AI emissions: Why the real challenge is supplier disclosure

Every new report on the sustainability credentials of artificial intelligence (AI) appears to contradict the last. AI is either presented as an environmental catastrophe or a technological breakthrough that could help solve some of the world's biggest sustainability challenges.

As sustainability reporting professionals, our instinct is often to ask: what is the carbon footprint? But that’s not the only question we should be asking. 

Why AI emissions are difficult to estimate 

I explored whether we could estimate the emissions associated with ICAS’ use of AI using publicly available information.  

What seemed like a straightforward exercise quickly became more complicated. Different recognised approaches produced widely different estimates because they measured different things and applied different assumptions and boundaries.  

This is not a new problem. Sustainability professionals are accustomed to exercising judgement over estimates.  

Anyone involved in Scope 3 reporting will be familiar with the challenges of incomplete data, proxy factors, supplier disclosures and methodological choices. AI is the latest example of this broader issue, albeit one currently attracting significant attention. 

Each participant in the disclosure ecosystem has a different role, but each must engage with the complex assumptions and judgements underpinning disclosures. Preparers and boards must understand the assumptions behind these estimates. Auditors and assurance providers must assess whether those assumptions are reasonable. 

The technical complexity of AI and its supporting infrastructure shouldn’t be underestimated. Understanding the assumptions embedded within these estimates increasingly requires both sustainability and wider technical expertise. 

A familiar Scope 3 problem 

Sustainability reporting is intended to help users compare performance across organisations, sectors and over time.  

But if different organisations use different estimation techniques, proxy datasets and provider disclosures, can the resulting figures be meaningfully compared?  

Without more consistent methodologies and greater transparency from providers, there’s a risk that comparability, one of the fundamental aims of reporting, becomes increasingly difficult to achieve. 

AI highlights a challenge that runs through many Scope 3 emission disclosures. We have activity data, such as how many licences we purchase or how frequently a tool is used. But we lack the supplier-specific information needed to estimate our share of emissions with confidence.  

In that sense, AI is not creating an entirely new problem. It’s shining a light on one that sustainability professionals have been grappling with for years.  

There is increasing focus on the energy, emissions and water associated with AI services. There is also growing attention on impacts that are harder to measure, such as noise pollution, energy security and economic impact.  

But the emerging information also highlights the complexity of defining the boundary of AI.  

Should estimates consider only the energy used to generate a response? Or should they include supporting infrastructure, cooling systems, idle capacity and associated water consumption?  

Depending on the approach, results can vary significantly. 

What should AI providers disclose? 

Can organisations estimate AI-related emissions without information from AI providers? To some extent, yes.  

Spend-based methods, economic proxy factors and activity-based estimates can all be used. These approaches are familiar to anyone involved in Scope 3 emissions accounting.  

However, the more specific, accurate and comparable we want our estimates to be, the more dependent we become on information that only the provider can supply.  

A recent paper on AI emissions accounting notes that limited first-party data from infrastructure providers remains one of the largest constraints on accurate emissions accounting. It suggests that greater transparency from providers could replace assumption-driven estimates with more auditable information. 

In this respect, AI feels less like a new carbon accounting challenge and more like a familiar supplier disclosure challenge. 

The sustainability profession has spent years trying to improve transparency and data quality across supply chains. We routinely rely on suppliers to understand purchased goods and services, transportation, financed emissions and other value chain impacts.  

AI may be the latest manifestation of the same issue.  

The difference is that the underlying data is concentrated among a relatively small number of global technology companies. Their services are increasingly embedded within day-to-day organisational operations.  

Rather than debating whether an individual AI prompt emits a fraction of a gram of CO₂e or slightly more, perhaps we should ask what sustainability information organisations should reasonably expect AI providers to disclose. 

Should providers publish product-level emissions information? Should methodologies be more transparent and consistent? Should environmental metrics be independently assured?  

And are disclosures provided only at a company or facility level useful to organisations trying to understand their own share of AI-related emissions?  

As AI becomes a core business service, sustainability reporting frameworks may also need to consider what good disclosure looks like in this area.  

I don't pretend to have the answers. But as AI adoption grows, sustainability professionals will increasingly rely on data generated by external technology providers. 

The challenge isn’t simply to calculate AI emissions. It’s to secure transparent, consistent and reliable information that allows organisation to understand and compare them. If AI is to become a routine part of modern business, perhaps the next stage of the conversation isn’t about whether emissions can be estimated, but what sustainability information organisations should reasonably expect the providers of those services to disclose. That's something I've been reflecting on recently, and I suspect it is a discussion the sustainability reporting community will be having a lot more of in the years ahead.