AI weather forecasting now beats the world’s leading physics-based model on 97.2 percent of the metrics scientists use to judge accuracy, and it does so in eight minutes on a single chip rather than hours on a supercomputer. That is not a marketing claim. It is the result Google DeepMind published in the journal Nature after testing its GenCast model against 1,320 separate forecasting targets, comparing it directly to the European Centre for Medium-Range Weather Forecasts’ ENS system, the operational standard that national weather services and airlines have relied on for decades.
The honest answer to whether AI can predict the weather better than humans, however, is not a simple yes. It depends entirely on what kind of weather you are asking about, and the sharpest limits on AI forecasting show up precisely in the events where an accurate warning matters most.
How AI Weather Forecasting Actually Works
Traditional numerical weather prediction, the approach meteorologists have used since the 1950s, solves a vast system of physics equations describing how air, moisture, and pressure move through the atmosphere, using measured current conditions as a starting point. It is accurate but computationally enormous, requiring supercomputers running for hours to produce a single forecast.
AI weather forecasting takes a fundamentally different approach. According to Google DeepMind’s own account of GenCast, the model was trained on four decades of historical weather data from ECMWF’s ERA5 reanalysis archive, learning statistical patterns in how atmospheric conditions typically evolve rather than solving physics equations from scratch each time. GenCast is a diffusion model, the same category of AI that powers image and video generation, adapted to the spherical geometry of the Earth. Rather than producing one deterministic forecast, it generates an ensemble of 50 or more possible weather trajectories, letting forecasters see not just a single predicted outcome but the full range of plausible scenarios and how likely each one is.
The Compute Cost Behind the Speed Advantage
Part of AI weather forecasting’s appeal is its efficiency: GenCast produces a 15-day ensemble forecast in roughly 8 minutes on a single Google Cloud TPU, compared to hours on a supercomputer running thousands of processors for a traditional ensemble. That efficiency claim is real and independently verifiable, but it describes inference, running an already-trained model, rather than the cost of building it in the first place.
Training a model like GenCast on four decades of reanalysis data at global scale requires substantial compute, and that upfront cost sits inside the same data centre infrastructure LiveAIWire has examined in its reporting on the water and energy demands of AI data centres, which found that efficiency gains at the model level do not automatically translate into lower total resource consumption once demand for that efficient model expands to match.
For weather forecasting specifically, that trade-off looks more favourable than it does for generative AI broadly, since a single trained model can generate an enormous number of operational forecasts relative to its one-time training cost, and the alternative it replaces, a physics-based ensemble run on a dedicated supercomputer, is not exactly power-light either. The efficiency case for AI weather forecasting is genuine. It is simply not costless, and treating it as such would repeat exactly the accounting gap LiveAIWire’s broader climate coverage has flagged elsewhere in the AI industry.
Where the Accuracy Numbers Actually Come From
The 97.2 percent figure is specific and independently reproducible, not a rounded-up headline statistic. DeepMind’s peer-reviewed paper in Nature tested GenCast against ECMWF’s full 51-member ensemble across forecasts of different weather variables at different lead times out to 15 days, and GenCast was more accurate on 97.2 percent of those combinations, rising to 99.8 percent for forecasts more than 36 hours out. The gains were not limited to routine conditions. GenCast’s own published results show it consistently outperforming ENS on extreme heat, extreme cold, and high wind speed forecasts, and on tracking the paths of tropical cyclones, information with direct implications for evacuation decisions and disaster preparedness timing.
GenCast’s predecessor, GraphCast, produced a similarly striking result the year before: more accurate forecasts than ECMWF’s deterministic HRES model on more than 90 percent of 1,380 test variables, rising to 99.7 percent within the troposphere, the lowest layer of atmosphere where most weather that affects daily life actually happens. Both models are now open source, with code and weights published for the wider meteorological community, and ECMWF itself has run live experiments using GraphCast’s forecasts on its own systems.
The Limit Nobody’s Marketing Slide Mentions
The accuracy figures above describe how AI weather forecasting performs on ordinary and moderately extreme conditions. A separate, more recent line of research asked a narrower and more consequential question: how do these models perform on weather that has never happened before, the record-shattering events that cause the most damage and the most deaths.
The answer, published in research covered in detail by Carbon Brief, is that physics-based models still win decisively where it matters most. Researchers led by Zhongwei Zhang at the Karlsruhe Institute of Technology tested three leading AI models, GraphCast, Pangu-Weather, and Fuxi, against ECMWF’s HRES on roughly 246,000 record-breaking heat, cold, and wind events recorded in 2020. HRES consistently outperformed all three AI models, and the gap was widest for the events that broke previous records by the largest margins. The AI models systematically underpredicted the intensity of heat records and overpredicted the intensity of cold records, with the error growing larger the further a record exceeded anything in the model’s training history.
