Skip to content
Get in touchFR
Updated in October 2026
Explained

How can you reduce the environmental impact of using AI?
10 actions ranked by effectiveness

A ranking of 10 actions to reduce the environmental impact of using AI, by what each avoids over a year: the actions that matter, those that help and those that change almost nothing, then what no individual action can fix.

Guillaume Pakula
By Guillaume Pakula, co-founder of Celsius. Since 2020, he has helped 80+ organisations with their Bilan Carbone® and climate strategy.
October 2026
Updated October 2026 · 11 min
Cutting back on video, choosing a smaller model or no longer saying thank you: for a single usage profile, we calculated what each of 10 actions avoids over a year. 3 of them can avoid more than 1 kg of CO2 equivalent (kgCO2e) per year, while 2 avoid less than 50 g. The profile as a whole emits about 3 kgCO2e per year from server electricity alone.
Key takeaways
  • 1Generating 2 times fewer videos avoids 300 g to 1.1 kgCO2e per year, depending on the generator.
  • 2Reasoning mode uses 1.9 to 30 times the energy of a simple question, depending on the study and what it compares.
  • 3An answer 2 times shorter uses 54% less energy, on the model measured.
  • 4Dropping 1 thank-you per day avoids 41 gCO2e per year, as much as a single video emits.

Not all actions to reduce the environmental impact of using artificial intelligence (AI) are equally effective. For someone who generates 1 video per week, generating 2 times fewer avoids about 1.1 kgCO2e per year (1.1 kg of CO2 equivalent) with our reference generator, equivalent to 7.5 km by car, or 300 g with a median generator. No longer thanking your assistant (1 thank-you per day) avoids 41 g, which is 26 times less.

Reduce the impact of AI

10 actions ranked by what they avoid in a year

10 questions per day, 5 images and 1 video per week. Server electricity only, global average.

The actions that matter

The actions that help

The actions that change almost nothing

Projet Celsius calculation based on Epoch AI (2025); Luccioni et al. (2024); Delavande et al. (2025 and 2026); Arcep and PEReN (2026); AI Energy Score (2025); UNESCO and UCL (2025); Barros et al. (2025); Saad-Falcon et al. (2026); Husom et al. (2026); Cloud Carbon Footprint; ADEME (the French Agency for Ecological Transition), Impact CO2; Ember, Global Electricity Review 2026 (AI Impact Database).

We quantified 10 actions for a usage profile that we chose: 10 questions per day, 5 images and 1 video of 5 seconds per week. Of these, 3 actions matter, 5 help and 2 change almost nothing. By our calculation, this profile emits about 3 kgCO2e per year from server electricity alone, equivalent to 21 km by car.

1The 10 actions to reduce the environmental impact of AI, from most to least effective

How the actions are grouped

An action "matters" if it can avoid at least 1 kgCO2e per year and "helps" if it can avoid at least 50 g. Among the actions that help, running a model (the software that produces the answers) on your own computer only lightens the footprint if it is smaller than the model behind the online service.

An editor seen in three-quarter rear view, in front of 3 screens, in a video editing suite with black curtains
A video editing suite in Hamburg in 2016. According to a 2025 study, the most downloaded video generator on Hugging Face uses about 90 Wh for 5 seconds of video, as much as 300 written questions by our calculation. Photo Klaus Eichler / Max Mittelbach, Wikimedia Commons, CC BY-SA 3.0 DE

How it is calculated

For each action we assume a behaviour of our own choosing, which serves as its reference value: generating 2 times fewer videos, shortening 1 answer in 2, dropping 1 thank-you per day. 7 actions are based on a published measurement; for the other 3 (whether AI is needed, follow-ups and stored conversations), none exists. Every figure in grams or kilograms in this article comes from our own calculation, using global average electricity.

This profile is more active than average: in July 2025, the average weekly active ChatGPT user sent 3.6 messages per day, according to our calculation from the figures in a study co-authored by OpenAI. No figure has been published for the number of videos generated per user.

