Carbon-neutral AI requires rethinking how intelligence is measured

Artificial intelligence is rapidly embedding in daily economic and institutional activity. While its environmental impact has drawn growing scrutiny, most assessments still rely on broad averages that mask how different uses of AI drive sharply different energy demands. A new peer-reviewed study finds that this lack of granularity is leading organizations to underestimate both the scale and the structure of AI’s carbon footprint.

The study, titled The Carbon Cost of Intelligence: A Domain-Specific Framework for Measuring AI Energy and Emissions, published in the journal Energies, introduces a new framework for measuring AI energy use and carbon emissions that accounts for what types of tasks AI systems are performing, rather than treating all queries as equal.

The findings challenge the prevailing assumption that AI sustainability can be evaluated using a single average metric. Instead, the research shows that energy consumption varies dramatically depending on the knowledge domain, workload mix, and deployment context.

Why one-size-fits-all AI energy metrics no longer work

Most benchmarks report average energy per query or per model, obscuring how real-world deployments operate across multiple domains with very different computational profiles. In practice, AI systems handle a mix of medical, legal, financial, technical, and general knowledge tasks, each with distinct energy demands.

Using GPT-4 as a case study, the researchers conducted a cross-domain benchmark based on the Massive Multitask Language Understanding dataset, measuring both accuracy and energy consumption across five representative domains: medicine, finance, law, computer science, and general knowledge. The results show that inference energy can vary by more than fourfold depending on the domain.

Legal queries emerged as the most energy-intensive, consuming over four times more energy per query than general knowledge tasks. This difference is driven largely by input complexity, with legal prompts tending to be longer and more detailed. By contrast, general knowledge queries achieve high accuracy with comparatively low energy use, making them the most efficient domain in accuracy-to-energy terms.

The study finds that inference, not training, dominates AI’s operational energy footprint. While training large models has historically attracted attention due to its high upfront emissions, inference operations account for the majority of ongoing energy use once models are deployed at scale. As AI adoption accelerates, especially in enterprise and public-sector settings, inference energy increasingly shapes AI’s long-term environmental impact.

The research also highlights a flaw in common aggregation methods. Simple weighted averages of accuracy and energy tend to overestimate efficiency when combining domains with widely different profiles. High-efficiency domains disproportionately skew the results, masking the carbon burden of more energy-intensive tasks. This bias becomes especially problematic for organizations with heterogeneous workloads.

To address this, the authors introduce a new metric known as the Carbon Cost of Intelligence, or CCI. The metric uses a weighted harmonic mean to aggregate accuracy-to-energy ratios across domains, producing a more conservative and accurate estimate of energy efficiency for a given workload mix. This approach mirrors methods used in financial analysis to prevent distortion from extreme values.

Workload mix drives AI emissions more than model choice

Workload composition has a greater impact on AI energy use than many commonly discussed factors, including model architecture alone. By simulating different real-world deployment scenarios, the researchers show how the same AI model can produce vastly different carbon footprints depending on how it is used.

A hospital-style workload dominated by medical queries consumes nearly half the energy per query of a law-firm-style workload dominated by legal tasks. This difference translates into a near doubling of carbon emissions per query under identical infrastructure conditions. Across scenarios, the study finds that optimizing workload composition could reduce emissions by up to 49 percent without eliminating AI usage in high-energy domains.

The analysis further demonstrates that accuracy does not scale linearly with energy consumption. Some domains achieve high accuracy at relatively low energy cost, while others consume significantly more energy without corresponding gains in performance. This finding challenges the assumption that higher energy use necessarily delivers better outcomes and underscores the importance of domain-aware optimization.

The study also examines how aggregation methods influence sustainability planning. In scenarios where domain efficiencies vary widely, simple averaging methods overestimate efficiency by as much as 12 percent compared to the harmonic mean approach. This gap is large enough to undermine carbon neutrality targets and mislead regulatory disclosures if not corrected.

Importantly, the researchers point out that workload optimization does not imply reducing critical AI functions in domains such as law or healthcare. Instead, the framework enables organizations to refine how these workloads are handled, through measures such as prompt optimization, staged querying, caching, and selective model routing. These changes preserve functional requirements while reducing unnecessary energy expenditure.

Geography further amplifies these differences. Carbon emissions per query vary significantly depending on data center location and grid carbon intensity. The same AI workload can produce several times more emissions when served from regions with carbon-intensive electricity compared to regions with cleaner grids. The framework allows organizations to parameterize calculations based on local energy mixes, making it applicable across global deployments.

A practical framework for carbon-neutral AI deployment

The Carbon Cost of Intelligence is a practical tool for organizations seeking to align AI deployment with climate commitments. The framework enables companies to calculate per-query energy use, carbon emissions, and monthly or annual totals based on their specific workload distributions.

This capability addresses a growing regulatory and corporate governance gap. As sustainability reporting standards evolve, organizations face increasing pressure to disclose the environmental impact of digital operations, including AI. Aggregate metrics are unlikely to meet emerging expectations for transparency and accuracy. Domain-specific measurement offers a path toward defensible reporting and informed decision-making.

The authors argue that AI sustainability must move beyond abstract discussions of model size or training costs. Instead, it requires operational metrics that reflect how AI systems are actually used. By linking energy consumption to workload composition, the framework supports strategic choices about where optimization efforts will yield the greatest carbon savings.

The study also contributes to broader debates about the role of AI in the energy transition. While AI is often framed as both a driver of energy demand and a tool for efficiency gains, the research shows that these outcomes depend heavily on deployment patterns. Without domain-aware measurement, efficiency gains in one area may be offset by unrecognized energy growth in another.

The researchers acknowledge limitations, including reliance on benchmark-derived energy estimates and a relatively small sample size per domain. However, they emphasize that the framework is model-agnostic and can be applied to other large language models as energy measurement data become available. Future research is expected to expand the approach to additional domains, larger datasets, and dynamic workloads that change over time.

Despite these limitations, the findings mark a shift in how AI’s environmental impact can be understood and managed.