AI Environmental Impact Model
An indicative disclosure scale for communicating the relative energy intensity of AI use. Applicable to any piece of work, any output type and any level of AI activity.
Three axes produce one grade. Task type and model category set the base grade in the grid below, and usage intensity moves it by one step in either direction.
- Axes
- 3
- Grades
- A to E
- Output
- indicative grade
- Intensity
- ± one grade
Published March 2026. Where a session involved multiple task types, score based on the highest-intensity task used.
EM v0.6 · Published · March 2026
This model is a disclosure tool, not a sustainability certification.
It communicates the relative energy intensity of different AI tasks rather than making verified carbon claims. Grade A does not mean zero impact; it means the indicative impact is negligible at the current state of measurement. Scores will be tightened at v1.0 as first-party provider data becomes available under EU AI Act requirements from August 2026.
The grid · v0.6
Select a base grade, then an intensity
Model category →
On-device or lightweight
Standard cloud
Frontier or large
Text
Extended reasoning or agentic
Image generation
Video generation
↑ Task type
Cell value = base grade before intensity
Usage intensity
Indicative grade · Text · On-device or lightweight · Light
A, Low energy intensity
Low energy use at session level. Disclose as part of standard transparency reporting.
Grade A held: light use cannot move it below A
Indicative grade · Text · On-device or lightweight · Moderate
A, Low energy intensity
Low energy use at session level. Disclose as part of standard transparency reporting.
Grade A held at moderate intensity
Indicative grade · Text · On-device or lightweight · Heavy or continuous
B, Low-moderate energy intensity
Low to moderate energy use. Disclose as part of standard transparency reporting.
Base grade A moved up one step to B
Indicative grade · Text · Standard cloud · Light
A, Low energy intensity
Low energy use at session level. Disclose as part of standard transparency reporting.
Grade A held: light use cannot move it below A
Indicative grade · Text · Standard cloud · Moderate
A, Low energy intensity
Low energy use at session level. Disclose as part of standard transparency reporting.
Grade A held at moderate intensity
Indicative grade · Text · Standard cloud · Heavy or continuous
B, Low-moderate energy intensity
Low to moderate energy use. Disclose as part of standard transparency reporting.
Base grade A moved up one step to B
Indicative grade · Text · Frontier or large · Light
A, Low energy intensity
Low energy use at session level. Disclose as part of standard transparency reporting.
Base grade B moved down one step to A
Indicative grade · Text · Frontier or large · Moderate
B, Low-moderate energy intensity
Low to moderate energy use. Disclose as part of standard transparency reporting.
Grade B held at moderate intensity
Indicative grade · Text · Frontier or large · Heavy or continuous
C, Moderate energy intensity
Moderate energy use. Disclose in deliverable transparency notes.
Base grade B moved up one step to C
Indicative grade · Extended reasoning or agentic · On-device or lightweight · Light
A, Low energy intensity
Low energy use at session level. Disclose as part of standard transparency reporting.
Base grade B moved down one step to A
Indicative grade · Extended reasoning or agentic · On-device or lightweight · Moderate
B, Low-moderate energy intensity
Low to moderate energy use. Disclose as part of standard transparency reporting.
Grade B held at moderate intensity
Indicative grade · Extended reasoning or agentic · On-device or lightweight · Heavy or continuous
C, Moderate energy intensity
Moderate energy use. Disclose in deliverable transparency notes.
Base grade B moved up one step to C
Indicative grade · Extended reasoning or agentic · Standard cloud · Light
B, Low-moderate energy intensity
Low to moderate energy use. Disclose as part of standard transparency reporting.
Base grade C moved down one step to B
Indicative grade · Extended reasoning or agentic · Standard cloud · Moderate
C, Moderate energy intensity
Moderate energy use. Disclose in deliverable transparency notes.
Grade C held at moderate intensity
Indicative grade · Extended reasoning or agentic · Standard cloud · Heavy or continuous
D, High energy intensity
High energy use. Disclose and consider lower-intensity alternatives for future work.
Base grade C moved up one step to D
Indicative grade · Extended reasoning or agentic · Frontier or large · Light
C, Moderate energy intensity
Moderate energy use. Disclose in deliverable transparency notes.
Base grade D moved down one step to C
Indicative grade · Extended reasoning or agentic · Frontier or large · Moderate
D, High energy intensity
High energy use. Disclose and consider lower-intensity alternatives for future work.