The ranking therefore depends on the profile. For an intensive user, the total rises to 15 kgCO2e per year, of which 11 kg comes from video, which stays at the top; 8 actions out of 10 can then avoid more than 1 kg. For an occasional user (2 questions per day, 1 image per week, no video), reasoning mode moves to the top: the total falls to 170 gCO2e per year, and no reference value exceeds 35 g. Within a group, the order carries no meaning where the ranges overlap, and the savings cannot be added together.

2The actions that matter most: video, reasoning mode and whether AI is needed

Generate fewer videos, and shorter ones: about 1.1 kgCO2e per year

In our profile, the 52 videos of the year emit 2.1 kgCO2e out of 3, more than 70% of the total. This figure rests on a single generator: the most downloaded one on Hugging Face (a platform for open models, meaning models that can be downloaded freely), and the second most energy-hungry of the 7 (in French) measured in the study it comes from. With the median generator, the one in the middle of the ranking (it produces lower-definition videos), generating 2 times fewer would avoid only about 300 g.

Duration matters even more than the number of videos: according to the same study, a video 2 times shorter needs about 4 times less computation, on the generator tested. ADEME, the French Agency for Ecological Transition, lists among its main principles of use for users that video generation must be limited and meet a specific need. Our article on the energy used by an AI-generated video goes into detail.

Do not leave reasoning mode on by default: from 0.46 to 15 kgCO2e per year

Reasoning mode makes the model work through intermediate steps before answering: it writes much more text, and therefore uses more energy. PEReN, the French state's digital regulation expertise centre, measured open models for Arcep, the French electronic communications regulator: with reasoning mode on (in French), they use 92% more energy on average, with an uncertainty range from -21% to +348%, and 849% more on average on programming exercises.

The Hugging Face AI Energy Score ranking compares 2 groups of models: those that reason consume 30 times more on average than the others. For 3 models tested with and then without reasoning, it even measures 150 to 700 times more.

Reducing the impact of AI

Reasoning uses 1.9 to 30 times the energy of a simple question, depending on the study and on what it compares

Energy of a question processed with reasoning, relative to one processed without. The area of each square is proportional to the energy.

1

Without reasoning

the baseline question

x1.9

Reasoning on, according to PEReN

up to 92% more on average, across the open models tested

x9.5

Same measure, on a coding task

the highest case PEReN recorded

x30

Reasoning models, according to AI Energy Score

on average, compared with models that do not reason

Arcep (the French electronic communications regulator) and PEReN (the French digital regulation expertise centre), Generative artificial intelligence: what environmental challenges? (May 2026, in French); Hugging Face, AI Energy Score v2 (December 2025). PEReN multiples: Projet Celsius calculation from +92% and +849%.

This action is a special case: our profile includes no question in reasoning mode, so the action is an avoided addition. Switched on for the 3,650 questions of the year, this mode would add between 460 gCO2e (with the PEReN figure) and 15 kgCO2e (if we apply the factor of 30, although that compares 2 different groups of models). This is the most uncertain item in the ranking. The action itself is to reserve this mode for the questions that call for it.

Ask yourself whether AI is necessary: from 0 to 3 kgCO2e per year

This is the first condition of frugal AI in the framework published by AFNOR, France's national standards body, as summarised by Ecolab, a French state service: having demonstrated that AI is needed rather than a solution that consumes less. The gain is that of the use avoided, to which we give no reference value, only a range from nothing to the whole profile. Our comparison of ChatGPT and a Google search quantifies the difference that a conventional search makes; with the reference values of this series, giving up a video avoids as much as 300 questions would emit.

3The actions that help: images, model size, short answers, complete requests and local models

Generate fewer images, and smaller ones: about 170 gCO2e per year

The 260 images of the year emit 350 gCO2e if we take the average of the generators measured. Generating 2 times fewer avoids 170 g, and 2 times less again with a median generator. Size matters too: according to a study from the University of Florence, an image 2 times smaller on each side needs 1.3 to 4.7 times less energy, depending on the generator. See our article on the footprint of a generated image.

