Do we know how to measure AI’s environmental impact?
Artificial Intelligence (AI) is among the most disruptive technologies of our times, garnering worldwide attention and investment. Generative AI, for example, pulled in USD 33.9 billion worth of private investment globally last year – 18.7 per cent more than in 2023, according to the Stanford University AI index.
Adoption by businesses surged, with 78 per cent of organizations reporting AI use in 2024, up from 55 per cent the previous year in the same index.
By automating processes and enhancing predictive capabilities, AI could address a range of societal challenges, including environmental sustainability and climate action.
But what about AI’s own impact?
AI can reduce energy consumption, emissions and other environmental damage across numerous sectors. But using it that way involves trade-offs, as AI’s own footprint keeps growing.
The International Telecommunication Union (ITU) and partners in the Green Digital Action initiative aim to understand and quantify those trade-offs.
Measuring what matters
The ITU-led initiative includes a Green Computing pillar that directly addresses the impact of AI and big data centres. A newly formed Sustainable AI working group brings ITU – with decades of experience in setting environmental standards for technologies – together with public and private partners to drive further research on how to assess AI’s environmental impact.
The group released its new report – Measuring what matters: How to assess AI’s environmental impact – at ITU’s annual AI for Good Global Summit on 10 July.
The report starts out by identifying which components of AI’s impacts are actively being measured. Synthesizing academic, corporate and environmental research findings, it also considers the transparency, measurement practices, relevance and actionability of current methodologies.
It also notes limitations and recommends actions for three key groups: AI developers or producers; AI users or consumers; and digital policy makers.
Key findings
The working group’s report identifies the following gaps in current measurements of AI environmental impact:
- Over-reliance on estimates: Current methods often depend on proxies and indirect estimates, particularly during the training phase. Real-time, empirical data is rarely used, leading to inaccuracies.
- Underreported lifecycle phases: While AI model training is widely discussed, the inference phase, user behaviour and Scope 3 emissions (supply-chain impacts) remain underexplored.
- Opaque water-use and infrastructure impacts: Key metrics related to AI development, like water consumption in data centres, are poorly tracked.
- Lack of standardized methodologies: Existing tools and methodologies for AI impact measurement are fragmented and inconsistent, without standard definitions, reporting units, and evaluation frameworks to ensure reliable and actionable assessments.
- Carbon-centric metrics: Environmental assessments focus heavily on greenhouse gas (GHG) emissions, neglecting broader impacts associated with technology growth, such as biodiversity loss, electronic waste, resource depletion and resource contention.
The report urges AI developers to integrate sustainability from model design and training onwards. Assessments need to capture impact across the entire AI lifecycle – not just during training.
Further research is needed to support a more holistic approach to AI impact analysis. Lifecycle assessments could be applied for all impacts, while real-time telemetry tools would help in modelling the impacts of user behaviour across widely distributed networks, the report suggests.
ITU environmental standards
ITU standards provide key technical guidance on energy efficiency, green data centres and telecommunication site infrastructure, along with circular economy principles and other groundwork for greening digital technologies.
Building on this foundation, the ITU Telecommunication Standardization Sector (ITU-T) Study Group 5 (Environment, climate action, circular economy and electromagnetic fields) continues exploring a broad range of topics, including sustainable design principles for AI and extended reality (XR) systems, as well as water, energy and material-use efficiency in digital infrastructure.
A standard currently under development by the study group will help assess efficiencies and emissions in AI systems. Grounded in existing ITU methodologies, such as L.1410 (life cycle assessment) and L.1480 (enabling effects for other sectors), the new standard should support transparent, objective comparisons between different AI systems, or between AI and non-AI approaches. It may also provide an initial basis for environmental scoring.
Looking ahead
The Green Digital Action Sustainable AI working group has also developed a comprehensive measurement and testing plan to identify specific use cases, scenarios, and workloads to examine for a deeper understanding of AI’s environmental impact.
Policymakers, industry leaders, and researchers must agree on standardized methodologies for lifecycle assessments across all stages of AI systems.
Consumers, as end users of AI-driven products and services, can also demand sustainable practices.
ITU encourages everyone to support green computing, help shape international standards and join the global call for Green Digital Action.
Get involved in Green Digital Action
Learn more about Green Digital Action and contact green@itu.int to get involved.
Header image credit: Adobe Stock
