← All food signals
Trust signal

The Fraud Testing Gap

Food regulators say their routine fraud tests miss adulterants in honey and olive oil. Labs are training machine-learning models to authenticate food, but none is in official use yet.

What is it?

Food fraud is proven in a lab, and regulators now admit their routine tests miss some adulterants. Government labs are training machine-learning authenticators to close the gap. None is yet in official control, so the AI in use still picks which shipment to test.

Why now?

In April 2026 the FDA said its honey test sees only a subset of added sweeteners. EU auditors found olive-oil labs skipping required parameters. NIST’s machine-learning authenticators remain in development, and UK advisers warn the same AI can forge the paperwork.

Who’s building it?

Government measurement labs build the authenticators; regulators decide when a trained model counts as proof.

Measurement institutesFood regulatorsReference laboratoriesAtomic-energy agenciesEU auditors
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — government labs training models to authenticate food chemistry.
Buyers — official control laboratories that need tests catching more adulterants than today’s methods.
Gatekeepers — food regulators deciding which method can stand as proof.
What’s at stake
Whether an official test result still means a product is genuine, once regulators admit it misses fraud.
Honest honey and olive oil undercut on price by fakes.
Certificates that AI can now forge as well.
What’s unsettled
Whether a machine-learning authenticator will ever be accepted as official proof.
Whether reference data can be shared widely enough to train models.
Whether regulators close the gap faster than fraudsters adopt the same AI tools.
02 · How the demand matureda progression, not a timeline
Admission
Regulators publish reports saying routine tests miss some honey and olive-oil fraud.
Research
Government labs train machine-learning models on isotope and trace-element data.
Targeting
The AI actually deployed picks which import to sample, and leaves the lab test itself unchanged.
Proof
Whether a trained model can stand as legal evidence decides when it reaches control.
03 · Where the market goesthe forks, not a prediction
→Lab or border: machine learning may move into official authentication, or stay confined to choosing which shipments get tested.
→Open-source authenticators may spread to every control lab, or stall on missing reference data.
→Regulators may simply add more parameters to existing laboratory tests instead of adopting trained models.
→Forgers using AI may widen the gap faster than labs close it.
Trained models can already sort genuine from suspect samples in research; the open question is whether a regulator will let one stand as proof in court.
04 · The Verification Stack deskadjacent markets, moving now
Proof Beneath the Fabric
Forensic tests are added where paperwork alone may not be enough.
Oritain says 90% of brands had a risk-consistent prohibited-cotton result in 2025.
Sourcing Journal (WWD)
The Toxic Counterfeit
Counterfeits can carry contamination that authentication must catch.
UK IPO testing found carcinogens and heavy metals in seized counterfeits.
UK Intellectual Property Office
The Surveilled Ingredient
AI flags which food chemicals get reopened.
FDA uses AI to identify safety signals on food chemicals already on the market.
FDA
05 · The Wire
01
FDA found 4 of 102 honey samples violative.
FDA ↗
02
EU auditors: only Spain analysed all 15 olive-oil parameters
European Court of Auditors ↗
03
NIST is building open-source machine-learning classifiers and authenticators to establish food provenance and authenticity.
NIST ↗
04
The UK Food Standards Agency’s science advisers say AI could both generate and detect fraud, and will attract criminals seeking to get round food controls.
GOV.UK (FSA Science Council) ↗
05
An EU coordinated action found 147 honey samples, 46%, suspicious of adulteration and said current methods are not developed enough.
EU Joint Research Centre ↗
06
FDA’s machine-learning model reads an imported seafood entry and returns a sample recommendation within seconds.
FDA ↗
The panel on this signal
They agreed on
–Regulators concede in their own published reports that routine honey and olive-oil tests miss some adulterants.
–Machine-learning authenticators are still inside government research labs.
–The AI regulators actually run chooses what to test, not what is inside.
The chemometrician disagreed on
–The bottleneck is reference data, not algorithms: a model trained on too few authentic samples learns the library’s gaps.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.