greenai.guide

Home › Good practice › Honest green claims about AI

Good practice

Honest green claims about AI

Updated · 5 min read

In short

Only make environmental claims about AI that are specific, measured on a stated method and true for the whole product, not just one part. Avoid vague words like "green" or "carbon neutral", show totals as well as efficiency gains, and check the rules where you sell: the UK CMA can now fine directly for misleading claims, and new EU rules have applied since 27 September 2026.

Key points

  • Specific, evidenced, comparable claims are safe; vague ones like 'green AI' or 'eco-friendly' are risky.
  • In the UK, the CMA's Green Claims Code sets six principles, and since April 2025 the CMA can fine up to 10% of global turnover for consumer law breaches.
  • In the EU, Directive 2024/825 has applied since 27 September 2026, banning generic environmental claims without proof and product claims of climate neutrality based on offsetting.
  • The proposed EU Green Claims Directive is not going ahead: the Commission's 2027 work programme lists it for withdrawal.
  • Efficiency is not sufficiency: a per-request saving is not a reduction if total use grows faster.

Why are green claims about AI risky?

Because AI's environmental impact is hard to measure, easy to describe selectively and increasingly regulated. A claim that sounds reasonable in a press release can mislead if it rests on an unstated boundary, an unrepresentative test or offsets.

In a poll at a 2026 UK green AI sandpit of mathematicians, computer scientists and industry, "greenwashing" was one of the most common one-word answers to the question of what green AI means. The sandpit's view was that without trustworthy numbers, a green claim is little more than a press release. Rules in the UK and EU now reflect that.

What can I claim, and what should I avoid?

Claim specific, measured changes with the method stated; avoid general labels and implied totals.

Usually fine, with evidence Risky or banned
"Energy per summary fell 40% between version 2 and version 3, measured with [method] on [workload]" "Green AI", "eco-friendly AI", "sustainable AI" with nothing behind it
"Runs in a region with lower grid carbon intensity than our previous host" "Carbon neutral" or "climate positive" based on offsets
"Our median text request uses X Wh, measured as follows…" "Zero-emission AI" because the host buys renewable certificates
"Uses a model one-tenth the size of our previous one, at the same accuracy on our test set" "Our AI saves the planet" or other unqualified benefit claims
"We report location-based and market-based emissions" Efficiency percentages with no baseline, workload or method

Be especially careful with:

  • Boundaries. Google's 2025 Gemini figures, for example, are stated for a defined boundary that excludes the external network, users' devices and training. Say what yours includes.
  • Comparisons. Compare like with like: same task, same quality, same measurement method.
  • Spin. The sandpit noted that an "X times more efficient" claim usually means either the old version was very wasteful or the claim is spin. Show the baseline.
  • Benefit claims. If you claim AI helps the environment elsewhere (optimising heating, logistics, grids), count the AI's own energy against the saving.

What do the rules say?

Where you sell matters. Check the rules in each market and take legal advice for significant claims. This section is a summary, not legal advice.

United Kingdom

  • The Competition and Markets Authority's Green Claims Code (2021) sets six principles: claims must be truthful and accurate; clear and unambiguous; must not omit or hide important relevant information; comparisons must be fair and meaningful; claims must consider the full life cycle of the product or service; and claims must be substantiated.
  • Since April 2025, under the Digital Markets, Competition and Consumers Act 2024, the CMA can decide itself whether consumer law has been broken, without going to court, and can fine up to 10% of global turnover (or £300,000 if higher).

European Union

  • Directive (EU) 2024/825 on empowering consumers for the green transition has applied since 27 September 2026. It adds to the blacklist of unfair commercial practices:
  • generic environmental claims ("green", "eco-friendly", "climate-friendly") unless the trader can show recognised excellent environmental performance relevant to the claim;
  • claims that a product has a neutral, reduced or positive climate impact based on offsetting emissions;
  • sustainability labels not based on a certification scheme or set up by public authorities. The Commission's guidance notes that companies can still describe investments in carbon credit projects, as long as they don't mislead.
  • The separate proposed Green Claims Directive, which would have required prior verification of claims, has stalled. The Commission signalled in June 2025 that it intended to withdraw it, and its 2027 work programme, adopted on 4 November 2026, lists it for withdrawal.

United States

  • The Federal Trade Commission's Green Guides (16 CFR Part 260, last revised in 2012) set out how environmental marketing claims should be qualified and substantiated. The FTC opened a review in December 2022, including on carbon offset claims; check whether revised guides have been issued before relying on the 2012 text.