Grade D held at moderate intensity
Indicative grade · Extended reasoning or agentic · Frontier or large · Heavy or continuous
E, Very high energy intensity
Very high energy use. Disclose and actively consider lower-intensity alternatives before committing to this approach.
Base grade D moved up one step to E
Indicative grade · Image generation · On-device or lightweight · Light
A, Low energy intensity
Low energy use at session level. Disclose as part of standard transparency reporting.
Base grade B moved down one step to A
Indicative grade · Image generation · On-device or lightweight · Moderate
B, Low-moderate energy intensity
Low to moderate energy use. Disclose as part of standard transparency reporting.
Grade B held at moderate intensity
Indicative grade · Image generation · On-device or lightweight · Heavy or continuous
C, Moderate energy intensity
Moderate energy use. Disclose in deliverable transparency notes.
Base grade B moved up one step to C
Indicative grade · Image generation · Standard cloud · Light
B, Low-moderate energy intensity
Low to moderate energy use. Disclose as part of standard transparency reporting.
Base grade C moved down one step to B
Indicative grade · Image generation · Standard cloud · Moderate
C, Moderate energy intensity
Moderate energy use. Disclose in deliverable transparency notes.
Grade C held at moderate intensity
Indicative grade · Image generation · Standard cloud · Heavy or continuous
D, High energy intensity
High energy use. Disclose and consider lower-intensity alternatives for future work.
Base grade C moved up one step to D
Indicative grade · Image generation · Frontier or large · Light
C, Moderate energy intensity
Moderate energy use. Disclose in deliverable transparency notes.
Base grade D moved down one step to C
Indicative grade · Image generation · Frontier or large · Moderate
D, High energy intensity
High energy use. Disclose and consider lower-intensity alternatives for future work.
Grade D held at moderate intensity
Indicative grade · Image generation · Frontier or large · Heavy or continuous
E, Very high energy intensity
Very high energy use. Disclose and actively consider lower-intensity alternatives before committing to this approach.
Base grade D moved up one step to E
Indicative grade · Video generation · On-device or lightweight · Light
C, Moderate energy intensity
Moderate energy use. Disclose in deliverable transparency notes.
Base grade D moved down one step to C
Indicative grade · Video generation · On-device or lightweight · Moderate
D, High energy intensity
High energy use. Disclose and consider lower-intensity alternatives for future work.
Grade D held at moderate intensity
Indicative grade · Video generation · On-device or lightweight · Heavy or continuous
E, Very high energy intensity
Very high energy use. Disclose and actively consider lower-intensity alternatives before committing to this approach.
Base grade D moved up one step to E
Indicative grade · Video generation · Standard cloud · Light
C, Moderate energy intensity
Moderate energy use. Disclose in deliverable transparency notes.
Base grade D moved down one step to C
Indicative grade · Video generation · Standard cloud · Moderate
D, High energy intensity
High energy use. Disclose and consider lower-intensity alternatives for future work.
Grade D held at moderate intensity
Indicative grade · Video generation · Standard cloud · Heavy or continuous
E, Very high energy intensity
Very high energy use. Disclose and actively consider lower-intensity alternatives before committing to this approach.
Base grade D moved up one step to E
Indicative grade · Video generation · Frontier or large · Light
D, High energy intensity
High energy use. Disclose and consider lower-intensity alternatives for future work.
Base grade E moved down one step to D
Indicative grade · Video generation · Frontier or large · Moderate
E, Very high energy intensity
Very high energy use. Disclose and actively consider lower-intensity alternatives before committing to this approach.
Grade E held at moderate intensity
Indicative grade · Video generation · Frontier or large · Heavy or continuous
E, Very high energy intensity
Very high energy use. Disclose and actively consider lower-intensity alternatives before committing to this approach.
Grade E held: heavy or continuous use cannot move it above E
01.
How scoring works
As AI use becomes routine in professional and creative work, its environmental footprint is a growing area of public and regulatory interest. Verified per-model energy figures are not yet available for most AI systems at the level of detail needed for precise accounting, and the methodologies providers do publish use incompatible approaches that make comparison difficult.
This model takes a practical approach to that challenge. Rather than claiming a precision the data cannot support, it uses task type, model category and usage intensity to produce an indicative A to E grade that any organisation can apply and disclose. The grade sits alongside the AI Transparency grade on the Report Card and can be included in deliverable footers as a standard transparency measure.
Every assessment combines three factors. Where a session involved multiple task types, score based on the highest-intensity task used.
