GraphCast matches best operational cyclone track models in 2023

AI system

System: GraphCast (graph neural network global weather model)
Model: GraphCast.
Release: Exact release not established.
Application developer: DeepMind
Role: GraphCast generated weather forecasts.
Configuration: The event concerned GraphCast’s live experimental forecasts on ECMWF charts in 2023. The CIRA evaluation used GraphCast-operational.
Inputs and outputs: Not assessed
Autonomy: Not assessed
Workflow: AI models initialized from operational NWP analyses and run alongside physics-based models. Verification using the NHC procedure, including early-model conversion and consensus experiments.
Human oversight: Not assessed
Relation to AI: The CIRA preprint isolates the AI models' contribution via NHC verification and consensus experiments. The nine-day Hurricane Lee landfall forecast rests on vendor statements and press quotes of ECMWF staff. ECMWF documents read (Newsletters 177 and 178, science blog, Severe Event Catalog) do not record it.
AI contribution: The CIRA preprint isolates the AI models' contribution via NHC verification and consensus experiments. The nine-day Hurricane Lee landfall forecast rests on vendor statements and press quotes of ECMWF staff. ECMWF documents read (Newsletters 177 and 178, science blog, Severe Event Catalog) do not record it.

Benefits

Outcome identifiers: Not assessed
Benefit: adding all four improved the NHC track consensus by 11% at 120 h
Category: Improved Task Accuracy
Status: Not assessed
Beneficiary groups: People. Organizations. 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: adding all four improved the NHC track consensus by 11% at 120 h
Comparison: the NHC track consensus without the four AI models, benchmarked against the long-term improvement of official forecasts
Results: Reference: consensus track errors without the AIWP models; official five-day track forecasts improving by about 2% per year since 2001. Result: consensus track errors reduced by up to 11% at the longer time periods.
Additional value from AI: Not assessed
Limitations: Not assessed
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: ECMWF (experimental live runs on public charts); CIRA/Colorado State University (evaluation).
Jurisdictions: Developer: United Kingdom. Operator: United Kingdom.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Follow-up

Conditions for replication: Track skill does not extend to intensity: the CIRA preprint reports a substantial low intensity bias. Independent case evaluation (Storm Ciaran) found peak winds underestimated by all AI models tested.
Replication attempts: 2 follow-up entries. Outcome unresolved: 1. Benefit recurred: 1.
Later findings: 30 March 2026. By the 2025 season NHC had brought AI-based models, including a newer Google DeepMind ensemble, into real-time operations, per its verification report.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Evidence

Source numbers: AIBD-0090-source-0001–0007
Citations: Numbered references below.
Language: Not assessed
Source relationships: Independent, Developer, Operator
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. Organizations. Society.
Depth: Not assessed
Reasons: People: Skill evaluation, no verified discrete event. Organizations: Skill evaluation, no verified discrete event. Society: Skill evaluation, no verified discrete event.
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

Google's AI-based earthquake alert system underestimated hazard magnitude; same developer family and hazard-prediction task as DeepMind's GraphCast cyclone forecasts, harming the exposed public.

Field

Weather forecasting / tropical cyclones. Meteorological research and operational forecast centers.

Purpose

Medium-range global weather forecasting by machine learning, including severe-event prediction such as tropical cyclone tracks.

Dates

Date basis unassessed: 2024 (precision unassessed). Updated: 22 October 2026.

Stage

Not assessed

Version

SHA-256: bb38f0333464…

References

Other identifiers: AI Incident Database: 1546.
Category mappings: Not assessed
Brief: Not assessed
Revisions: 8 revisions. Latest: 22 October 2026.

Sources

Report Independent author

Cangialosi, J. P., Martinez, J., Reinhart, B., Papin, P. (2026). National Hurricane Center Forecast Verification Report: 2025 Hurricane Season. NOAA/NHC, 30 March 2026.
Preprint Independent author

DeMaria, M., Franklin, J. L., Chirokova, G., Radford, J., DeMaria, R., Musgrave, K. D., Ebert-Uphoff, I. (2024). Evaluation of Tropical Cyclone Track and Intensity Forecasts from Artificial Intelligence Weather Prediction (AIWP) Models. arXiv:2409.06735, submitted 8 Sep 2024; the preprint's disclaimer reads 'This work has been submitted to the Artificial Intelligence in the Earth Sciences American Meteorological Society journal'. The quote is on p. 23 of the full PDF, not on the abstract page.
Preprint Developer

Lam, R., Sanchez-Gonzalez, A., Willson, M., et al. (2023). GraphCast: Learning skillful medium-range global weather forecasting. arXiv:2212.12794 (published in Science, 2023).
Preprint Independent author

Charlton-Perez, A. J., Dacre, H. F., Driscoll, S., et al. (2024). Do AI models produce better weather forecasts than physics-based models? A quantitative evaluation case study of Storm Ciaran. arXiv:2312.02658 (University of Reading).
News report Independent author

MIT Technology Review (2023). Google DeepMind's weather AI can forecast extreme weather faster and more accurately. 14 November 2023.
Press release Developer

Google DeepMind (2023). GraphCast: AI model for faster and more accurate global weather forecasting. DeepMind blog, 14 November 2023.
Press release Deployer

Rabier, F. (2023). The AIFS is launched (editorial). ECMWF Newsletter No. 177, Autumn 2023; the page states 'Published in October 2023' and its metadata dates it 16 October 2023.