What a honey authenticity database is

What the instrument actually tells you

Authenticity analysis divides into two kinds. Targeted analysis looks for something specific — a named marker compound, a particular syrup residue — and knows in advance what it is asking. Untargeted analysis records a broad signal instead: a nuclear magnetic resonance spectrum, an infrared profile, a mass-spectrometric fingerprint. Nothing in that signal is labelled genuine or adulterated. It becomes an authenticity finding only when it is compared against a body of data from honeys believed to be real.

This is the step most reporting skips. A headline saying a laboratory found a honey to be adulterated is usually describing a sample whose profile sat outside the range a database holds for honeys of that description. That is a real and useful finding. It is not the same kind of statement as a measurement of added syrup, and it inherits every weakness of the collection it was compared against.

Why the transparency problem became a legal problem

The largest reference collections were built commercially, at considerable cost, and are not published. That is a defensible commercial position and it creates a specific difficulty: if the comparison set cannot be examined, neither the representativeness of the comparison nor the interpretation drawn from it can be independently checked. Findings resting on those databases have supported significant commercial and enforcement decisions, and their opacity has been identified as a barrier to their acceptance and a cause of legal dispute.

Defra, the Food Standards Agency, Food Standards Scotland and the Government Chemist have been working on the scientific side of this since 2018, following a Government Chemist seminar on honey authenticity in 2019. The route they took was not to build a public database or to reject proprietary ones. It was to write down how such a database should be interrogated when its conclusions are challenged, with safeguards for confidential data so that scrutiny does not require publication.

What a database is interrogated for

The framework published in November 2025 groups the questions under five headings. They are worth knowing even outside a dispute, because they are also the reasons a database-based result can be right, wrong, or simply not applicable to the sample in front of it.

1. 1ScopeWhat was this database built to answer, and was that decided before it was assembled or discovered afterwards? A collection begun to detect one kind of adulteration and later used for geographical origin is being asked a question it was not designed for, and stating what a database is not suitable for is treated as part of declaring its scope.
2. 2CompositionWhat is actually in it, and on what basis were samples admitted or excluded? A reference database is only as authentic as its criteria for deciding that a sample was genuine in the first place — a circularity that has to be broken by evidence outside the analysis, such as collection directly from a known beekeeper.
3. 3MetadataWhat is recorded alongside each sample: botanical source, geographical origin, bee species, year and season, processing, and the traceability of the sample itself. Without this the collection cannot be subset, and a sample can only be compared against the whole thing rather than against the honeys it actually claims to be.
4. 4RepresentativityDoes the collection capture the natural range of what it purports to cover, including seasons, years, regions and production practices? If it claims to represent honey globally, does it reflect where honey is actually produced? Is it curated and reviewed as the world changes?
5. 5ValidationWere the analytical methods properly validated before being used, and are they accredited to an international standard such as ISO/IEC 17025? A database is only as good as the measurements that populated it.

What a database match can and cannot show

Models, accuracy figures and the two ways a test fails

Most whole-profile methods end in a statistical model: a classifier trained on authentic samples, and sometimes on deliberately adulterated ones, that sorts a new sample into a class. A headline such as 'the model identified origin with 98% accuracy' describes that model on that study's samples. It does not say how the model performs on honeys from another region, another year, another season, another processing line or another instrument, and the published figure is often measured on samples drawn from the same collection the model was built on. External validation — testing on independent samples the model has never seen — is the step that most often is missing.

• A false positive: a genuine honey flagged as suspicious. Natural variation can resemble adulteration; an unusual floral source, a rare region or a honeydew honey can sit outside the reference set; and processing such as heating changes some markers.
• A false negative: an adulterated honey passed. A C3 syrup passes a C4-focused isotope screen; a syrup without the markers sought passes a marker screen; and a low adulteration fraction can sit below what a method detects.
• Both kinds shrink when independent methods are combined and when the reference set represents the honeys actually being tested. Neither disappears.

Where this is going

The framework is a method rather than a rule: it carries a disclaimer that it represents its authors' views and not government policy, and its own conclusion is that database scrutiny needs standardised approaches while avoiding a counsel of perfection no real database could meet. Its authors expect it to be generalised to authenticity databases for other foods.

Two things are worth watching. The first is whether a mechanism for confidential data transfer is established, which is what would let a database be examined without being published. The second is the European side: Directive (EU) 2024/1438 empowers the Commission to introduce harmonised methods for detecting sugar adulteration and a uniform methodology for tracing honey origin, and if those arrive they change what a reference collection has to be able to do.