AI carbon footprint: governance checklist for energy and environmental accountability

The story: a checklist for AI's environmental impact

On September 1, 2026, a group of eight researchers published an operational checklist for measuring energy consumption and carbon footprint of machine learning and artificial intelligence applications in Earth System Modeling. The document appears on arXiv with identifier 2609.00847^[1].

The checklist organizes questions along successive phases of the model development pipeline.

The stated objective consolidates principles scattered across many publications. The authors translate ethical and environmental recommendations into concrete metrics. Each item references examples and suggestions drawn from recent literature.

Eight authors sign the work, with Filippo Dainelli as the lead author. The research emerges from the Earth System Modeling community, a domain with high computational intensity, where climate models consume considerable computing resources and AI amplifies that demand.

Why computational consumption enters governance

Training large-scale AI models consumes energy in significant quantities. Each iteration of experimentation multiplies consumption, with direct effects on costs and emissions.

Organizations that declare sustainability goals face a concrete tension. AI research accelerates, computational demand grows, environmental commitments remain fixed. The checklist offers a method to reconcile these forces with verifiable data.

Measurement precedes control: you manage what you quantify, and this tool makes every phase quantifiable.

From scattered principles to operational list

Before this document, environmental best practices for AI lived fragmented. Studies and commentaries exposed them in scattered form, difficult to apply to daily work.

The checklist changes the perimeter. It consolidates a body of knowledge into a verifiable tool, phase by phase. The relevant delta concerns operationality: it shifts from aspiration to a checklist.

The governance signal: AI sustainability becomes a matter of process, measurable and assignable. A generic ESG principle acquires precise metrics.

Frameworks that name what to measure and when to measure it produce governance. Those that list general principles produce intentions. The document also establishes common language between climate scientists and machine learning specialists.

Who is accountable for environmental footprint

The structural question remains identical in every governance framework. Which named role within the organization is accountable for the model's environmental footprint, by name, in writing, before deployment?

The checklist provides the metrics. Assigning the role falls to the organization. Accountability without a name remains documentation, distant from real governance.

Teams that connect each checklist item to a specific owner obtain a defensible audit trail. Teams that file the checklist as an attachment obtain paper.

A compliance posture calibrated to generic ESG principles appears oversized for the new operational context. Audit remains required, but the perimeter shifts toward measured metrics.

What technical teams verify

Development teams face choices at every phase with direct environmental impact. Architecture selection, model size, number of experiments and computing infrastructure all weigh on final consumption.

The checklist guides these choices with operational questions. It indicates where to reduce consumption and where to document measurement. Each phase produces data that feeds upstream reporting.

The cited metrics estimate project energy consumption and emissions, linked to concrete examples for each question. The complete version and associated discussions are also available on alphaXiv^[2].

The document spans 12 pages, one figure and two tables, presented at the GREEN-AI workshop of the ECML PKDD 2026 conference in Naples. This academic context defines the status: peer-reviewed tool in a research environment.

The predictable objection

A recurring objection views these checklists as a formal exercise. The criticism holds that environmental documentation produces paper, distant from results.

The reply lies in structure. A checklist linked to metrics and named owners generates verifiable evidence. The difference between compliance theater and real governance lies precisely in this assignment of responsibility.

Organizations that adopt the tool with defined owners transform obligation into advantage. Those that treat it as a formality obtain an inert archive.

Three decisions for the board and the European framework

Three decisions for the board. Here are the questions that General Counsel, Chief Risk Officer and audit committee must resolve now.

1. General Counsel: what exposure emerges from public sustainability declarations compared to actual consumption of AI models?
2. Chief Risk Officer: which line item in the risk framework incorporates energy footprint as a measured metric?
3. Board Audit & Risk Committee: which environmental disclosure on AI projects proves verifiable with documented evidence?

Each decision requires a named owner. The framework provides for assignment before deployment, consistent with checklist logic.

The European framework adds weight to this tool. The EU AI Act imposes transparency and technical documentation obligations on providers of high-risk AI systems.

A European Commission decision on AI Act implementation defines enforcement timelines and scope. Environmental metrics enter corporate reporting through ESG directives. The checklist offers a ready tool for populating that reporting with measured data.

The EU AI Act proceeds in phases of application, with staggered deadlines through 2026 and beyond. This checklist maintains the character of a voluntary technical tool, distinct from legal obligation.

Regulatory horizon

Regulatory horizon. The checklist has been available since September 1, 2026 as an arXiv publication, addressed to the climate research and AI community.

The tool maintains a voluntary nature at present. Its adoption depends on internal policies of research organizations. Competitive advantage belongs to teams that build structured governance now, with named accountability and audit trail.

Regulatory enforcement of AI in Europe advances by stages. Organizations that align environmental metrics and documented responsibility gain a margin of 18-24 months before application becomes binding.