AI-Assisted Auditing of Climate Code and Statistical Methods: What Was Found, What Was Not, and Lessons for the Climate Community

Climate research increasingly rests on shared open-source software and on statistical analyses that few readers ever re-run. Over about two weeks in November 2026, the author, a retired atmospheric scientist, worked with a commercially available AI assistant to audit more than 40 widely used climate codes, datasets and high-profile papers. The AI wrote and ran code, rebuilt published analyses from raw data and searched the literature. The author chose the targets, judged every apparent error, and contacted developers and authors before anything was made public. Confirmed problems were found in about 15 targets. They include a regression in a standard drought index that returned minus infinity for dry months; a default in two trend-test packages that declares about half of trend-free series significant; a ranking error in a bias-adjustment package that inflated adjusted extreme rainfall several-fold for dry-biased models; a forecast-skill reference that overstated the value of initialization; a double-counted area factor in the ocean module of two simple climate models; and position-shifted duplicate ship reports still present in the main marine archive 35 years after they were first reported. Several fixes are now in released code, two of them as pull requests written by the author. Most targets were clean, several apparent errors did not survive checking, and one re-analysis largely confirmed the original paper while showing that its evidence had been overstated. The most effective techniques were re-running rather than reading, conservation and null-calibration checks, and contacting authors first. Expert verification remained essential throughout. AI-assisted auditing is now cheap enough to be routine for community software and high-impact studies.