For sustainability professionals tasked with the monumental challenge of tracking corporate emissions, artificial intelligence has long been heralded as the "holy grail." The promise is seductive: by simply inputting a bill of materials or a brief product description, AI platforms can instantly generate a Product Carbon Footprint (PCF), bypassing the months of tedious data collection, supplier outreach, and spreadsheet management that currently define the industry.
However, a sobering new study from researchers at Watershed—a leading carbon accounting platform—has cast a long shadow over this optimism. While AI models can arrive at impressively accurate final numbers by chance, their underlying logic is often fundamentally flawed. The study suggests that relying on these automated tools without rigorous oversight is not just a shortcut; it is a potential liability that could derail genuine decarbonization efforts.
The Main Facts: The Promise vs. The Reality
The core appeal of AI in carbon accounting lies in efficiency. Currently, measuring a product’s footprint—the total greenhouse gas emissions generated from raw material extraction to manufacturing—is an arduous task. It requires granular data on every component, energy usage, and logistical step. According to data from PwC, this complexity has resulted in a stagnation of progress: 69 percent of companies have managed to conduct PCFs for less than a quarter of their entire product portfolios.
Companies like Makersite and Terrascope have entered the market promising to disrupt this status quo. Makersite claims it can automate "accurate LCAs [Life Cycle Assessments] across your entire product portfolio in seconds," while Terrascope asserts that it can achieve 70 percent accuracy without requiring direct data input from suppliers.
Yet, the Watershed study, led by Krishna Rao, suggests that the "black box" nature of these models hides a significant deficit in reliability. While AI models from industry giants like Anthropic, DeepSeek, Google, and OpenAI could generate a final PCF estimate within a reasonable margin of an expert-calculated figure 77 percent of the time, the process by which they arrived at those figures was often riddled with errors.
A Chronology of the Research
The Watershed team sought to move beyond anecdotal claims of AI efficacy by conducting a systematic benchmark. Their study followed a structured, multi-phase methodology:
- Phase 1: Dataset Construction: The researchers compiled a diverse set of 175 products spanning complex sectors, including chemicals, textiles, electronics, and raw materials.
- Phase 2: The "Expert Baseline": For every product in the dataset, the team established an "expert answer"—a verified, scientifically rigorous PCF calculation based on verified industry databases.
- Phase 3: The AI Challenge: The team prompted state-of-the-art Large Language Models (LLMs) to perform the same task. The models were asked to break down the products into constituent parts and assign emissions values to each step of the manufacturing lifecycle.
- Phase 4: Evaluation of Logic: The researchers compared not just the final total, but the intermediate steps—the "reasoning" the AI used to build its estimate.
- Phase 5: The Gap Analysis: The team published their findings in August 2026, revealing a startling disconnect between the models’ "lucky" final guesses and their flawed internal reasoning.
Supporting Data: When Accuracy Is a Coincidence
The most alarming finding from the study is the discrepancy between final results and procedural accuracy. When the models were asked to simply output a total number, they performed surprisingly well. However, when the researchers forced the models to show their work—decomposing the product into its component materials and estimating the emissions for each—the accuracy plummeted to as low as 37 percent.
"What surprised me most was the size of the gap," says Rao. "We saw instances where the AI would overestimate the footprint of a raw material by a factor of ten, but then underestimate the energy required for assembly by the same margin. The two errors canceled each other out, leaving a final number that looked plausible, but was based on entirely wrong premises."
This phenomenon, often referred to in AI research as "hallucination," suggests that LLMs are not actually performing life cycle assessments. Instead, they are using statistical pattern matching to predict what a "typical" footprint for a product might look like, rather than calculating the unique footprint of the specific product in question.
Official Responses and Industry Stance
The industry response to the study has been nuanced. Providers of AI-powered sustainability tools argue that these models are still in their infancy and that the speed they offer provides a "good enough" starting point for companies that previously had zero data.
"Our goal is to provide directional guidance," says a spokesperson for a leading carbon accounting tech firm. "In a world where companies are operating with 0% visibility into their scope 3 emissions, an AI estimate that is 70% accurate is a massive leap forward compared to an educated guess on a napkin."
However, Watershed’s study serves as a necessary reality check. By making their testing methodology open-source, the researchers are effectively inviting other firms—including their own—to subject their tools to this level of scrutiny. The implication is clear: transparency in the "intermediate steps" is the only way to build trust in these systems.
Implications for Corporate Sustainability
The implications of these findings are profound for the C-suite and sustainability departments.
1. The Danger of Misguided Decisions
The primary purpose of a PCF is not just to report a number for a sustainability report; it is to identify "hotspots" where emissions can be reduced. If an electronic manufacturer decides to switch vendors for a specific chip because an AI model flagged it as a high-carbon component, but that flag was a result of an AI error, the company has wasted time and capital. They may have inadvertently switched to a more carbon-intensive vendor while thinking they were doing the opposite.
2. The Rise of "Audit-Ready" Requirements
As regulatory bodies—such as the SEC in the U.S. and the EFRAG in the EU—tighten their requirements for climate disclosure, "black box" numbers will likely be rejected. Companies will be expected to produce an audit trail. If a company cannot explain how it arrived at its product emissions, those figures will be effectively worthless in a compliance audit.
3. The Need for "Human-in-the-Loop"
The study suggests that AI should be treated as a junior assistant rather than an expert consultant. Sustainability professionals must maintain a "human-in-the-loop" workflow. This means using AI to handle the tedious data ingestion, but requiring expert analysts to verify the logic and the intermediate data points before the final footprint is integrated into corporate strategy.
4. The Future of Benchmarking
Rao’s work has established a new standard for how these tools should be evaluated. Moving forward, providers of carbon accounting software will likely be forced to offer more transparency into their algorithms. Companies purchasing these services should now demand "proof of reasoning" rather than just "proof of results."
Conclusion: Caution Over Convenience
The automation of Product Carbon Footprints is a tantalizing goal, and there is no doubt that AI will eventually play a critical role in decarbonizing the global economy. The ability to process millions of data points across global supply chains is a task for which humans are ill-equipped.
However, the Watershed study serves as a critical reminder that when it comes to sustainability, the path to the result is just as important as the result itself. If we accept "impressive-sounding" numbers without interrogating the logic behind them, we risk building our climate strategies on a foundation of digital sand. For now, the best advice for any organization utilizing AI for sustainability reporting is simple: use the tools to save time, but never delegate the thinking.
