Skip to main content
Bureau of AI
Category: Trustworthy AI Principles

Environmental Impact

Simply put

Environmental impact refers to any change to the environment that results, wholly or partially, from human activity, including a company's operations, products, or services. These changes can be either negative (harmful) or positive (beneficial). In the context of AI systems, this term is commonly applied to effects such as energy consumption associated with model development and use, though the specific scope depends on how it is defined in a given framework.

Formal definition

Environmental impact is defined as any alteration to the environment, whether adverse or beneficial, resulting wholly or partially from an organization's activities, products, or services. The concept encompasses both directions of effect (positive and negative) and may arise from human activity that creates environmental imbalance. Assessment of these consequences prior to a decision on a plan, policy, program, or project is typically conducted through an Environmental Impact Assessment (EIA). The precise boundaries, measurement methodologies, and applicable criteria for environmental impact vary by jurisdiction, sector, and governing framework, and are not standardized across all contexts.

Why it matters

Environmental impact has become a focal point in AI governance discussions because the development and operation of AI systems involve activities—such as energy consumption—that produce changes to the environment, whether adverse or beneficial. As organizations expand their use of AI, understanding and accounting for these effects is increasingly treated as part of responsible corporate activity, sustainability reporting, and broader environmental accountability. Because environmental impact can be either negative or positive, assessing it requires organizations to look at both the harms their activities may cause and any beneficial effects, rather than assuming impact refers only to damage.

The significance of the term also stems from its variability. The precise boundaries, measurement methodologies, and applicable criteria for environmental impact are not standardized across all contexts; they vary by jurisdiction, sector, and governing framework. This means that two organizations can describe their environmental impact very differently depending on which activities, products, or services they include in scope and how they choose to measure change. For professionals relying on exact language, this variability is a practical concern: claims about environmental impact are only meaningful when the underlying scope and methodology are made explicit.

Because of this lack of standardization, environmental impact is a concept that must be interpreted within its governing framework rather than as a fixed, universal measure. Treating it as a settled or comparable metric across organizations or jurisdictions risks overstating what a given assessment actually demonstrates.

Who it's relevant to

Sustainability and ESG Professionals
Those responsible for environmental reporting rely on a clear scope for what counts as environmental impact, since the concept spans both adverse and beneficial changes and its boundaries vary by framework. They must document which activities, products, or services are included and which measurement methodologies are applied, because these choices are not standardized across contexts.
AI Governance and Model Risk Practitioners
Professionals overseeing AI systems may encounter environmental impact in relation to effects such as the energy consumption of model development and use. They should treat the term functionally and specify the applicable framework, rather than assuming a single authoritative definition or measure applies across all jurisdictions and sectors.
Compliance and Policy Specialists
Those interpreting environmental requirements need to recognize that the criteria and boundaries for environmental impact differ by jurisdiction and governing framework. Where advance evaluation is required, an Environmental Impact Assessment may be the mechanism used, but its scope and methodology should be confirmed against the specific instrument that applies.
Auditors and Assurance Providers
When reviewing environmental impact claims, auditors must confirm the stated scope, direction of effect, and measurement methodology, because the absence of standardization means claims are only verifiable against the framework and boundaries the organization has explicitly defined.

Inside Environmental Impact

Compute and energy consumption
The electricity used to train and operate AI models, often cited as a primary driver of an AI system's environmental footprint. Estimates vary widely depending on model size, hardware efficiency, data center design, and duty cycle, and figures reported in the literature are not directly comparable across studies.
Carbon emissions (operational and embodied)
Greenhouse gas emissions associated with an AI system. Operational emissions arise from energy used during training and inference and depend heavily on the carbon intensity of the local electricity grid. Embodied emissions arise from manufacturing hardware and building infrastructure. These are distinct categories and are frequently conflated.
Water and resource use
Water consumed for data center cooling and resources associated with hardware manufacturing. These impacts are less commonly measured than energy use and reporting practices are not standardized.
Lifecycle scope
The boundary of what environmental impact assessment covers, which can range from a single training run to the full lifecycle including data collection, training, deployment, inference at scale, and hardware disposal. The chosen scope materially affects reported figures, so stating the boundary is essential.
Reporting and disclosure practices
The methods and metrics organizations use to estimate and communicate environmental impact. As commonly observed, there is no single universally adopted measurement methodology, and disclosures may be voluntary, inconsistent, or absent depending on jurisdiction and sector.
Governance intersection
Where environmental considerations enter AI governance structures, such as procurement policies, model approval criteria, and sustainability oversight. This is distinct from model risk management, which is concerned with risks arising from model use rather than environmental externalities, though the two may overlap in organizations that treat sustainability as a governed risk.

