A new generation of AI systems called "world models" could help solve some of climate science's most stubborn problems, from hurricane landfall prediction to carbon cycle uncertainty, by learning how Earth's systems actually behave rather than just describing them. Unlike ChatGPT or other large language models (LLMs) that predict the next word in text, world models learn the underlying dynamics of physical systems by observing them, then simulate forward to test what happens next. Major AI labs including Google DeepMind, OpenAI, and researchers like Yann LeCun are now organizing their research around this architecture.

What Makes World Models Different From Today's AI?

The distinction matters because it addresses a fundamental gap in how AI currently tackles climate science. Today's weather forecasts are dramatically better than they were five years ago, thanks to neural networks trained on historical data. Google's flood forecasting system now runs in over 150 countries. DeepMind's GraphCast and similar neural weather models have matched or beaten traditional physics-based forecasts at a fraction of the computational cost.

But these successes mask a deeper problem. The hardest climate questions remain unsolved: predicting what a hurricane will do when it reaches land, knowing when the next drought will break, understanding how ocean circulation will shift as ice melts. Sub-seasonal weather forecasts, which drive water, energy, and agricultural planning decisions, remain weak. The forests, soils, and vegetation that absorb roughly one-third of global emissions carry the largest uncertainty in the entire carbon budget.

The bottleneck isn't computing power or data. It's representation. The Earth system is modeled in pieces, atmosphere and ocean and ice and land treated separately, which strips away the signals that live in the coupling between them. The variables that matter most are largely unobserved: root-zone soil moisture, the deep ocean, the cavities under ice shelves. A world model trained on how a system actually behaves, rather than how it looks, could learn these unwritten dynamics while respecting the physics we trust as hard constraints.

Why Are Major AI Labs Betting on This Architecture?

Three forces have converged to make world models viable now. The architecture has matured enough to train at scale. The Earth is now instrumented at a level that makes learning dynamics across multiple systems possible. And the capital that built the large language model era has begun to reallocate toward this new frontier.

Yann LeCun, who left Meta in late 2025 to launch Advanced Machine Intelligence Labs, has built his research program around world models. Demis Hassabis, who runs Google DeepMind, has made them central to its push toward more general AI. Fei-Fei Li raised a billion dollars for her company World Labs to pursue what she calls "spatial intelligence." Jensen Huang is building the simulation platforms and compute infrastructure behind the next wave of AI, much as NVIDIA did for large language models.

Some applications are already in operational use. AlphaFold, NVIDIA's Earth-2, and GraphCast are deployed across biology and weather forecasting. They work where physics is partly understood and observations are rich. But none yet learns the dynamics of the open systems whose uncertainty has barely narrowed: sea level rise, the carbon cycle, the coupled behavior of a warming planet.

What Could World Models Actually Deliver for Climate?

World models will not deliver certainty. They will not collapse the sea-level range to a single point estimate. Their value is narrower but real: tighter, more honest ranges of plausible futures, the band that coastal planners and banks actually need for trillion-dollar adaptation choices.

There is also a hard limit. A world model cannot forecast a regime the Earth has never entered, like the world after an Atlantic circulation collapse, because there is no data from the far side to learn from. No method can. But the relevant question is how close we are to it, which these systems can begin to help answer.

The potential for solving problems in climate, oceans, the biosphere, and the biology of disease is vast. But what these models train on shapes what they become. A world model trained predominantly on warehouse logistics, driving footage, and engineering data learns a particular kind of physics. One trained also on planetary observation, cell biology, and grid dynamics learns something else.

How Can Schools and Communities Address AI's Climate Impact Today?

While world models represent a long-term research frontier, immediate questions about AI's environmental footprint are already reaching classrooms and school boards. Portland Public Schools is grappling with how to address AI use in an era of climate commitments, revealing a gap between ambitious climate policies and the rapid deployment of energy-intensive technology.

Students at Portland schools have raised concerns about the environmental cost of generative AI. A 2026 United Nations report suggests that if the data centers used to power AI were considered a country, they would rank 11th globally in electricity consumption. Google's data centers in The Dalles, Oregon, consumed nearly 550 million gallons of water in 2025, according to reporting by The Oregonian, nearly 40 percent of the entire city's water use.

In June, Portland's Climate Crisis Response Committee delivered a recommendation suggesting AI use in the classroom be limited "as much as possible," and that the district publicly track, monitor, and codify AI use. The committee also proposes that any AI use in classes should be accompanied by education about its environmental impact.

In June

• Track AI Usage: Schools should catalog existing AI tools in use and establish baseline metrics for energy consumption and water use associated with these systems.
• Educate About Environmental Costs: When AI is used in classrooms, pair it with explicit instruction about the electricity and water required to train and run these models.
• Evaluate Alternatives: Before adopting new AI tools, schools should assess whether non-AI solutions or lower-impact technologies can achieve the same educational goals.
• Engage Students in Decision-Making: Include student voices in technology adoption decisions, particularly those concerned about climate impact, to ensure policies reflect community values.

Alma Valls, a rising senior and president of Cleveland High School's environmental club, became aware of AI's climate impact through conversations with friends. She now serves as a student representative on the Climate Crisis Response Committee. "I was really confused, and I hadn't realized the environmental effects of AI," Valls explained. "Being in friend groups and in a school that's very progressive on climate shifts has caused me to know more about it, and be more anti-AI, and I'm definitely hearing other people concerned about it".

"I was really confused, and I hadn't realized the environmental effects of AI. Being in friend groups and in a school that's very progressive on climate shifts has caused me to know more about it, and be more anti-AI, and I'm definitely hearing other people concerned about it," said Alma Valls.

Alma Valls, rising senior and environmental club president, Cleveland High School

Ian Ritorto, a just-graduated Roosevelt High School senior who served as the School Board's student representative, told reporters that PPS's student body feels "overwhelmingly negative" about AI. When bringing up AI's use in classrooms to fellow students and teachers, almost all express a level of outrage.

Portland Public Schools adopted its Climate Crisis Response Policy in 2022, aiming to cut the district's emissions in half by 2030 and reach net zero by 2040. The policy is one of the most progressive in the nation, explicitly prohibiting fossil fuel infrastructure in new buildings and requiring climate considerations in procurement. But the policy passed well before the explosive rise of AI, and therefore offers no guidance for how the district should navigate such software.

The district has opted not to renew its two generative AI contracts, which included Amira Learning and Lumi Story AI. Superintendent Dr. Kimberlee Armstrong stated that "right now, there's no generative AI or AI product that we have in Portland Public Schools that's moving forward in the fall." The School Board also placed a pause on new generative AI contracts and asked district officials to catalog existing AI in the district.

Dr. Kimberlee Armstrong

Angela Long, a former Climate Crisis Response Committee member and PPS parent, takes a more moderate position than some advocates. She doesn't think AI needs to be eradicated from the classroom, but wants PPS to adopt a policy that helps teach students responsible use. "They're not just using it to do homework, they're using it for everything," Long noted. "There's no boundary on when you should be using it and not be using it, and that's where I think, as an educator, it's your role to help students know the difference between how to use this tool and when to use the tool".

The tension between AI's potential and its environmental cost reflects a broader challenge facing institutions committed to climate action. As world models and other advanced AI systems mature, the decisions made now about what data they train on, who builds them, and how they're deployed will shape what AI can do for climate science for years to come.