AI system
- System
-
Google DeepMind AI hurricane forecast model, ensemble mean designated GDMI in NHC guidance
- Model
-
Not assessed
- Release
-
Not assessed
- Application developer
-
DeepMind - Role
-
Not assessed
- Configuration
-
Not assessed
- Inputs and outputs
-
Not assessed
- Autonomy
-
Not assessed
- Workflow
-
Forecasters received the experimental DeepMind ensemble in real time and combined it with conventional forecasts. They retained responsibility for the official track and intensity forecasts.
- Human oversight
-
Not assessed
- Relation to AI
-
NHC forecasters state in Discussions 8, 10, 12, 13 and 14 that official track and intensity forecasts were blended or aligned with Google DeepMind's ensemble. The official forecast remains a human synthesis of several aids. So the AI contribution to outcomes in Jamaica is not isolated.
- AI contribution
-
NHC forecasters state in Discussions 8, 10, 12, 13 and 14 that official track and intensity forecasts were blended or aligned with Google DeepMind's ensemble. The official forecast remains a human synthesis of several aids. So the AI contribution to outcomes in Jamaica is not isolated.
Benefits
- Outcome identifiers
-
Not assessed
- Benefit
-
National Hurricane Center forecasters used Google DeepMind’s guidance in their official forecasts for Hurricane Melissa.
- Category
-
Improved Task Accuracy
- Status
-
Not assessed
- Beneficiary groups
-
People. Organizations. Society.
- Beneficiaries
-
National Hurricane Center forecasters. People in Melissa’s path were the intended beneficiaries of the forecasts.
- Reach
-
No count of people who benefited from the AI guidance is established.
- Distribution and access
-
Not assessed
- Duration
-
Not assessed
- Expected?
-
Yes
- Costs and harms
-
Not assessed
Measurement
- Study design
-
Retrospective verification of operational forecasts against the storm’s observed track and intensity. [8] [9] [10]
- Measurement
-
NHC compared forecasts at matching lead times with its final estimates of the storm’s position and intensity. It compared official forecasts with model guidance on a common sample. [8] [9] [10]
- Setting and sample
-
Hurricane Melissa in the Atlantic, 21–31 October 2025. The report compares forecasts at lead times through 120 hours. [8] [9] [10]
- Comparison
-
NHC’s official forecasts, conventional model guidance, and climatology-persistence forecasts for the same storm. [8] [9] [10]
- Results
-
GDMI’s track forecasts were comparable to NHC’s official forecasts through 96 hours. NHC performed slightly better at 120 hours and had lower intensity errors than every guidance model through 96 hours. [8] [9] [10]
- Additional value from AI
-
NHC credits GDMI and regional models jointly with increasing forecasters’ confidence at unusually long lead times. It does not isolate GDMI’s effect on the official forecast. [8] [9] [10]
- Limitations
-
The sources establish that forecasters used the guidance. They do not establish that it changed protective actions, reduced losses, or prevented deaths. [8] [9] [10]
- Denominators
-
Not assessed
- Materials for replication
-
Not assessed
Money
- Kind of value
-
Not assessed
- Amount
-
Not assessed
- Currency and prices
-
Not assessed
- Period and unit
-
Not assessed
- Realization
-
Not assessed
- Recipients and payers
-
Not assessed
- Gross or net
-
Not assessed
- Included costs
-
Not assessed
- Comparison
-
Not assessed
- Method
-
Not assessed
- Return and payback
-
Not assessed
People and places
- Participants
-
Developer:
DeepMind. Operator: 
NOAA National Hurricane Center. - Jurisdictions
-
Benefit: New Hope,
Jamaica. Developer: 
United Kingdom. Operator: 
United States. - Ownership and control
-
Not assessed
- Dependencies
-
Not assessed
- Cross-border connections
-
Not assessed
- Geopolitical implications
-
Not assessed
Follow-up
- Conditions for replication
-
Forecasters need timely model output, checks before operational use, and conventional guidance for comparison. Later observations are needed to assess forecast accuracy.
- Replication attempts
-
2 follow-up entries. Benefit recurred: 2.
- Later findings
-
NHC published its Tropical Cyclone Report on 25 February 2026 and revised it on 19 May. The report evaluates the forecasts against the storm’s observed track and intensity.
- Availability
-
Not assessed
- Institutional decision
-
Not assessed
- Grounds
-
Not assessed
- Action and result
-
Not assessed
- Success criterion
-
Not assessed
Evidence
- Source numbers
-
AIBD-0088-source-0001–0010
- Citations
-
Numbered references below.
- Language
-
Not assessed
- Source relationships
-
Operator
- Original evidence
-
NHC’s operational advisories document its use of DeepMind guidance. Its later verification reports assess forecast accuracy. IFRC reports provide humanitarian context without isolating AI’s effects. [8] [9] [10]
- Access and versions
-
The NHC interview, 2025 verification report, and revised Melissa report were accessible in full. Earlier access limits remain attached to the individual IFRC reports. [8] [9] [10]
- Supported findings
-
Not assessed
- Evidence links
-
Not assessed
- Evidence strength
-
Not assessed
- Missing evidence
-
Not assessed
Grades
- Assessed benefits
-
People. Organizations. Society.
