Climate science has spent decades making the planet computable. The newer move is making it legible. A 2025 review in Nature Communications by Camps-Valls et al. maps how artificial intelligence now sits on both sides of that line. It extracts features from noisy, sparse observations to predict floods, droughts and heatwaves, and it reshapes how those predictions reach people. AI-generated inundation maps, photorealistic flood scenes, and language-based warnings tailored to different populations are all offering a new way to reimagine climate risk.
That convergence is the premise of a 2025 paper in Environmental Sciences Europe by Leal Filho et al., which argues foundation models can widen public engagement by generating interactive visualisations and simplified, expert-validated explanations of abstract climate concepts. Whether such explanations hold up is what Atkins et al. (2024) already tested in Communications Earth & Environment. Prompting GPT-3.5 and GPT-4 about hazard exposure across 191 countries and comparing the answers to established risk indices, they found broad agreement for floods and cyclones, weaker performance on droughts, and a bias towards Western, English-language framing. GPT-4 erred less than its predecessor, yet accuracy stayed uneven across exactly the regions where climate literacy matters most.
The harder finding concerns reception rather than accuracy. In three experiments (N = 2,580), Eppler (2026) found that highly realistic AI-generated disaster images do not raise support for climate action. They intensify emotion, but they also erode trust in the source, an effect strongest when the artificial origin is suspected or disclosed. When viewers sensed an image was synthetic but were not told, their willingness to make personal sacrifices fell. The realism that makes a flooded street feel close is the same property that, once detected, reads as manipulation.
This is why the rendering problem is finally an institutional one. Writing in Communications Earth & Environment on the IPCC's options for AI, Buck et al. (2026) locate its function not in summarising literature faster but in sustaining a shared basis of reality, and sketch a scenario in which public backlash against AI erodes precisely that. The papers converge on an awkward asymmetry. AI can make climate data more legible to more people than any prior medium, but legibility bought through synthetic fidelity neagtes the trust it depends on. The constraint is not effectiveness in having the message arrive but how credible it ends up being.