AI in Food Labelling: What Is Actually Real
A wave of funded startups promises automated nutrition panels and allergen compliance. Some of that is real and useful; some of it is a demo. This separates the two, and explains why the last step of a label stays human whatever the marketing says.
What AI genuinely automates well
The pattern: AI is excellent at getting data in and at flagging anomalies. Both are genuine time savings, and for a small producer transcribing supplier specs is hours a month.
What is being oversold
The claim to treat carefully is end-to-end label generation from a photograph or a recipe description, with no verified ingredient data underneath. A model can produce a plausible ingredient list. Plausible is not the standard — a label has to be traceable to what you actually bought.
- Allergen determination without supplier specs. The model knows what hummus usually contains, not what your supplier’s reformulated batch contains.
- Nutrition panels from recipe text alone. Composition varies by product and cooking loss is process-specific.
- Regulatory interpretation as a decision. Useful for orientation; not a compliance position.
- “Fully automated compliance.” Compliance is a responsibility, and responsibility does not transfer to a vendor.
Why the last mile stays human
UK and US food law both put responsibility on the food business operator. No software vendor assumes liability for your label, and none of the funded startups do either — read their terms. That is not a gap in the technology, it is the structure of the regulation.
Practically, it means the useful design is AI that accelerates data capture and checks for omissions, with a human confirming the record and the software deriving the label from confirmed data. FoodCore is built that way deliberately, and we say so on the AI recipe tools page.
What to ask a vendor
- Where does allergen data come from — a model, or my supplier specifications?
- Is the label derived from stored records, or generated as text each time?
- Can I see and edit the ingredient record the label came from?
- What is sent to the AI provider, and what is not?
- Who is liable if the label is wrong? (The answer is always you — check they say so plainly.)
Frequently asked questions
Can AI generate an FDA nutrition facts panel? It can assist with data capture and formatting. The underlying composition data has to be sourced properly, and the responsibility stays with the business. See creating an FDA-compliant label.
Is AI labelling software worth it for a small producer? The data-capture part, often yes — transcribing specs is genuinely slow. The decision-making part, no.