- Depth
-
People: Not assessed. Organizations: 2. Society: 3.
- Reasons
-
GDMI and regional models increased forecasters’ confidence in predicting rapid intensification. NHC’s later review found that GDMI track forecasts performed comparably to its official forecasts through 96 hours. No study isolates the AI’s effect on protective actions or losses.
- Observed or intended
-
Not assessed
- Rubric
-
Not assessed
- Assessor and date
-
Not assessed
- Alternative assessments
-
Not assessed
- Fill
-
The deployer's account
Entry details
- Kind
-
Not assessed
- Related entries
-
Google-built hazard early-warning system underestimated the Kahramanmaras earthquake and failed to warn millions; same developer family and public-warning task as DeepMind's Melissa forecasts.
- Field
-
Weather forecasting / tropical cyclones. Public-sector meteorological service.
- Purpose
-
Experimental AI tropical cyclone forecast model providing ensemble track and intensity guidance to NHC forecasters.
- Dates
-
Date basis unassessed: 2025 (precision unassessed). Updated: 25 October 2026.
- Stage
-
Not assessed
- Version
-
SHA-256: 7494545dd8ae…
References
- Other identifiers
-
AI Incident Database: 1546.
- Category mappings
-
Not assessed
- Brief
-
Not assessed
- Revisions
-
11 revisions. Latest: 25 October 2026.
Sources
-
Report Deployer
National Hurricane Center (2025). Tropical Storm Melissa Discussion Number 1, 1100 AM EDT Tue Oct 21 2025, forecasters Cangialosi/Bucci. NOAA/NWS.
Original operational advisory establishing that DeepMind ensemble guidance contributed to the official forecast from the first advisory.
The NHC official track forecast lies between the Google DeepMind ensemble mean track and the correct consensus aid, HCCA.
-
Report Deployer
National Hurricane Center (2025). Tropical Storm Melissa Discussion Number 2, 500 PM EDT Tue Oct 21 2025, forecaster Cangialosi. NOAA/NWS.
Original operational advisory recording the blend of DeepMind and other guidance.
This prediction is a little left of the previous forecast and remains a blend of the latest Google DeepMind ensemble mean, the corrected consensus aid HCCA, and simple consensus TVCA.
-
Report Deployer
National Hurricane Center (2025). Tropical Storm Melissa Discussion Number 8, 500 AM EDT Thu Oct 23 2025, forecaster Cangialosi. NOAA/NWS.
Original operational advisory recording a later adjustment toward DeepMind and HCCA guidance.
The NHC track forecast is shifted a little to the right of the previous one, trending toward a blend of the latest Google DeepMind ensemble mean and HCCA solution.
-
Report Deployer
National Hurricane Center (2025). Tropical Storm Melissa Discussion Number 10, 500 PM EDT Thu Oct 23 2025, forecaster Papin. NOAA/NWS.
Original operational advisory recording track and intensity guidance. It does not isolate downstream public benefits.
the latest NHC track was only nudged slightly southward from this morning, blending the reliable HCCA and GDMI aids
-
Report Deployer
National Hurricane Center (2025). Tropical Storm Melissa Discussion Number 13, 1100 AM EDT Fri Oct 24 2025, forecaster Papin. NOAA/NWS.
Original operational advisory recording the ensemble’s Category 4-or-higher signal.
Every single Google DeepMind member shows the system become a Category 4 hurricane or higher
-
Report Deployer
National Hurricane Center (2025). Tropical Storm Melissa Discussion Number 14, 500 PM EDT Fri Oct 24 2025, forecaster Papin. NOAA/NWS.
Original operational advisory recording the intensity forecast’s agreement with GDMI.
The NHC intensity forecast is on the high side of the guidance envelope, and closest too the Google DeepMind ensemble mean (GDMI).
-
Report Deployer
Original landfall update establishing storm context, not the AI system’s effect on public outcomes.
CATEGORY 5 MELISSA MAKES LANDFALL IN JAMAICA...ONE OF THE MOST POWERFUL HURRICANE LANDFALLS ON RECORD IN THE ATLANTIC BASIN
-
Press release Deployer
National Weather Service (n.d.). AI in Hurricane Forecasting at the National Hurricane Center: A Q&A with NHC Science Operation Officer Wallace Hogsett. NOAA/NWS News Around NOAA.
Original interview with NHC’s Science Operations Officer. NHC deployed the model and partnered in its development; this is not an independent evaluation.
The AI models honed in very early on the likely track and intensity and provided very valuable guidance to complement our traditional NWP guidance.
-
Report Deployer
Original seasonal verification by NHC and a cooperating researcher. It evaluates forecasting performance, but it is not an independent evaluation of AI-attributable humanitarian outcomes.
the NHC track errors are notably lower than average for major hurricanes, including Humberto and Melissa in 2025
-
Report Deployer
Printed pp. 16–17, Track Forecast and Intensity Forecast, Tables 5b and 6b.
Original retrospective verification by the agency that used the guidance. The report compares track and intensity forecasts for Melissa and explains how GDMI and regional models contributed to forecasters’ confidence. It does not isolate the AI model’s effect on humanitarian outcomes.
The NHC OFCL forecasts were comparable to GDMI through 96 h, but performed slightly better than GDMI at 120 h.