Transparency as a condition, not an afterthought

HOLiFOOD's recommendations emphasize that any AI tool used in food safety must be transparent, explainable and built on well-documented, representative data. Researchers, authorities and other users need to understand how predictions are produced, assess their limitations and retain responsibility for interpreting the results.

"Better data and new technologies can strengthen food safety, but they are only part of the answer. We also need expert interpretation, effective communication with citizens and policy decisions that consider consequences across the whole food system," says Jeanne-Marie Membré, senior research scientist at Inrae and one of HOLiFOOD's project partners.

New detection methods stuck at the door

Alongside the database work, HOLiFOOD investigated technologies that could complement established food safety controls. These include metagenomics (sequencing all genetic material in a sample to identify microbes), portable PCR systems that can detect pathogens on-site, biosensors, spectroscopy and AI-supported chemical analysis.

These technologies show potential, but wider implementation will require further validation and clear regulatory pathways before they reach routine use. The project therefore calls for harmonized validation frameworks, official guidance for AI-supported analytical methods and interoperable data infrastructures that allow information to be compared and shared across countries.

When recognizing a risk isn't enough

The project also examined the human side of risk. HOLiFOOD research indicates that risks tied to long-term drivers such as climate change can feel psychologically distant: people may acknowledge a threat exists without understanding how it could affect them personally or what they can realistically do about it.

The recommendation that follows is practical. Emerging risks should be communicated through locally and personally relevant examples. Citizen observations and social media analysis can provide early signals of emerging events and public concerns, but both come with important limitations. Citizen-generated reports need verification and connection to institutional surveillance and response systems, and social media discussions should not be treated as representative of the wider population.

Food safety decisions ripple outward

The fourth of HOLiFOOD's policy briefs addresses trade-offs that arise when food safety, nutrition, environmental sustainability, food security and affordability are considered in isolation from one another. The tensions are concrete. More intensive thermal processing, for instance, may reduce microbiological risks while increasing energy consumption or contributing to the formation of process-related contaminants.

The project recommends involving policymakers, risk assessors and experts from different disciplines at an early stage, so that potential consequences across the wider food system can be identified and evaluated transparently before decisions are locked in.

The bigger picture

Together, the database and the four policy briefs sketch a sober vision of AI's role in food safety: useful, promising, but bounded. Machine learning models trained on hundreds of millions of records can flag where risks may be drifting. New detection technologies can catch hazards faster. Neither, in HOLiFOOD's framing, substitutes for expert judgment, honest communication with the public or policymaking that accounts for the food system as a whole.

The 392 million records in CHEFS are now publicly available. What researchers build with them — and how carefully authorities interpret those tools — will determine whether the database becomes a genuine instrument for anticipating food safety problems rather than simply a very large archive of past ones.