	No. Every recipe was requested as complete and balanced.

---

Sources

• Rosa MBS, França J, Siqueira MTS, Raineri C. Use of Artificial Intelligence in obtaining canine diets: nutritional inadequacy and the need for technical knowledge. Vet Anim Sci 2026: in silico analysis of 11 chatbot recipes; no animals fed. Funded by a Brazilian government research agency

Always confirm with your vet before acting on anything on this page.