We mislead decision-makers and potentially lose public trust
In the effort to map the invisible weight humanity places on the atmosphere, precision is not merely a technical virtue — it is a moral one. A team at Northern Arizona University has found that Climate TRACE, one of the world's most prominent AI-driven emissions databases, may be undercounting vehicle carbon dioxide in American cities by roughly 70 percent on average, with some cities showing gaps exceeding 90 percent. When the instruments we use to diagnose a crisis are themselves miscalibrated, the remedies we design risk missing the wound entirely. The findings arrive at a moment when governments are staking billion-dollar climate decisions on exactly this kind of data.
- A 70 percent undercount in urban vehicle emissions is not a rounding error — it is a distortion large enough to redirect entire climate strategies toward the wrong targets.
- Indianapolis and Nashville appear in the Climate TRACE data at more than 90 percent below rival estimates, signaling that the problem is structural, not incidental.
- This is not Climate TRACE's first credibility challenge: Gurney's team previously found similar miscounts in the database's power plant figures, raising the specter of a systemic AI blind spot across fossil fuel categories.
- Cities and nations acting on underestimated emissions data may chronically underinvest in transit, congestion pricing, and other interventions precisely where they are most needed.
- The researchers are not calling for the project's abandonment but for the scientific guardrails — expert review, transparency, and rigorous methodology — that any tool shaping public policy demands.
- As AI becomes the backbone of global climate monitoring, the stakes of getting the fundamentals right have compounded far beyond any single database.
Kevin Gurney's research team at Northern Arizona University has identified a significant and consistent flaw in Climate TRACE, the high-profile emissions inventory co-founded by Al Gore. Published in Environmental Research Letters, their study found that the database underestimates vehicle CO2 emissions in American cities by an average of 70 percent when compared against Vulcan, a rival database built on official traffic counts and fuel consumption records with an uncertainty margin of around 14 percent.
Across 260 U.S. cities, the gap was not random noise but a persistent pattern. In Indianapolis and Nashville, Climate TRACE's figures came in more than 90 percent below Vulcan's estimates. Postdoctoral researcher Bilal Aslam noted the undercount's consistency across the sample, while research associate Pawlok Dass raised the possibility that the same distortions may extend beyond U.S. borders.
The likely culprit is Climate TRACE's reliance on artificial intelligence to generate emissions estimates at scale — a promising approach that, the researchers argue, has been deployed without sufficient scientific discipline. This is not an isolated finding: Gurney's team previously flagged comparable problems in how Climate TRACE handles power plant emissions, suggesting that AI-driven undercounting may be affecting more than half of fossil fuel emissions tallied for American cities.
Gurney was careful to acknowledge that perfect accuracy is unattainable, but insisted the current gap falls far short of acceptable standards. Without reliable data, he warned, policymakers are misled and public trust in climate science erodes. A city that believes its vehicle emissions are lower than they are may underinvest in public transit or congestion pricing; a nation may misallocate climate finance on a massive scale.
The study stops well short of dismissing Climate TRACE's mission — the world urgently needs better emissions monitoring — but it makes clear that powerful new tools require the same rigorous review and transparency that science has always demanded. As governments lean ever more heavily on AI-powered climate systems, the cost of miscalibration is no longer abstract.
Kevin Gurney's team at Northern Arizona University has uncovered a troubling gap in one of the world's most influential climate databases. The Climate TRACE emissions inventory, co-founded by former Vice President Al Gore, is systematically undercounting the carbon dioxide that cars and trucks pump into city air—by roughly 70 percent on average, according to research published in Environmental Research Letters.
This matters because governments and cities are increasingly using high-resolution emissions data like Climate TRACE to design climate policy and measure progress toward their carbon goals. If the numbers are wrong by that magnitude, the policies built on them may be solving the wrong problem, or solving it at the wrong scale. Gurney, a professor in Northern Arizona University's School of Informatics, Computing, and Cyber Systems, compared Climate TRACE's vehicle emissions estimates against Vulcan, a competing database developed in his own lab that relies on official traffic counts and fuel consumption records. Vulcan carries an uncertainty margin of about 14 percent. The gap between the two databases was far wider.
