Accounting for AI Inference in Corporate GHG Inventories: A Four-Tier Methodology for Scope 3 Category 1 Reporting
This paper proposes a standardized four-tier methodology for accurately reporting AI inference emissions under Scope 3 Category 1 of the CSRD, replacing inaccurate generic economic factors with precise, token-based physical calculations that reveal significantly lower carbon footprints and highlight critical water-carbon trade-offs in data center location strategies.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine your company is like a large household. For years, you've been tracking how much electricity you use for your lights and fridge (Scope 1 and 2). But now, you've started buying "smart services" from outside companies—like AI chatbots, automated writing tools, and smart assistants.
The new European rules (CSRD) say you must also report the environmental cost of these outside services. The problem? No one knows how to measure the "carbon footprint" of a digital conversation.
This paper, written by Guillermo Llopis, acts as a user manual for solving this mystery. It argues that current methods are like trying to weigh a feather by guessing its weight based on the price of a bag of sand—they are wildly inaccurate. Instead, the author proposes a four-tier ladder of accuracy, depending on how much data your company can actually find.
Here is the breakdown of the paper's core ideas using simple analogies:
1. The Problem: The "Bag of Sand" Mistake
Currently, many companies guess the carbon cost of AI by looking at how much they spent. They use a generic formula that says, "Every euro spent on computer services creates X amount of pollution."
The paper calls this the EEIO method. It's like trying to calculate how much water a specific shower uses by looking at your total water bill and the price of a plumber's visit. It includes the cost of the plumber's time, the truck, and the profit, not just the water.
- The Result: This method overestimates AI pollution by 10 to 40 times. It's like saying your digital chatbot is as dirty as a coal power plant just because you paid for it.
2. The Solution: The Four-Tier Ladder
Instead of guessing, the paper suggests climbing a ladder. You start at the top (most accurate) and only go down if you don't have the data.
- Tier 3 (The Gold Standard): The AI provider gives you a verified, audited report saying, "This specific service used exactly this much energy."
- Status: Currently, no one offers this yet. It's the "dream tier."
- Tier 2a (The Receipt Check): You have access to the AI's billing portal. You can see exactly how many "tokens" (the digital words the AI processed) you used.
- How it works: You take the token count, multiply it by a known energy benchmark (like a car's MPG), and multiply that by the local power grid's cleanliness.
- Analogy: It's like looking at your gas pump receipt to see exactly how many gallons you put in, rather than guessing based on your car's make and model.
- Tier 2b (The Best Guess): You don't have the token count, but you know how many messages were sent or how many people used the tool.
- How it works: You estimate the tokens based on the number of messages (e.g., "1 message = 400 tokens").
- Analogy: It's like estimating your gas usage by counting how many times you drove to the store, assuming a standard trip length. It's less precise, but better than nothing.
- Tier 1 (The Last Resort): You only know how much you paid for a software subscription that includes AI, but you can't see the AI usage at all.
- How it works: You use the old, inaccurate "spend-based" method.
- Warning: The paper says this number is likely a massive overestimate (like the "bag of sand" method). You should report it as a "worst-case scenario" rather than a fact.
3. The Big Surprise: Location Matters More Than Size
The paper discovered a counter-intuitive fact: Where the AI lives matters more than what the AI does.
- The Analogy: Imagine two identical cars. One is driven in Sweden (where the electricity comes from clean hydro power), and the other in Singapore (where electricity comes from fossil fuels).
- The Finding: The car in Singapore pollutes 13 times more than the one in Sweden, even if they drive the exact same distance.
- The Lesson: A company using a "small" AI model in a dirty grid region might pollute more than a company using a "giant" AI model in a clean grid region.
4. The Hidden Trade-Off: The "Clean but Wet" Paradox
The paper also looked at water usage. Data centers need water to cool down, and making electricity often uses water too.
- The Paradox: Sweden has the cleanest air (lowest carbon) because of its hydroelectric dams. But, those dams use a massive amount of water.
- The Result: Sweden is the "greenest" place for carbon, but the "wettest" place for water usage. Ireland, on the other hand, is a sweet spot: it has decent carbon numbers and uses very little water.
- The Takeaway: If you only look at carbon, you might pick the wrong location. You have to look at both carbon and water to make a truly green choice.
5. The Bottom Line: It's a Paperwork Problem, Not a Monster Problem
The paper ran a simulation with a typical 200-person European company.
- The Result: The total pollution from their AI usage was less than 1 ton of CO2.
- The Context: That is less than the pollution from a few employees commuting to work for a month.
- The Real Challenge: The emissions aren't huge, but the rules are strict. Companies cannot say, "We didn't know how to measure it." They must use the ladder method to find the best data they can.
In summary: This paper gives companies a clear, step-by-step recipe to stop guessing and start measuring their AI pollution accurately. It warns them that spending money doesn't equal pollution, that location is the biggest lever they can pull, and that they need to watch out for hidden water costs. The data is already there; they just needed a map to find it.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.