Axis 1
Task type
Text
Writing, researching, editing, analysing, translating or generating code
Extended reasoning or agentic
Using an AI tool's advanced research or reasoning mode, or running multi-step automated workflows
Image generation
Creating images or visual assets using AI
Video generation
Creating video content using AI
Axis 2
Model category
On-device or lightweight
A lightweight or on-device tool
Standard cloud
ChatGPT, Claude, Gemini or similar used for everyday tasks
Frontier or large
An advanced or specialist tool, or a general assistant in its most powerful mode
Axis 3
Usage intensity
Light
Under ~30 minutes, single task
Move one grade down (minimum A)
Moderate
~1 to 2 hours, several tasks
No change. Base grade applies.
Heavy or continuous
Throughout a working day or project
Move one grade up (maximum E)
01.
The A to E grade scale
This is the one scale in the framework where the colour carries a direction. One end of it genuinely is worse than the other, and the grade is a prompt to consider a lower-intensity approach rather than a band a reader should be indifferent between.
- Green, negligible
- Amber, moderate
- Red, very high
Low intensity
Low energy intensity
Low energy use at session level. Disclose as part of standard transparency reporting.
Typical use: Brief text query, on-device tools, single query to any text AI.
Low-moderate
Low-moderate energy intensity
Low to moderate energy use. Disclose as part of standard transparency reporting.
Typical use: Full working day on text tasks with Claude or ChatGPT; occasional AI image generation.
Moderate
Moderate energy intensity
Moderate energy use. Disclose in deliverable transparency notes.
Typical use: Continuous frontier model use across a multi-day project; moderate-volume image generation campaign.
High
High energy intensity
High energy use. Disclose and consider lower-intensity alternatives for future work.
Typical use: Frontier image generation campaign; extended deep research mode use; any video generation.
Very high
Very high energy intensity
Very high energy use. Disclose and actively consider lower-intensity alternatives before committing to this approach.
Typical use: Heavy AI video production; frontier reasoning at high volume; large-scale automated content generation.
02.
JSON schema
The canonical machine-readable definition of this model, including the base grade grid, axis definitions and methodology sources, is published as a JSON schema file under CC BY 4.0.
dttt-environmental-v0.6.json
AI Environmental Impact Model, current published version
01.
Methodology sources
The grid is a reading of the published evidence rather than a measurement. These are the sources it reads, and the reason the model states a band rather than a figure.
2024
IEA
Electricity 2024: Analysis and Forecast to 2026
Data centre electricity consumption projections including AI workloads.
iea.org
2024
Goldman Sachs
AI is Poised to Drive 160% Increase in Data Center Power Demand
AI power demand projections and infrastructure investment context.
2025
Gemini environmental impact disclosure
First major first-party per-query environmental disclosure by an AI company. Reports median text query energy use of 0.24 Wh. Provides the most credible single anchor point for standard cloud text task energy estimates currently available.
2025
Epoch AI
How Much Energy Does ChatGPT Use?
Third-party estimate of GPT-4o query energy at approximately 0.3 Wh, broadly consistent with Google's first-party Gemini figure. Provides context for frontier model text task energy intensity and explains the role of mixture-of-experts architecture in efficiency.
epoch.ai
2025
Hugging Face / Luccioni and Alves da Costa
What kind of environmental impacts are AI companies disclosing? (And can we compare them?)
Critical analysis of 2025 environmental disclosures from Google, OpenAI and Mistral. Establishes that current company disclosures use incompatible methodologies and cannot be meaningfully compared. Supports the decision to base DTTT scoring on task-type and model-category bands rather than per-model figures.
huggingface.co
2025
Bailey / Medium
The Energy Impact of LLMs — Where Are We in 2025?
Synthesis of task-type energy differentials. Estimates image generation at up to 60 times more energy than simple text queries, and video generation at approximately 2,000 times more. Supports the task-type hierarchy in the base grade grid.
nathanbaileyw.medium.com
2025
MIT Technology Review
We did the math on AI's energy footprint
Documents the lack of first-party disclosure from closed model providers and the structural reasons why per-model verified figures are not currently available. Supports the disclosure-based rather than measurement-based approach of this model.
technologyreview.com
2025
EU AI Act / European Commission
EU General-Purpose AI Code of Practice
Signed by OpenAI and Anthropic in July 2025. Requires comprehensive energy consumption metrics as part of model documentation. Enforcement begins August 2026 for new models. Expected to materially improve first-party disclosure data available to update this model's methodology sources at v1.0.
Grade boundaries are revised at v1.0 rather than adjusted between releases, so a disclosure issued under v0.6 remains readable against the version that produced it.