The mechanism behind that gap is not a flaw specific to any one model. It is a structural consequence of how these systems learn. Study co-author Sebastian Engelke told Carbon Brief that AI models “depend strongly on the training data” and are “relatively constrained to the range of this dataset.” A record-breaking heatwave, by definition, produces conditions the model has never seen in training. Physics-based models are not limited this way, since they solve the underlying equations of atmospheric behaviour directly rather than pattern-matching against historical examples, which means they can in principle represent a genuinely unprecedented event even when nothing quite like it appears in the data they were built from.
Why This Distinction Matters More Than the Headline Accuracy Number
Study co-author Erich Fischer, a climate scientist at ETH Zurich, called the finding a “warning shot” against replacing traditional forecasting models with AI ones too quickly, specifically because accurate forecasts are needed most in the run-up to exactly the events AI currently forecasts worst. Early warning systems exist to trigger evacuations and mobilise disaster response ahead of the most damaging weather, not the routine kind, and a systematic underprediction of a heat record’s true intensity is precisely the failure mode that costs lives if a warning system relies on it uncritically.
It is worth noting that the AI models tested in that study, GraphCast, Pangu-Weather, and Fuxi, are deterministic models from an earlier generation. Newer probabilistic models like GenCast, which generate a full ensemble of possible outcomes rather than a single prediction, have not yet been tested against the same record-breaking-extremes protocol, and researchers in the field consider it a genuinely open question whether the ensemble approach narrows this specific gap. What is well established is that no current AI weather model, deterministic or probabilistic, has been shown to match physics-based forecasting specifically for genuinely unprecedented extremes.
How Forecasters Actually Use This Today
The practical answer major weather agencies have converged on is neither replacing physics-based forecasting with AI nor ignoring AI’s genuine strengths. Operational meteorologists at ECMWF, NOAA, and the UK Met Office run AI models alongside traditional numerical weather prediction and ensemble products, treating AI output as one additional voice in the forecasting discussion rather than a replacement for the others. When AI and physics-based models disagree, that divergence itself becomes useful information, signalling elevated forecast uncertainty that prompts closer scrutiny rather than being resolved by simply picking whichever model ran fastest.
NOAA’s National Hurricane Center has explicitly incorporated AI model guidance, including GraphCast and Pangu-Weather output, into its internal forecast discussions for tropical cyclone tracking, where AI models have shown genuine skill. That integration reflects the actual state of the field: AI weather forecasting has earned a real, evidence-backed seat at the table for the large majority of routine and even moderately extreme forecasting, while physics-based models remain the standard specifically for the rare, record-breaking events where the stakes of getting it wrong are highest.
The Genuine Success Story Behind the Skepticism
None of the limitations documented above should obscure how unusually well-verified AI weather forecasting is compared with most other claims made about AI’s real-world benefits. LiveAIWire’s investigation into machine greenwashing across the AI industry found that 74 percent of the climate-benefit claims major tech companies make about AI are unproven, with more than a third citing no evidence at all. Weather forecasting was one of the rare exceptions the investigation specifically called out: GraphCast and DeepMind’s flood and cyclone forecasting tools have been validated against independent meteorological benchmarks and published in peer-reviewed journals, not merely asserted in a corporate sustainability report.
That same pattern of genuine, audited progress with real gaps extends into related disaster forecasting work. LiveAIWire’s coverage of AI flood forecasting closing the urban warning gap found nearly identical dynamics: a model that performs remarkably well where it has dense historical data to learn from, and measurably worse in regions and situations where that data is thin, precisely the same training-data dependency that limits AI weather models on record-breaking extremes. The same dependency shows up again in AI wildfire detection satellites, where LiveAIWire’s reporting found the fastest coverage gains still concentrated in well-monitored regions, while the newly settled or historically under-surveyed areas that need earlier warning most remain the hardest for any training-data-dependent model to serve reliably.
What This Means for You
For everyday forecasts, the five-day outlook on your phone, the chance of rain this weekend, AI weather forecasting is very likely already contributing to what you see, whether through Google Search and Maps or the underlying models national weather services increasingly run alongside their traditional systems, and the evidence strongly supports trusting it as much as you trust any modern forecast.
For the events that actually threaten lives and property, an unprecedented heatwave, an unusually intense cold snap, a wind event beyond anything on record, the current evidence supports treating official warnings from national meteorological agencies, which blend AI and physics-based guidance rather than relying on either alone, as more reliable than any single AI model’s output on its own.
The field is moving quickly, and the specific gap documented in record-breaking extremes is exactly the kind of problem AI weather research is now actively working to close, through hybrid approaches that combine physics-based simulation with AI’s speed and pattern recognition. Until that gap closes and is independently verified the way GenCast’s routine-weather performance already has been, the safest reading of “can AI predict the weather better than humans” is: yes, for almost everything, and not yet, for the weather that matters most.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.