When the researchers examined 260 American cities, they found that Climate TRACE's vehicle emissions figures ran roughly 70 percent lower than Vulcan's estimates. Some cities showed even starker discrepancies. Indianapolis and Nashville each came in more than 90 percent lower in the Climate TRACE data. The pattern suggests the problem is not confined to a handful of outliers but baked into the system itself. Bilal Aslam, a postdoctoral researcher on the study, noted the consistency of the undercount across the sample. Pawlok Dass, a research associate, flagged the possibility that similar undercounting may be occurring outside U.S. borders, potentially affecting climate assessments worldwide.
The root of the problem appears to lie in Climate TRACE's reliance on artificial intelligence to estimate emissions. While AI offers promise as a tool for environmental monitoring at scale, the researchers argue it requires scientific discipline to work reliably. The study is not the first to find problems with Climate TRACE's methodology. Gurney's team previously identified comparable issues in how the database estimates power plant emissions. Taken together, these findings suggest that AI-driven climate monitoring systems may be systematically underestimating more than half of fossil fuel emissions in American cities.
Gurney emphasized that perfect accuracy in emissions accounting is impossible, but the standard should be higher than what Climate TRACE currently delivers. "We must ensure that the data shared with policymakers and the public is unbiased and meets best practices and the most rigorous scientific standards available," he said. "Without this, we mislead decision-makers and potentially lose public trust in our ability to tackle climate change." The researchers offered recommendations for improving Climate TRACE's methodology, stressing the need for expert review, transparency, and adherence to established scientific practices.
The implications ripple outward. Cities and nations making billion-dollar decisions about where to invest in emissions reductions may be working from a distorted picture of where the problem actually lies. A city that thinks its vehicle emissions are lower than they truly are might underinvest in transit infrastructure or congestion pricing. A nation might misallocate climate finance. The study does not question the importance of the Climate TRACE project itself—the world needs better emissions monitoring—but it does expose the risk of deploying powerful new technologies without the guardrails that science demands. As governments lean more heavily on AI-powered climate tools, the pressure to get the fundamentals right has never been greater.
Notable Quotes
When combined with our previous study on Climate TRACE power plant CO2 emissions, our results suggest that the Climate TRACE data significantly underestimate over half of U.S. fossil fuel-based CO2 emissions in cities.— Kevin Gurney, Northern Arizona University
We must ensure that the data shared with policymakers and the public is unbiased and meets best practices and the most rigorous scientific standards available. Without this, we mislead decision-makers and potentially lose public trust in our ability to tackle climate change.— Kevin Gurney
The Hearth Conversation Another angle on the story
Why does a 70 percent undercount matter so much? Isn't some data better than no data?
It depends on what you do with it. If a city thinks its vehicle emissions are 30 percent of what they actually are, it might decide the problem is smaller than it is and spend less money fixing it. You're not just getting a number wrong—you're potentially misdirecting billions in climate spending.
How did this slip through? Climate TRACE is a major project with serious backing.
The database relies on artificial intelligence to estimate emissions from satellite imagery and other remote data. AI is powerful, but it needs to be checked against ground truth—actual traffic counts, fuel sales, that kind of thing. Climate TRACE apparently didn't do enough of that validation work before publishing.
Is this a failure of the whole approach, or just Climate TRACE?
It's a warning about the approach. AI can monitor environmental systems at a scale humans never could before. But if you deploy it without rigorous scientific review, you get confident-sounding numbers that are actually quite wrong. The technology isn't the problem. The lack of guardrails is.
What does Gurney's own database, Vulcan, do differently?
Vulcan is built on official data—traffic counts from transportation agencies, fuel consumption records from energy companies. It's calibrated to reality from the ground up. It's not perfect either, but its uncertainty is known and much smaller. That's the difference between AI that's been validated and AI that hasn't.
Could this affect other parts of Climate TRACE, or just vehicles?
The researchers already found similar problems with power plant emissions estimates. They suspect the issue is broader. When you're using the same methodology across an entire database and it's failing in multiple places, you have to ask whether the whole system needs rethinking.