How this site is made

Food Additive Index is built and written with an AI model, Anthropic's Claude, working to instructions and rules set by the person who runs the site. This page says which parts the AI writes, which parts it does not, and what has and has not been checked. The wider method is on Method.

What scripts read from sources, with no AI in the loop

The legal facts on the 1189 additive pages are not written by anyone. Scripts fetch each regulator's published list — the eCFR, EUR-Lex, legislation.gov.uk, Health Canada, the Federal Register of Legislation, e-Gov, South Korea's notification, the Codex and JECFA databases, EFSA and the FDA — and store, for each row, the source, the provision and the date it was retrieved. A page prints those rows and links the instrument. The status of an additive in a country, the maximum level in a food category, an acceptable daily intake and the identifiers (E, INS, CAS) are all of this kind.

A change to what a source says is detected on the next weekly run and recorded with the date. A change that looks too large, or that the source does not explain, is held back and not published until it has been looked at and released or discarded with a note. The structure diagrams are PubChem's records; the tables are generated from the data. Nothing on the site is an AI-generated image.

What the AI drafts

• The short summary on each additive page, and the description of what it ismade from, with the dietary flags that follow from that description.
• The explanations on the guide pages — this one, Method, theglossary, the labelling notes and the note on eachjurisdiction saying what its regulator publishes.
• The dated entries on Changes and the summaries of studies and regulators' readings of them, each written from the instrument or paper it cites.
• The code, the importers and the tests that produce everything above.

What the AI does not do

• It does not decide whether an additive is permitted anywhere, or at what level. Those come from the rows above.
• It does not rate, rank or call any additive safe or unsafe, and does not give health advice. Where regulators disagree, both readings are shown, attributed.
• It does not treat “not listed” as “prohibited”. They are separate values in the data.
• It does not translate: names in other languages are the names the Official Journal prints.
• It is not given a name, a face or credentials. This site has no named author, and none is implied.

How it is checked

• Every fact in a table carries its source and retrieval date, so a reader can check it against the instrument without trusting this site.
• The prose is drafted to fixed rules — what a substance is, what it is made from, where it is used; no verdicts, no advice — and was compared with the status, identifiers and functions each page holds. A pass in September 2026 removed statements the data did not support and corrected chemistry errors.
• The importers, the matching of a regulator's name to a page and the change detector have automated tests, and a check of the built site flags a page with a missing or duplicate title, no structured data, or no canonical address.
• Two pages are merged, and the old address redirects to the one that remains, when they turn out to be one substance.

What has not been checked

No error rate has been measured for the prose, and no second reader, whether a toxicologist, a food lawyer or a regulatory specialist, is claimed for any page. The prose is short, and where it is wrong the table beneath it, with its citations, is the authority. Nothing here is legal, medical or regulatory advice. Before relying on an entry for a decision, read the instrument it links.

The pages were added in batches as each source was matched, not all at once, and a page is created only when a regulator names the substance and there is a source to cite for it.