How we check
Every claim on this site names its source. If we cannot confirm something from a named source, we leave it out.
The research
We search PubMed for papers on each ingredient in food: the European Food Safety Authority's assessment where there is one, then reviews and studies in people, up to fifteen per ingredient. An AI model writes one plain sentence per paper from its abstract. A second model, from a different company, checks that the sentence says only what the abstract says. A sentence that gives a verdict, or fails a check, is not shown. Findings from animal or lab studies say so.
Most findings link to our own summary of the paper, written by an AI model from the full paper where we can read it free, and checked the same way. Title, authors, journal and funding come from PubMed, not from the model. We read a full paper only from a free, lawful copy, and quote a paper only when its licence allows any use; that licence is named on the summary.
The signal
The word at the top of a page follows fixed rules from the findings: how many found harm, how many found none, and whether they were in people or only in animals and cells. Findings in people always outrank findings in animals. Something found in a food, a single overdose, an allergy, or an expert panel's view is shown but not counted. The signal describes the research on the page, not how much of the ingredient you eat.
- Concerning, in studies of people. At least one finding of harm in people, and none in people that found no harm or a benefit.
- Mostly concerning, in studies of people. More than two thirds of the findings in people point to harm. The rest found no harm or a benefit, and they are shown too.
- Concerning, in animal and lab studies. Harm was found only in animal or lab studies. Findings in people, where they exist, outrank these.
- Conflicting. Findings in people point both ways: as many show harm as show none, or at least two show harm and each side has at least a third.
- Mostly no concerns. Most findings in people found no harm or a benefit, and under a third of them, or just one, point to harm. Those are shown too.
- No concerns. At least one finding in people, and none of them found harm.
- Unclear. The findings in people only say the question is not settled.
- Not enough. No finding we could check yet. This says nothing about the ingredient, only about what we have read so far.
The regulators
Code reads each regulator's own list and writes the status line. No AI model writes these. The lists are the FDA's rules in Title 21, the EU's list of authorised additives in Regulation 1333/2008, the Food Standards Agency's list for Great Britain, and Health Canada's Lists of Permitted Food Additives. Where an ingredient is not on a list, we show nothing for that place rather than guess.
Sources and licences
Paper records come from PubMed through the National Library of Medicine's E-utilities (NCBI disclaimer and copyright notice). Free full texts are found through OpenAlex, Semantic Scholar, Europe PMC and Crossref. Opening lines draw on English Wikipedia, under CC BY-SA 4.0.
If a page uses your work in a way you object to, tell us through Report an error. We reply within 48 hours and take the passage down while we look.
The products
Product counts and examples come from Open Food Facts, for products sold in the US, Canada and the UK, used under the Open Database Licence.
Skin pages
Some ingredients also go into creams, sunscreens, makeup and soaps. Their skin pages are built the same way, from research on skin use. We have not checked cosmetic regulations yet.