Choose the smallest model that does the job: about 140 gCO2e per year

According to a study from Université Côte d'Azur, choosing the smallest sufficient model everywhere would reduce the energy consumed by AI in the use phase by 27.8% overall. We apply this global average to our questions, for want of a better one. The gap between 2 models can be much larger: small models specialised in one task (such as summarising, translating or answering questions) consume 15 to 50 times less than a general-purpose model according to UNESCO, although they do not suit every use. Our ranking of ChatGPT, Claude and Gemini compares the models of the main providers. At Anthropic, Fable 5.2 or Opus 5.5 shows what moving up to the most powerful model costs, and how much energy and water Fable 5.2 and Opus 5.5 use shows the effect of the effort setting.

Ask for short answers: about 140 gCO2e per year

UNESCO and University College London measured the effect of length on an open model: an answer 2 times shorter (in French) uses 54% less energy, whereas a question 2 times shorter saves only 5%.

Reducing the impact of AI

Shortening the answer saves 54% of the energy, shortening the question 5%

Energy of one answer by length of question and length of answer, measured on an open model.

Baseline

400-word question, 400-word answer

Question 2 times shorter

200-word question, 400-word answer

Answer 2 times shorter

400-word question, 200-word answer

Answer 4 times shorter

400-word question, 100-word answer

On the same model, measured on another machine by Hugging Face researchers, writing 1 unit of text takes about 35 times more energy than reading 1, according to our calculation.

UNESCO and UCL, Smarter, Smaller, Stronger (2025), Annex A.3: a single model (LLaMA 3.1 8B), desktop graphics card. Delavande, Pierrard and Luccioni (2026) for the ratio of 1 to 35, Projet Celsius calculation.

On the same model run on another machine, writing a token (a word or part of a word) takes about 35 times more energy than reading one, according to our calculation from the measurements of Hugging Face researchers. We assume that 1 answer in 2 is shortened by half, which avoids 140 gCO2e per year. ADEME also considers it beneficial to specify the length or quality of the content you want.

Write one complete request rather than follow-ups: about 100 gCO2e per year

Each follow-up makes the model write a new answer. According to ADEME, precise queries will require fewer repetitions and new queries. We count 1 follow-up avoided for every 5 questions, or 100 gCO2e per year. This is an assumption, as no measurement has been published.

Attaching a long document costs more: according to an estimate by the research institute Epoch AI, having the model read 100,000 tokens, or about 80,000 words, uses about 40 Wh (watt-hours), as much as 130 questions by our calculation. According to the same source, this cost is paid only once per conversation, so attach only the pages you need.

Run a small model on your own computer

With the same model, no gain has been established: handling one request at a time, a computer chip and a chip in a data centre, the building where the servers run, use about the same energy, and a data centre card that handles 64 requests at once uses 12 to 20 times less per request. The gain comes from a smaller model: according to our calculation based on the measurements in another study, a lightweight model on a laptop consumes about 4 times less than a ChatGPT question, our reference, for an answer of equal length.

A laptop open, seen close up, its screen showing a terminal full of lines of text
A laptop and its terminal, in 2026. According to a study updated in 2026, a lightweight model on a laptop consumes 0.03 Wh for an answer of 220 tokens, not counting the screen or the rest of the machine. Photo Solijon Solayev, Wikimedia Commons, CC BY-SA 4.0

This gives a gain from 0 to 380 gCO2e per year, for the same electricity, according to our calculation. In France, electricity emits 11 times less than the world average, 41 gCO2e per kWh against 458 g in 2025, according to data from Ember, an energy think tank: the balance becomes favourable if the online service runs on electricity close to that average. Our article on AI presented as eco-friendly examines this case further.

4The actions that change almost nothing: thank-yous and deleted conversations

Stop saying thank you: 41 gCO2e per year

In April 2025, a user wondered how much electricity people's thank-yous cost OpenAI. Its chief executive, Sam Altman, replied, as the press reported and without any calculation to back it up: "tens of millions of dollars well spent - you never know".

In January 2026, researchers at Hugging Face published a measurement on an open model: a thank-you consumes 0.245 Wh, including the model's answer. 1 thank-you per day adds up to 41 gCO2e per year, or 290 metres by car. A year of daily thank-yous uses about as much energy as a single video of 5 seconds, which equals about 370 thank-yous.

Reducing the impact of AI

It takes about 370 thank-yous to use as much energy as one generated 5-second video

On an open model, one thank-you uses 0.245 Wh, including the model's reply; a 5-second video, about 90.0018 Wh.

