ALERTCalifornia AI Wildfire Detection System Generated Excessive False Alarms

The ALERTCalifornia network, operated by UC San Diego in partnership with California utilities and fire agencies, deployed over 1,000 AI-powered cameras across the state to provide early wildfire detection. The system uses computer vision algorithms to analyze camera feeds in real-time, automatically flagging potential fire signatures for human review. However, throughout 2022, the system generated significant numbers of false positive alerts, particularly during foggy conditions and periods of industrial activity. The AI models struggled to distinguish between actual wildfire smoke and visually similar phenomena including marine layer fog rolling inland, low-hanging clouds in mountainous areas, dust from construction or agricultural activities, and emissions from industrial facilities. In some regions, false alarm rates reached 70-80% of all AI-generated alerts during certain weather conditions. Fire departments reported that these false alarms required dispatch of personnel and equipment to investigate, diverting resources from other emergency responses. The high sensitivity settings, while designed to ensure no actual fires were missed, created operational challenges for fire agencies already stretched thin during California's extended fire seasons. Some departments reported responding to multiple false alarms per day during peak fog season, with crews traveling significant distances only to find clear conditions. The false alarms also triggered unnecessary public warnings in some cases, causing confusion and eroding public trust in the alert system. In response to these challenges, UC San Diego researchers worked throughout 2022 and 2023 to refine the AI algorithms, incorporating additional training data that included common false positive scenarios. They also enhanced the human verification process, requiring trained operators to confirm AI detections before issuing public alerts or dispatching resources. Weather data integration was improved to help the system account for atmospheric conditions that commonly cause false readings. By late 2023, system operators reported that algorithm improvements and enhanced human oversight had reduced false alarm rates to approximately 30-40%, though this remained higher than desired. The incident highlighted the challenge of balancing sensitivity in safety-critical AI systems, where missing a real fire could have catastrophic consequences, but excessive false alarms undermine system effectiveness and waste critical resources.