Common questions

Answers to the questions practitioners most commonly ask about Environmental Impact.

Does measuring a model's environmental impact just mean estimating the energy used during training?
No. Training energy is only one component. As commonly framed, environmental impact assessment spans the full model lifecycle, which can include data collection and storage, training, fine-tuning, and ongoing inference. For many deployed systems, cumulative inference over time can contribute substantially to total impact, so focusing only on training risks understating the overall footprint. The relative weighting varies by system, workload, and deployment scale, and there is no single authoritative methodology that applies across all contexts.
Is environmental impact an AI governance concern or a model risk management concern?
It can appear in both, and the two should not be collapsed. AI governance typically addresses environmental impact through organizational policies, accountability structures, disclosure practices, and oversight of whether such impacts are considered in system decisions. Model risk management, as historically framed, concerns risks arising from model use and would generally treat environmental factors only where they bear on the risks it covers. Environmental impact is not a traditional core element of model risk management guidance, so its treatment there is often organization-specific rather than settled practice.
What information do we need to collect to estimate the environmental impact of a model?
Estimation typically draws on inputs such as compute time, hardware type, energy consumption, and the carbon intensity of the electricity supplying the relevant data centers. Because these inputs are often not directly observable, organizations frequently rely on estimates or provider-reported figures, which introduces uncertainty. Documenting assumptions, data sources, and known gaps is generally more defensible than presenting a single precise number, and the appropriate level of rigor depends on the intended use of the estimate.
Who within an organization should be accountable for tracking environmental impact?
Accountability assignment varies by organizational structure and is not universally standardized. In many frameworks, environmental impact tracking is embedded within broader AI governance responsibilities, with roles clarified across the lines of defense so that those developing or operating systems, those overseeing them, and those providing independent assurance have distinct duties. The specific allocation should be defined in policy rather than assumed, and it may intersect with existing sustainability, procurement, or IT functions.
How can environmental impact be incorporated into model documentation?
It can be recorded alongside other lifecycle information, for example within model documentation or model cards, noting the scope of what was measured, the estimation method, key assumptions, and the boundaries of the assessment. Because methodologies differ and inputs are often estimated, documentation should state limitations explicitly rather than imply precision that the underlying data does not support. What to include is generally shaped by internal policy and any applicable disclosure expectations.
How does environmental impact assessment differ between training a new model and using an existing one?
The distribution of impact typically shifts. Building or training a model tends to concentrate impact at a point in time, whereas using an existing model concentrates impact in ongoing inference that accumulates with usage. This distinction can inform practical choices such as reuse versus retraining, though the appropriate approach depends on the specific system, workload, and how impact is being measured. No single rule determines which path has a lower footprint across all cases.

Common misconceptions

A model's environmental impact can be captured by a single headline number, such as the carbon cost of training.
Reported figures depend on the chosen lifecycle scope, grid carbon intensity, hardware, and assumptions. Training-only estimates typically omit inference at scale and embodied emissions. Without a stated boundary and methodology, headline numbers are not comparable across models or studies.
Training is where nearly all of an AI system's environmental impact occurs.
While training can be energy-intensive, inference performed repeatedly at scale over a deployed system's life can, in many cases, contribute substantially to total impact. The relative weight of training versus inference varies by system and usage, so neither should be assumed to dominate without measurement.
Environmental impact assessment is a settled, mandated component of AI governance and model risk management frameworks.
Environmental disclosure is treated unevenly and its regulatory treatment is evolving and jurisdiction-dependent. It is not a traditional element of model risk management as historically framed, and where it appears in governance it is often through separate sustainability or procurement policies rather than a universal requirement.

Best practices

State the assessment boundary explicitly, specifying whether reported figures cover training only, inference, or the full lifecycle including embodied hardware emissions, so results are interpretable and comparable.
Distinguish operational emissions from embodied emissions and report the carbon intensity of the electricity grid used, since operational impact varies significantly by location and time.
Use qualified, methodology-transparent language when disclosing figures, noting assumptions and known limitations rather than presenting single estimates as authoritative.
Account for inference and deployment-scale energy use alongside training, and revisit estimates as usage patterns change over the system's operational life.
Situate environmental considerations within the appropriate governance mechanism, such as procurement, sustainability oversight, or model approval criteria, without conflating them with model risk management controls that address risks arising from model use.
Track the evolving and jurisdiction-specific status of environmental disclosure expectations, and avoid presenting voluntary or proposed reporting practices as settled legal requirements.