Integrating hazard into the foodome: a critical review of AI pipelines for chemical, nutritional and safety characterization of foods, with a FAIR audit of data infrastructures

The AI-foodome, the machine-readable integration of food chemistry, nutrition and safety through artificial intelligence, is the declared objective of a fast-growing literature, yet the data infrastructures invoked as its foundation have not, to our knowledge, been tested at the architectural level. A systematic search of the reproducible corpus assembled for this review (3205 records, 2015–April 2026) identified no prior per-indicator FAIR audit of these infrastructures; the audit reported here evaluates ten infrastructures with three standardised instruments allocated by technical applicability (F-UJI, O'FAIRe and the RDA FAIR Data Maturity Model). They are FAIR-by-narrative, not FAIR-by-design: across the seven infrastructures assessed on a common indicator scheme, Accessibility reaches 86.9% compliance while Interoperability falls to 33.3%, a gap that tracks the technical substrate rather than implementation choices. Declared compliance does not correspond to technical exemplification: no typed inter-axis link was traversable among the audited infrastructures, and a single food entity, wheat, could not be reconstructed across axes. The fragmentation is read here as structural: at the observed rate of cross-axis convergence, reaching 50% multi-axis integration would require approximately 73 years (95% CI: 53–198). The audit locates the integration deficit in the data architecture rather than in the algorithms: the constraints documented here are located below the level at which model refinement operates. From this diagnosis the review derives a four-pillar framework (semantic unification, FAIR-by-design infrastructures, cross-axis provenance traceability and inclusive contribution) specifying the conditions under which agencies and infrastructure operators can query composition, intake and hazard data as a single evidence base.