How we check

What an ingredient page shows

Four things, each from its own source: what the ingredient is and why it is used, what published research on it found, where regulators in four places stand, and which products list it. A part of a page with no data is left out, not filled in. A missing section is a fact about what we have read so far, not about the ingredient.

Where the research comes from

We search PubMed, the US National Library of Medicine's index of medical research, for papers with the ingredient's name in the title and a food or diet angle. For additives we look for, in order:

• the European Food Safety Authority's own safety assessment of the additive, where one exists;
• review papers and meta-analyses from the last twenty years;
• studies in people.

We read up to fifteen papers per additive. We do not search for single studies in animals, because a page built from those alone would lean towards alarm. Some single lab studies still come through, because PubMed files studies on human cells under human research. Animal and lab findings are labelled as such wherever they appear.

How a finding is written and checked

An AI language model writes one plain sentence per paper, from that paper's abstract. We do not trust it. Before any sentence appears on a page, it must pass every one of these checks:

1. A second AI model, from a different company, confirms the paper is about the ingredient as it is eaten in food. Papers about industrial uses, medicines or farm animals are dropped.
2. The sentence is short, plain, and names the ingredient.
3. It gives no verdict. A sentence that calls the ingredient safe or harmless is rejected. A finding may still report an official daily intake limit.
4. It says nothing about what is allowed or banned. Only code writes those sentences, from the regulators' own lists.
5. Every number in it appears in the abstract.
6. A finding from animal or lab studies says so.
7. The second model confirms the sentence says only what the abstract says.

A sentence that fails a check gets one rewrite, then must pass every check again. If it still fails, it is not shown. Each finding links to its paper, so you can read the source yourself.

Paper summaries

Most findings link to our own summary of the paper behind them. An AI language model from DeepSeek writes each summary from the paper's full text when it is free to read through PubMed Central, and from its abstract when it isn't. Each page says which.

The same kinds of check apply. Every number must appear in the paper. The site's own voice gives no verdicts and says nothing about what is allowed or banned. Anything the authors conclude or advise is credited to them. A second AI model, from Qwen, confirms that the one-line summary and each finding say only what the paper says. A summary that fails a check is rewritten once, and a paper whose summary still fails gets one fresh attempt. If that fails too, there is no summary page and the finding links to the paper instead.

The title, authors, journal, funding, conflicts of interest and where the authors work come from the paper's PubMed record, not from the model. No person has read every summary yet.

The signal

Two AI models, from different companies, each sort every finding by what it reports: harm, no harm, a benefit, or an open question. If they disagree, the finding counts as an open question. The same two models also say whether each finding comes from studies in people or only from animals and cells. A finding counts as evidence in people only when both say so. When they disagree, or the sentence does not say, the finding is shown on the page but not counted. The word at the top of the page follows fixed rules from those counts:

• Concerning, in studies of people. At least one finding of harm in people, and nothing in people that points the other way.
• Concerning, in animal and lab studies. Harm was found only in animal or lab studies. Findings in people, where they exist, outrank these.
• Conflicting. Some findings in people show harm, and others found none or left it open.
• No concerns. Two or more findings in people, and none of them found harm.
• Unclear. The findings in people only say the question is not settled.
• Not enough. Fewer than two findings we could check. This says nothing about the ingredient, only about what we have read so far.

Findings in people always outrank findings in animals. The signal describes the research on the page. It is not a safety rating, and it does not account for how much of an ingredient you eat.

What regulators say

Code reads each regulator's own list and writes the status sentence. No AI model writes these.

• United States: the FDA's rules in Title 21 of the Code of Federal Regulations, parts 73, 172, 182, 184 and 189.
• European Union: the list of authorised food additives in Regulation (EC) No 1333/2008.
• Great Britain: the Food Standards Agency's list of approved additives. Northern Ireland follows the EU list under the Windsor Framework.
• Canada: Health Canada's Lists of Permitted Food Additives.

Where we could not find an ingredient in a list, we show nothing for that place rather than guess.

Products

Product counts and examples come from Open Food Facts, an open database of food labels, for products sold in the US, Canada and the UK. It is used under the Open Database Licence. Open Food Labels is independent and not affiliated with Open Food Facts.

What we have not done yet

• No registered dietitian has reviewed these pages yet.
• Findings are written from abstracts, not full papers.
• Many common additives have little published research in people. Their pages say "Not enough" rather than filling the gap.
• The signal counts findings. It does not weigh a large trial above a small one.

This site is a reference, not medical advice. If you have a condition, an allergy or a diet to follow, ask your doctor or a dietitian.