1 thank-you

0.245 Wh

1 video of 5 seconds

90.0018 Wh, or about 370 thank-yous

Delavande, Pierrard and Luccioni, Small Talk, Big Impact (2026); Delavande, Pierrard and Luccioni, Video Killed the Energy Budget (2025). Ratio: Projet Celsius calculation.

Delete your conversations: less than 10 gCO2e per year

A year of questions and answers takes up around ten megabytes of text, and the generated images and videos a few hundred. Based on the consumption of a data centre disk and 3 copies of each file, keeping this data for a year takes about 17 Wh for the disk alone, which is less than 10 gCO2e according to our calculation and our assumptions. There can be other reasons to delete your conversations, such as protecting your privacy.

5In a company: choosing tools and setting rules of use

For an organisation, the AFNOR framework on frugal AI, published in June 2024, lists 31 best practices, grouped into 7 themes, which the Ecolab summary ranks by ratio of gain to effort, from best to weakest. One theme concerns the need: checking that AI is a relevant way to meet it.

Our guide explains how to count AI in your company's Bilan Carbone® (the French carbon accounting method), another quantifies AI's contribution to an SME's footprint, and our calculator gives a first order of magnitude; to measure and manage usage over time, our article on the providers that measure AI in companies compares the options. AI comes on top of computers, screens and phones, which our guide on a company's digital carbon footprint covers.

A wide shot of a training room: participants seated around tables with their laptops, a female trainer standing in front of a screen
A workshop of Les sans pagEs, a Wikipedia contribution project, at La contemporaine in June 2023. One of the 7 themes of best practices in AFNOR's framework on frugal AI is skills. Photo MarianneLC, Wikimedia Commons, CC BY-SA 4.0, cropped

6What individual actions do not fix: missing provider data, default AI and data centre growth

Measurements that providers do not publish

This ranking combines an estimate (the energy of a question), researchers' measurements made on open models, and our assumptions. For the most widely used assistants, there is no measurement published by their providers of the energy used by an image or a video, and PEReN did not test the reasoning mode of closed models such as ChatGPT's. Users therefore cannot check the effect of their own actions. The calculation also leaves out the stages where AI has an environmental impact before the answer: chip manufacturing and model training.

AI switched on by default

Some AI is not something the user asked for: answers generated at the top of search engine results, and assistants built into software and phones. In its July 2026 opinion, ADEME asks that, in traditional services, users be able to switch off these features if they wish. It also considers training and awareness crucial to using AI thoughtfully. To raise a team's awareness, our article on choosing a speaker compares talk formats.

The growth of data centres

One person's actions amount to a few kgCO2e per year. Data centres' electricity consumption rose by 17% in 2025, an increase of about 70 TWh (terawatt-hours), according to the International Energy Agency (IEA): our article on data centre electricity consumption gives the orders of magnitude, and the one on the 6 environmental impacts of AI looks beyond carbon alone.

7Key takeaways

  • 3 actions matter: generating fewer videos, and shorter ones; not leaving reasoning mode on by default; asking whether AI is necessary.
  • 5 actions help: fewer images, a smaller model, short answers and complete requests avoid from 100 to 170 gCO2e per year for those with a reference value, while a local model helps if it is smaller.
  • 2 actions change almost nothing: stopping the thank-yous (41 gCO2e per year) and deleting your conversations (less than 10 g).
  • This ranking applies to people who generate videos: the profile emits about 3 kgCO2e per year from server electricity alone, more than 70% of it from video.

This ranking does not say whether you should use AI: our article on the arguments for and against boycotting AI addresses that question, and the one on digital pollution beyond AI places AI among other uses.

Further resources

Frequently asked questions

or: guillaume@projetcelsius.com

More articles to read

View all →
French environmental cost label for clothing: what changed on 1 October 2026
Regulation

French environmental cost label for clothing: what changed on 1 October 2026

6 min read
What is the French textile eco-score?
Who displays it, and has it become mandatory?
Regulation

What is the French textile eco-score? Who displays it, and has it become mandatory?

