UW researchers build AI agents that estimate electronic device carbon footprints in minutes

University of Washington researchers have developed an AI system that automatically estimates the carbon footprint of electronic devices like laptops and smartphones. The system uses two AI agents that work together to scrape public data from sources like FCC databases and iFixit teardowns, then reference life cycle assessment databases to convert component lists into carbon estimates. It achieves an average error rate of 5% to 19%, similar to human experts, but completes the analysis in about a minute instead of days or months. The team also developed a nearest neighbors method to estimate carbon footprints for devices or materials without direct data. For materials, this approach had an average error of 23%, compared to 143% for human experts picking the closest entry. The researchers note that running the AI models generates emissions roughly equivalent to brewing a cup of tea per device estimate. They plan to work with companies to automate sustainability workflows.