SACRAMENTO, CA — As a fourth atmospheric river in three weeks dumped an estimated 11 inches of rain on the Central Valley Saturday, the state’s $48 million predictive flood model — trained almost exclusively on data collected between 2012 and 2022, a period California climatologists refer to as ‘the part where it didn’t rain’ — issued a confident green-status advisory indicating that conditions across the watershed remained ‘nominal to slightly arid.’
The model, developed in partnership with a Bay Area infrastructure firm and praised at its 2024 launch as ‘the future of climate resilience,’ continued to recommend voluntary water conservation as homeowners in Guerneville filmed their living rooms from the second floor.
‘The system is performing exactly as designed,’ said Priya Anand, a senior resilience architect at Beacon Strategy, which consulted on the rollout. ‘It was designed during a drought. It thinks droughts are what weather is. When you show it a flood, it interprets that as a very wet drought, which is still, statistically, a drought.’
Officials emphasized that the model is a ‘decision-support tool, not a decision-making tool,’ a distinction first introduced Saturday morning at 9:14 a.m., approximately ninety minutes after the levee in Wilton failed.
According to a Department of Water Resources spokesperson, the AI did flag one anomaly during the storm: an unusually high volume of what it categorized as ‘reflective surface in agricultural zones,’ which engineers later identified as Sonoma County.
The model’s training set, internal documents show, includes 4,200 days of recorded California weather, of which 11 contain meaningful precipitation and zero contain the phrase ‘atmospheric river,’ a term that did not enter common forecasting use until after the dataset was frozen. Asked Saturday what the system would call the current event, an engineer at the vendor paused for several seconds before suggesting ‘a humidity spike.’
Residents of affected counties expressed mixed feelings about the rollout. ‘My phone told me to take a shorter shower at 6 a.m.,’ said Doreen Vasquez, 58, of Rio Vista, speaking from a Red Cross cot. ‘Then a horse went past the window.’
The state has announced plans to retrain the model on more recent data once the more recent data stops happening. In the interim, a Cabinet-level working group has been convened to determine whether California is still in a drought, a flood, both, or some third category the model would find easier to process.
At a press availability outside a half-submerged pumping station near Modesto, a Department of Water Resources deputy director was asked whether the agency still had confidence in the system. He nodded, looked at his phone, and confirmed that the app was currently recommending he drink more water.