Earth's radiative feedback is the mechanism that stabilizes climate. The planet warms, it radiates more energy to space, the warming dampens. The feedback parameter lambda quantifies this self-correction. Higher lambda means stronger correction. Lower lambda means the system is more sensitive to forcing.
Van Loon, Rugenstein, Zelinka, and Andrews (arXiv:2603.12515) find that lambda has weakened significantly since the mid-1990s. Using a convolutional neural network trained on climate model ensembles and applied to observed sea surface temperature reconstructions, they show the climate system's self-correcting mechanism has become less effective over three decades. The mid-1990s represented a peak of stability. Since then, the thermostat has been loosening.
The culprit appears to be warming in the subtropical Northeast Pacific, confirmed by targeted perturbation experiments in climate models. This is not a model artifact — independent CMIP6 simulations corroborate the trend. The physical mechanism involves cloud and surface albedo feedbacks in the Pacific that become less stabilizing as the ocean warms in that region.
The counterintuitive implication: you might expect a warming planet to approach a new equilibrium — the feedback should strengthen as the system adjusts. Instead, the feedback is weakening. The planet is becoming more sensitive to forcing even as it warms. Each additional unit of forcing produces more warming than the same unit would have produced in the 1990s.
This connects to a structural pattern: systems that should self-regulate can lose their regulatory capacity precisely when regulation matters most. The thermostat weakens under the conditions that make thermostasis essential.
Van Loon, Rugenstein, Zelinka & Andrews, “Recent Weakening of the Global Radiative Feedback,” arXiv:2603.12515 (March 2026).
Note (2026-11-11): checked against the full text of the paper (Van Loon, Rugenstein, Zelinka and Andrews, arXiv 2603.12515). The numbers are as the essay says: the feedback parameter reached its most stabilizing value of about minus 3 W per square metre per kelvin around the mid-1990s and has weakened to about minus 2, estimated with a convolutional network trained on climate-model output and applied to observation-based temperature reconstructions, with warming in the subtropical Northeast Pacific an important driver confirmed in targeted E3SMv2 experiments. Five points differ. (1) The paper treats the feedback as varying on decadal time scales with changing surface-temperature patterns (the 'pattern effect'), computed over moving 30-year windows, and says each value can come from internal variability, external forcing or both; the essay reads the weakening as the planet 'becoming more sensitive to forcing', with each unit of forcing producing more warming than in the 1990s, which is the essay's inference about warming and not a statement the paper makes. (2) The essay says 'independent CMIP6 simulations corroborate the trend'. The paper's training and test data are CMIP5 and CMIP6 atmosphere-only and coupled model runs, and it says it confirms the result with climate-model simulations extended to 2022; it also reports a spread of about plus or minus 0.24 W per square metre per kelvin (99 percent interval) from reconstruction uncertainty in the satellite era, and notes the network gives a model-average pattern effect. (3) The mechanism in the essay, 'cloud and surface albedo feedbacks in the Pacific', is not what I found: the paper discusses low-cloud feedbacks tied to eastern Pacific marine cloud decks and a weakened positive feedback from the eastern Pacific, and I did not find a surface-albedo mechanism for this region. (4) The paper says the weakening is not significantly affected by El Nino or the Pacific Decadal Oscillation, which the essay omits. (5) The paper says previous estimates had already indicated weakening after 2014, and that coordinated model runs end in 2014, which is why it extends the estimate to 2025; the essay's closing pattern about systems losing regulation 'precisely when regulation matters most' is the essay's.