Google flood alerts in Bihar cut reported illness by 16.6%

AI system

System: Google flood forecasting with Android push alerts (Google Public Alerts), plus community dissemination by Yuganter volunteers
Model: Not assessed
Release: Not assessed
Application developer: Google
Role: Not assessed
Configuration: Not assessed
Inputs and outputs: Not assessed
Autonomy: Not assessed
Workflow: Google/CWC flood alerts available to all 319 study panchayats (both arms); the randomized element was volunteer dissemination. Trained community volunteers relaying smartphone alerts in villages with low literacy and limited location-services use. Repeated flood seasons (2022 was more intense than 2023) allowing effects to be measured.
Human oversight: Not assessed
Relation to AI: The AI forecast generated the alerts. The randomized contrast, however, is between community dissemination of those alerts and alerts alone. So the measured effect is attributable to the dissemination of AI forecasts rather than to forecast accuracy itself. A 2019 pilot in which the only added treatment was raising alert awareness with a single local leader found no treatment impact.
AI contribution: The AI forecast generated the alerts. The randomized contrast, however, is between community dissemination of those alerts and alerts alone. So the measured effect is attributable to the dissemination of AI forecasts rather than to forecast accuracy itself. A 2019 pilot in which the only added treatment was raising alert awareness with a single local leader found no treatment impact.

Benefits

Outcome identifiers: Not assessed
Benefit: 16.56% fewer illness symptoms and about 30% lower medical expenditure
Category: Improved Physical Health
Status: Not assessed
Beneficiary groups: People. Society.
Beneficiaries: Not assessed
Reach: Not assessed
Distribution and access: Not assessed
Duration: Not assessed
Expected?: Yes
Costs and harms: Not assessed

Measurement

Studies: Not assessed
Study design: Not assessed
Measurement: Not assessed
Setting and sample: Not assessed
Measures: 16.56% fewer illness symptoms and about 30% lower medical expenditure
Comparison: control communities, where households had access to the smartphone alerts but no community volunteers relaying them
Results: Reference: control households (reference group). Result: 0.36 fewer illness symptoms (a 16.56% decrease); roughly INR 6,500 less on illness treatment (a 32.87% decrease).
Additional value from AI: Not assessed
Limitations: Not assessed
Denominators: Not assessed
Materials for replication: Not assessed

Money

Kind of value: less spent on illness treatment per household over the flood season
Amount: INR 6,500
Currency and prices: INR, stated in the source. Nominal source amount; price year not separately stated.
Period and unit: 2022–2023 flood-season study; the prose does not isolate the monetary recall window. Estimated difference per household in illness-treatment spending.
Realization: Retrospective estimate.
Recipients and payers: Recipient: Households in treatment communities in Bihar, India. Payer: Households paying for illness treatment.
Gross or net: Estimated household medical-expenditure reduction, not a full net-welfare or program-profit estimate.
Included costs: Not assessed
Comparison: Randomized community dissemination versus control communities, both with access to the AI forecasting system.
Method: Not assessed
Return and payback: Not assessed

People and places

Participants: Developer: Google. Operator: Central Water Commission of India with Google; last-mile dissemination by Yuganter (NGO) under the Jagnani-Pande evaluation.
Jurisdictions: Benefit: Bihar (twelve flood-prone districts), India. Developer: United States. Operator: India.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Follow-up

Conditions for replication: Alerts reach households before water arrives (46.5% more likely in treatment communities in 2022). Households trust alerts despite more false positives (8.09% vs 3.64%). Dissemination capacity is maintained. From 2023 two volunteers per panchayat replaced three, supplemented by direct WhatsApp/SMS messages to treatment households enrolled in the 2022 endline survey.
Replication attempts: 1 follow-up entries. Benefit recurred: 1.
Later findings: Neither cited source reports a development after the 2022-2023 evaluation seasons.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Evidence

Source numbers: AIBD-0031-source-0001–0002
Citations: Numbered references below.
Language: Not assessed
Source relationships: Independent
Original evidence: Not assessed
Access and versions: Not assessed
Supported findings: Not assessed
Evidence links: Not assessed
Evidence strength: Not assessed
Missing evidence: Not assessed

Grades

Assessed benefits: People. Society.
Depth: Not assessed
Reasons: People: Effects are cohort averages across roughly 3.6 million people's communities (fewer illness symptoms, lower medical spending). No identifiable household with an attributable changed outcome is documented. Society: The study estimates population-level preparedness and health effects from a randomized design. It reports no discrete event such as a named evacuation.
Observed or intended: Not assessed
Rubric: Not assessed
Assessor and date: Not assessed
Alternative assessments: Not assessed
Fill: Independently studied

Entry details

Kind

Not assessed

Related entries

Same developer and channel: Google's Android hazard alerts; flood alerts cut illness in Bihar, while the Android Earthquake Alerts system underestimated the Kahramanmaraş quake and failed to warn millions of Android users in Turkey.

Field

Disaster risk reduction / flood early warning. Rural households in riverine flood zones.

Purpose

The forecasting system was built to warn people of riverine floods before water arrives so they can prepare or evacuate. The measured preparedness and health effects are that intended benefit, but only at cohort level.

Dates

Date basis unassessed: 2022 (precision unassessed). Updated: 21 October 2026.

Stage

Not assessed

Version

SHA-256: 1bfab83fdd45…

References

Other identifiers: AI Incident Database: 1546.
Category mappings: Not assessed
Brief: Not assessed
Revisions: 7 revisions. Latest: 21 October 2026.

Sources

Report Independent author

Abdul Latif Jameel Poverty Action Lab (J-PAL) (n.d.). Disseminating Flood Warnings to Increase Disaster Preparedness in India. Evaluation summary (PDF).
Preprint Independent author

Jagnani M, Pande R (2024). Forecasting Fate: Experimental Evaluation of a Flood Early Warning System. Working paper, Tufts University and Yale University (Economic Growth Center), 3 April 2024.