Climate sensitivity range
How much the planet warms in the long run for a doubling of CO2 is still uncertain: IPCC says 2.5-4 °C.
Open in the interactive tree →IPCC AR6 (2021) judged equilibrium climate sensitivity likely 2.5-4 °C with a best estimate of 3 °C, narrowing the 1979 Charney range of 1.5-4.5 °C. Cloud feedback, especially of low clouds, is the largest uncertainty. Some newer models run 'hot' and are given reduced weight.
As of October 2026
The AR6 range remains the official benchmark in 2026. Research on emergent constraints, paleoclimate and satellite cloud data keeps refining it: a June 2026 paper in Earth System Dynamics constrains the transient climate response to 1.81 K (very likely 1.28-2.33 K) and concludes that the record-warm years 2023-2024 do not justify raising sensitivity estimates.
What is missing
- Observations and models of low-cloud feedback at process scale
- Better knowledge of how much aerosol pollution masks warming
- A longer satellite record of Earth's energy balance
- Paleoclimate proxies with smaller error
- Global storm-resolving models run for centuries
Becomes possible once solved
- Narrower projections of warming for each emissions path
- A more precise remaining carbon budget
- Better regional climate risk estimates for planning
Open steps
- Low-cloud feedback Medium AI leveragePin down how low clouds change as the planet warms, using cloud-resolving simulations and satellite data.
- Aerosol masking of warming Low AI leverageQuantify how much past warming pollution aerosols hid, which sets how large the greenhouse response really is.
- Learned constraints on sensitivity Medium AI leverageFind observable quantities in model ensembles and satellite records that reliably narrow the sensitivity range, and test them out of sample.
- Century-long storm-resolving runs High AI leverageRun global storm-resolving models for centuries, or emulate them, so cloud and circulation feedbacks appear without hand-tuned parameters.
- Tighter paleoclimate constraints Low AI leverageReduce proxy and dating error for past warm and cold climates so they can constrain sensitivity.
Where AI could help
Medium AI leverage. AI emulators make large ensembles and cloud-resolving physics cheaper, but cloud feedback is limited by physics and observations, not compute alone.
- Emulate cloud-resolving physics inside climate models to run many more long, high-resolution simulations
- Learn emergent constraints from model ensembles and satellite records to narrow sensitivity ranges
- Combine satellite cloud and energy-balance data to estimate low-cloud feedback directly
- Diagnose why some models run hot, using explainable-AI tools on the ensembles
Shown so far
- In July 2024, NeuralGCM (Nature) combined a differentiable atmosphere solver with learned physics and tracked climate metrics over decades at 140 km resolution with prescribed sea-surface temperatures. source
- In September 2018, a deep network trained on a cloud-resolving model reproduced subgrid convection inside a climate model at a fraction of the cost (Rasp et al., PNAS). source
Prerequisites
- Earth observation by satellite1960
- Climate models1967
- Ice cores1987
- Argo Ocean Float Network2000-2007Ocean heat content from Argo constrains Earth's energy imbalance