Why AI carbon footprinting tools are prone to misleading results

A new study looked at AI tools that estimate the carbon footprint of manufactured products, and found the results can be misleading even when they look accurate on the surface. Researchers from Watershed, a carbon accounting company, tested AI models from Anthropic, DeepSeek, Google, and OpenAI on 175 products across materials, chemicals, and textiles. The best model landed within a reasonable range of expert estimates 77 percent of the time. But when researchers checked the model's actual work — how it broke a product into parts and estimated the emissions from each piece — accuracy dropped to as low as 37 percent. In other words, the AI often got the right answer for the wrong reasons.

This matters because these tools are being pitched as a fast way to skip a slow, manual process. Normally, figuring out a product's carbon footprint means tracking down data on every material and component, often from multiple suppliers, which is why most companies have only done this for a small share of their products. AI tools promise to shortcut that work, sometimes claiming near-instant results. But the study suggests that trusting the final number alone is risky, since flawed data on individual components can hide inside an otherwise plausible-looking total.

This isn't a story about home energy upgrades directly, but it's a reminder for anyone researching the environmental footprint of products, appliances, or building materials using AI tools: a confident-sounding answer isn't the same as a reliable one. If you're comparing the embodied carbon of materials for a renovation, it's worth checking how a figure was calculated, not just what the final number says.