Google FINDER flags restaurants 3.1 times as likely to be unsafe

AI system

System: FINDER (machine-learned foodborne illness detection model)
Model: Not assessed
Release: Not assessed
Application developer: Google
Role: Not assessed
Configuration: Not assessed
Inputs and outputs: Not assessed
Autonomy: Not assessed
Workflow: Aggregated web search and location logs covering a city's restaurant patrons. Health departments able to act on a daily list. Inspection outcome records to compare against routine and complaint-based inspections.
Human oversight: Not assessed
Relation to AI: FINDER generated the daily restaurant lists. Inspectors did not know the source of the prompt. The comparison is observational (FINDER-prompted versus other inspections during the same period), not randomized. Google employees conceived and implemented the model and co-authored the peer-reviewed evaluation. Health department staff oversaw implementation.
AI contribution: FINDER generated the daily restaurant lists. Inspectors did not know the source of the prompt. The comparison is observational (FINDER-prompted versus other inspections during the same period), not randomized. Google employees conceived and implemented the model and co-authored the peer-reviewed evaluation. Health department staff oversaw implementation.

Outcome identifiers: Not assessed
Benefit: 52.3% were unsafe versus 24.7% of 10,786 other inspections (odds ratio 3.06)
Category: Improved Task Accuracy
Status: Not assessed
Beneficiary groups: Organizations. Society.
Beneficiaries: Not assessed
Reach: Not assessed
Distribution and access: Not assessed
Duration: Not assessed
Expected?: Yes
Costs and harms: Not assessed

Studies: Not assessed
Study design: Not assessed
Measurement: Not assessed
Setting and sample: Not assessed
Measures: 52.3% were unsafe versus 24.7% of 10,786 other inspections (odds ratio 3.06)
Comparison: all other restaurant inspections conducted in the two cities during the deployment period (BASELINE)
Results: Reference: 24.7% of BASELINE restaurants deemed unsafe. Result: 52.3% of FINDER-identified restaurants deemed unsafe (3.06 times as likely).
Additional value from AI: Not assessed
Limitations: Not assessed
Denominators: Not assessed
Materials for replication: Not assessed

Participants: Developer: Google. Operator: Southern Nevada Health District (Las Vegas) and Chicago Department of Public Health.
Jurisdictions: Benefit: Las Vegas and Chicago, United States. Developer: United States. Operator: United States.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Conditions for replication: Access to anonymized, aggregated search and location data held by Google. Tolerance of the incubation-period lag before inference. Inspectors unaware of the prompt source.
Replication attempts: 1 follow-up entries. Benefit recurred: 1.
Later findings: No report of FINDER being deployed again or adopted by other health departments after the 2018 paper was found.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Assessed benefits: Organizations. Society.
Depth: Organizations: 2. Society: Not assessed.
Reasons: Organizations: 69 of 132 FINDER-prompted inspections (52.3%) found an unsafe restaurant versus 2,662 of 10,786 (24.7%) baseline inspections. Odds ratio 3.06 (95% CI 2.14–4.35), adjusted for city and risk level. Society: The paper argues that corrective actions “would prevent” spread of foodborne illness but documents no reduction in illness.
Observed or intended: Not assessed
Rubric: Not assessed
Assessor and date: Not assessed
Alternative assessments: Not assessed
Fill: The makers' claim

Kind: Not assessed
Related entries: Not assessed
Field: Food safety / public health. Restaurant inspection targeting.
Purpose: Identify restaurants likely to be causing foodborne illness from anonymized, aggregated web-search and location data so that health departments can target inspections.
Dates: Date basis unassessed: 2018 (precision unassessed). Updated: 22 October 2026.
Stage: Not assessed
Version: SHA-256: ce8fdf81a945…

Peer-reviewed study Developer

Sadilek A, Caty S, DiPrete L, Mansour R, Schenk T Jr, Bergtholdt M, Jha A, Ramaswami P, Gabrilovich E. Machine-learned epidemiology: real-time detection of foodborne illness at scale. npj Digital Medicine 1:36 (2018). doi:10.1038/s41746-018-0045-1