Other markets have their own rules; the same principles of specific, substantiated, complete claims apply almost everywhere.

How does the rebound effect change what I can claim?

An efficiency gain is not a reduction if total use grows faster. When AI becomes cheaper and quicker, people and systems use more of it. This is the rebound effect, or Jevons paradox.

The scale is visible in the data. Google reports a 33-fold fall in energy per median Gemini text prompt over a year to May 2025, while the International Energy Agency projects data-centre electricity use roughly doubling to about 945 TWh by 2030. Both can be true at once.

So:

  • report totals alongside per-unit figures;
  • don't present a per-request saving as if it were an absolute reduction;
  • say whether freed capacity was retired or filled with more work;
  • see efficiency is not sufficiency.

How should I report?

Report enough that a sceptical reader could check your figures.

  1. Method: tool, version, and whether figures are measured or estimated (measure AI energy).
  2. Boundary: what is included and excluded.
  3. Functional unit: per request, per task, per user.
  4. Carbon basis: location-based, and market-based if you use it. The GHG Protocol asks companies with contractual instruments to report both. The SCI specification does not let offsets or certificates reduce a score.
  5. Totals and trends over time, not just the best single figure.
  6. Uncertainty: ranges, number of runs, assumptions.
  7. Offsets reported separately, never netted off silently.
  8. Independent review for important public claims.

Checklist

  • [ ] The claim is specific: what, how much, compared with what
  • [ ] Measured or estimated, with method and boundary stated
  • [ ] Comparison is like for like, with the baseline given
  • [ ] No generic terms ("green", "eco", "sustainable") without recognised evidence
  • [ ] No climate-neutral claims based on offsets
  • [ ] Totals shown alongside efficiency gains
  • [ ] Evidence kept on file: workload, tool versions, results
  • [ ] Checked against the rules in every market where the claim appears

What to do next

  1. Collect every public environmental claim your organisation makes about AI, including on websites and in sales material.
  2. Test each against the checklist; rewrite or remove the ones that fail.
  3. Agree an internal method so new claims rest on the same evidence base (measure AI energy).
  4. Add an approval step for environmental claims to your green AI policy.
  5. Ask suppliers for the evidence behind their claims; see sustainable AI procurement.
  6. For help, look in the consultants directory; for the wider argument, read efficiency is not sufficiency.

Frequently asked questions

Can I call my AI product green?

Not on its own. A bare 'green' or 'eco-friendly' is a generic claim that EU rules now ban unless you can show recognised excellent environmental performance, and that the UK CMA treats as likely to mislead without clear evidence. Make a specific, measured claim instead.

Can I say my AI service is carbon neutral because we buy offsets?

In the EU, claims that a product has a neutral, reduced or positive climate impact because of offsetting are now banned in consumer marketing. Elsewhere they are high-risk. Describe what you have reduced, and report offsetting separately.

Is it greenwashing to say our model is 30% more efficient?

Not if it is true, measured on a stated method and baseline, and you don't imply that total impact has fallen when it hasn't. Say what was compared, how and on what workload.

What is the rebound effect?

When efficiency makes something cheaper or easier, people tend to use more of it, so total consumption can stay the same or rise. It is also called the Jevons paradox.

Sources

  1. Green Claims Code: making environmental claims, Competition and Markets Authority (2021)
  2. How the CMA uses its direct consumer enforcement powers, Competition and Markets Authority (2025)
  3. Directive (EU) 2024/825 on empowering consumers for the green transition, EUR-Lex (2024)
  4. Empowering Consumers for the Green Transition: questions and answers, European Commission (2026)
  5. Green Claims Directive withdrawn, FoodTimes (2026)
  6. Guides for the Use of Environmental Marketing Claims (16 CFR Part 260), US Federal Trade Commission (eCFR) (2012)
  7. Guides for the Use of Environmental Marketing Claims: request for public comment, Federal Register (2022)
  8. Software Carbon Intensity (SCI) specification, Green Software Foundation (2024)
  9. Measuring the environmental impact of delivering AI at Google scale, Google (arXiv 2508.15734) (2025)
  10. GHG Protocol Scope 2 Guidance, GHG Protocol (2015)
  11. Energy and AI, International Energy Agency (2025)

Join us

Help cut AI's impact on the planet

Get new guides, funding calls and directory listings by email, about once a month. Or add your product, course, PhD or research to the directories.

No tracking. Unsubscribe in one click. Other ways to join · Privacy