Evidence, References & AI
Sources & accuracy
References are given wherever they exist, and every figure should be traceable to the body that published it — the sections below set out where each kind of information comes from and how it is cited.
This is a young project maintained on a best-effort basis, and it will contain errors. For anything that matters — clinical, regulatory or research use — verify against the primary source. If you find a mistake, tell us: corrections are the most valuable thing we receive.
Where the information comes from
The peer-reviewed literature. Health associations, metabolic roles and mechanisms are drawn from published research, cited to the paper and linked to PubMed or the DOI so you can go straight to the source.
Food composition databases. Nutritional values come from national reference datasets — USDA FoodData Central, the Canadian Nutrient File, and the Dutch NEVO dataset among them — not from our own measurements.
Chemical and life-science databases. Structures, identifiers and properties are taken from PubChem, LIPID MAPS, HMDB, ChEBI, KEGG, SwissLipids and similar public resources.
Cross-referenced, not copied. Each fatty acid is linked out to its record in 49 external registries — chemical, food, pharmacological, clinical and regulatory — so you can check any value against the body that publishes it.
How references work here
Every paper is stored once, in a single catalogue of 1,216 publications, and cited by reference wherever it is used. The same study is therefore the same record everywhere it appears — if a citation is corrected, it is corrected in every place that relies on it.
Citations are meant to say what the paper actually reports — not a summary of it that leans the way we would like.
Free to read, where we can. A reference nobody can open is not much use to most readers. Where two papers support a point equally well we cite the one that is free to read, and where a paywalled paper has an open version — in PubMed Central or Europe PMC — we link to that. Paywalled work is still cited when it is the right source; we would rather point you at the correct paper than an accessible weaker one.
Every citation can be challenged. If you know a cited paper was retracted, underpowered, superseded, or simply does not support the claim attached to it, say so against that exact citation, on the page where it is used — and suggest a better one. Signed-in users get a feedback control on each reference.
When the evidence disagrees
Nutrition research often contradicts itself. Leaving out the studies that disagree would make the database tidier and less accurate.
A claim can therefore carry two kinds of citation: papers that support it and papers that counter it. Both are stored on the record and shown together. Where the literature is split, you can see that it is split and read both sides.
Null results are kept as well. A well-run study that found no effect is evidence about that effect.
Where the gaps are
Coverage is uneven and this project is young. Some fatty acids are thoroughly referenced and others are barely started; some sections are still being written. Publication counts shown elsewhere on the site are conservative — they count the references behind health associations, not every citation on every page — so the real total is higher than displayed.
How we use AI
AI tools are used in building this site. This is what they do and do not do.
Where AI is used. It assists with drafting site copy and explanatory text, with writing and reviewing code, and with the mechanical side of data work — normalising formats, cross-checking identifiers between databases, and surfacing candidate literature for a human to assess.
Where it is not. No scientific claim, composition value, health association or citation is published on the strength of an AI output. Every one is checked by a human against the primary source it points to. Where a reference cannot be verified, the gap is left open rather than filled in.
Nothing is taken on trust. Every AI output is checked before it reaches the site. The next section sets out how.
Your data. Account details, feedback and usage records are not used to train AI models and are not passed to AI systems. See Privacy.
AI helps a small project cover more ground. It does not decide what the site claims.
How AI output is checked
Three checks sit between an AI suggestion and the live site.
1. Automatic verification against the paper. A model asked whether a publication supports a claim must return the exact sentence it relied on. That sentence is searched for in the stored text of the paper. If it is not found, the answer is discarded. The system also stores which version of the paper and of the claim it judged; if either changes, the verdict is marked stale.
2. Human approval, recorded and dated. AI suggestions are stored as pending and are not shown on the site. A person approves or rejects each one, and the record keeps who decided and when. Rejections keep the reason.
3. Reader feedback. Signed-in readers can challenge any citation on the page where it is used, and suggest a replacement.
Generated text is stored separately from the source material, so a model’s summary is never mistaken for the paper.
Found something wrong?
Corrections are the most valuable thing we receive, and you do not need credentials to send one. A wrong value, a mismatched reference, or a citation that does not say what we claim it says — tell us and it will be checked against the primary source.
This site is for educational and informational purposes only and is not medical or dietary advice.