Major AI Climate Database Undercounts City Vehicle Emissions by 70%, Study Finds

High-resolution emissions data is the bedrock of modern climate policy. To cut carbon effectively, municipal governments must first know exactly where it is coming from, down to the specific neighborhood and highway interchange.[7]

For the past several years, the Climate TRACE consortium—co-founded by former US Vice President Al Gore—has been celebrated as a revolutionary tool in this effort. By combining satellite imagery with artificial intelligence, the platform promised to provide real-time, granular emissions data for the entire globe, bypassing the notoriously slow process of traditional carbon accounting.[4]

But a new peer-reviewed study has cast serious doubt on the accuracy of this AI-driven approach at the local level. Published in the journal Environmental Research Letters, the research claims that Climate TRACE is undercounting urban vehicle pollution by a massive 70.4% average across 260 United States cities.[1]

The study, led by Professor Kevin Gurney at Northern Arizona University, suggests that the world's most widely used climate emissions estimates could be missing far more pollution than anyone realized, raising red flags for policymakers who rely on the data.[2][5]

To understand the discrepancy, it is necessary to examine how carbon emissions are actually measured. The Northern Arizona University team compared the Climate TRACE data against the Vulcan Project, a government-funded database developed in Gurney's laboratory.[2][6]

Vulcan uses a "bottom-up" methodology. It calculates emissions by aggregating highly localized, physical data: municipal fuel consumption records, vehicle registration databases, road network topology, and travel demand models.[6]

Because Vulcan places the emissions exactly where the fuel was physically burned, its on-road uncertainty has been independently verified at approximately 14%. While not perfect, this margin of error is considered standard and reliable for high-fidelity physical modeling.[1][2]

Climate TRACE, by contrast, relies heavily on a "top-down" approach for its road transportation model. It takes large-scale data—such as national emissions inventories—and uses algorithms to scale that data down to the city level.[1][3]

According to the researchers, this is where the system broke down. The study identified that Climate TRACE applied a single, uniform "activity scale factor" of 0.3255 to align its on-road emissions with country-scale inventories.[1]

Applying a national average to granular urban environments fundamentally distorts reality, the researchers argue. The study found that in cities with heavy commuter traffic, such as Indianapolis and Nashville, the Climate TRACE estimates were more than 90% lower than the physical fuel-burn data recorded by Vulcan.[2]

The Climate TRACE consortium has strongly disputed the study's framing and conclusions. In a public response following the paper's publication, the organization stated that the analysis relied on an outdated version of their dataset.[3][4]

According to Climate TRACE, the data version analyzed by the researchers contained a temporary aggregation bug that affected city-level outputs. The consortium claims this bug was identified and patched in mid-2025, well before the study was published.[3]

Using their current, corrected data, Climate TRACE reports that its city road-transportation totals differ from the Vulcan database by only about 6% on average—a variance they describe as entirely normal between two independent emissions inventories.[3][4]

However, the scientific dispute has hit a verification roadblock. Gurney notes that since his team submitted their paper for peer review, Climate TRACE removed its original city-scale output from public access, replacing it with a broader county-scale dataset.[1][7]

Because a matching methodology document for this new county-scale output has not been published, independent researchers cannot directly verify if the underlying algorithmic flaw was actually fixed, or if the data was simply obscured by zooming out to a lower resolution.[1][5]

The stakes of this data dispute extend far beyond academic modeling. Municipal governments rely on high-resolution emissions baselines to make billion-dollar infrastructure decisions, from zoning for density to building electric vehicle charging networks.[5][7]

If a city planner believes their vehicle emissions are 70% lower than reality, their climate action plan will inevitably misallocate resources, under-investing in public transit while over-indexing on less impactful sectors.[7]

Ultimately, the controversy highlights a growing tension in climate science: the rush to deploy scalable AI monitoring versus the rigorous, slower process of physical ground-truthing. While artificial intelligence is essential for global visibility, researchers warn that it must be bound by strict scientific guardrails to ensure algorithms do not overwrite physical reality.[5][7]