What is the environmental impact of AI image generation?

Understand what is being counted

Training builds a model. Using it produces a result. A full check can also include making the computers and running the building. A number for one part is not automatically the total impact.

The IEA explains that data centres need electricity for servers, storage, networks and cooling. Measuring a graphics card alone counts less than measuring the whole service. Check that difference before repeating a striking comparison.

Read more: IEA: Energy demand from AI

Why one number per image can mislead

A 2025 experiment compared 17 image-generation models and found large differences in energy use. The effect of resolution also varied by model. Those are findings from a defined experiment, not a live meter for every online editor.

Your workflow may include several candidates, edits and upscales before one image is accepted. Count those attempts when describing the job. A single successful preview does not tell you the resources used for the finished deliverable.

Read more: Bertazzini and colleagues: Energy consumption of AI image generation

Water and electricity are connected

Berkeley Lab’s data-centre report separates water used at a facility from water associated with generating its electricity. Cooling design, location and the electricity source affect the result. A system that uses less water on site may use more electricity.

This is why a fixed “glass of water per image” claim needs strong evidence for the exact service and setting. Do not turn a national data-centre estimate into a precise number for your Photoshop edit.

Read more: Berkeley Lab: 2024 US Data Center Energy Usage Report, water methodology

Ask better questions about green claims

A fair comparison checks the same task and counts the same things. A cheaper or faster image does not prove a lower environmental impact. A broad claim about a company buying renewable power is not enough either.

If the needed data is missing, say you cannot rank the services. Avoid giving a tool a green label based on unrelated figures.

• Which model, hardware, output size and number of results were measured?
• Does the figure include cooling, idle equipment and training?
• Where and when was the work done?
• Does it measure electricity, emissions, water or several of these?
• Is it a measured result or an estimate with stated assumptions?

Reduce work you do not need

You can control the number of requests you make even when you cannot measure their exact footprint. Use these steps to avoid work with no clear purpose. They are workflow choices, not a promised percentage saving.

1. Write down the required change and the final output size.
2. Reuse an approved image when it already meets the brief.
3. Review a small batch before requesting more variations.
4. Repair a simple local issue with an appropriate editing tool instead of regenerating the entire scene.
5. Upscale the selected result when needed, rather than every rejected candidate.

Describe your choices without inventing savings

For a client report, you can record the model, number of attempts, requested sizes and reused assets. For example: “We approved one background before making the final layout sizes.” That is a process fact you can verify.

Do not turn fewer requests into a carbon-saving number without a matching measurement. If the environment matters to your purchase, ask the service for evidence. Check its date and what was measured.