How to Adapt Recipes for Food Allergies With AI

A language model is good at the pattern-matching side of allergy cooking. It has absorbed many substitution discussions, so it can tell you that a flax egg (one tablespoon of ground flaxseed mixed with about three tablespoons of water) binds a muffin batter but will not whip into a meringue, while aquafaba, the liquid from canned chickpeas, can. It also knows common hidden sources: whey and casein in some products marketed as non-dairy, wheat in regular soy sauce, barley in malt vinegar and malt flavoring, anchovies in Worcestershire sauce, and egg in some fresh pastas and glazes. What it cannot do is see the package in your hand. Formulations change, the same product can differ between countries, and a model's training data may be years old. In the United States, the Food Allergen Labeling and Consumer Protection Act requires major allergens to be declared, and the FASTER Act added sesame as the ninth major allergen, with labeling required from January 1, 2023. The European Union requires 14 allergens to be emphasized on labels. But precautionary statements such as "may contain" or "made in a facility with" are voluntary in the US, so a label without one is not proof that a product is safe. Cross-contact is the other gap. An allergen can move from a shared toaster, fryer oil, cutting board or bulk bin into food that contains no allergenic ingredient at all. AI can remind you of these risks, but only you can control your kitchen. One common misconception is that "gluten-free" and "wheat-free" mean the same thing. A wheat-free product can still contain barley or rye. Another is that a substitute that tastes right is automatically safe. The substitute's own ingredient list still has to be checked.