12 min read
Removable batteries and battery passports: who is affected in 2027?
Regulation

Removable batteries and battery passports: who is affected in 2027?

14 min read

A senior consultant reads every message

Tell us where you stand and what is pressing, and we will reply within 24 working hours with a first view on how to proceed.

Write to a consultant →
Our toolkit

A full range of free assessment tools

Cost estimator, obligation and eligibility checkers, footprint calculator.

Estimate costs and impacts
Ref. 2026·BC·0184REPORT · FINANCIAL YEAR 2026Bilan Carbone®complete 1·2·3Presentation to the Executive Committee · 2027 action planCONTENTS01Executive summaryp. 402Scope and methodologyp. 1203Emissions by scopep. 24042027 action planp. 5605Methodological appendicesp. 78VOLUME84 pages · ConfidentialSTANDARDISO 14064 · GHG ProtocolQUOTATION · BC.2026.0184Issued 05·05·2026 · Valid for 30 daysCelsius consultancy · Paris 3rd arrondissementINDICATIVE COST ESTIMATEYour companyMid-sized company · 180 employees · CSRD 2027SERVICEBilan Carbone®Scope 1·2·3 · 13 weeksBPIFRANCE FUNDING · - 60%INDICATIVE RANGE EXCL. VAT€16,800to €22,400PROJECT DETAILS01 · SCOPING3 weeks~€4,20002 · MODELLING8 weeks~€9,80003 · PRESENTATION2 weeks~€4,800Indicative estimate · Refined after scopingSIRET 891 234 567 00012

Bilan Carbone® cost estimator

Your price range in 1 minute, based on 2026 market rates and public funding.

1 minStart
Estimate costs and impacts
AI FOOTPRINT · MONTHLY USEYOUR TOTAL FOOTPRINT4.2 kgCO₂eq for 1,000 typical prompts · May 2026MODEL COMPARISON · 1,000 TYPICAL PROMPTSGPT-4oOpenAI6.8 kgClaude OpusAnthropic2.1 kgGemini ProGoogle4.5 kgMistral LargeMistral AI · FR1.6 kg

AI carbon footprint calculator

The climate impact of your AI queries, by model and task. Ecologits methodology.

3 minStart
Check an obligation
DIAGDÉCARBON'ACTIONACCELERATE THETRANSFORMATIONOF YOUR COMPANYIN ENERGY ANDTHE ENVIRONMENTOFFICIAL SCHEME · BPIFRANCE × ADEMECHECKER · ELIGIBILITYELIGIBLEYou meet the 3 Diag Décarbon'Action criteria.YOUR FUNDING€10,000 excl. VAT€6,000excl. VATAmount payable by you after the Bpifrance subsidyFUNDING BREAKDOWNBPIFRANCE 40%YOU 60%€4,000€6,000SCHEME SUBSIDISED BY

Diag Décarbon'Action eligibility

Check in 30 seconds whether your business is eligible for Bpifrance funding covering 40% of a Bilan Carbone® assessment.

30 secStart
Check an obligation
FRENCH REPUBLICMinistry for the Ecological TransitionLEGAL REQUIREMENTGreenhouse GasEmissions AssessmentArticle L. 229-25 of the French Environmental CodeDecree 2022-982 · published 1 July 2022FREQUENCY · 4 YEARSFILING · ADEMECHECKER · BEGES REQUIREMENTREQUIRED720 employees · BEGES requirementPOTENTIAL PENALTY€50,000maximum fine for failing to complete itart. L. 229-25OFFICIAL REFERENCE

BEGES checker

Does the BEGES requirement apply to you? An immediate answer, with the deadline and penalty.

30 secStart
Check an obligation
9:42SCANSCANNING DPP...DPP IDENTIFIEDOrganic cotton T-shirtSKU TX-CB-220 · Batch L-26-04781ESPR COMPLIANT42 DATA POINTS · 5 SECTIONSORIGINIndia · GOTSOrganic cottonFOOTPRINT5.2 kg CO₂eqLCA ISO 14040RECYCLABILITY85%Cotton recycling schemeREPAIRABILITY7.5 / 10Accessible partsEND OF LIFETextile recycling scheme (TLC) · Bordeaux/FRRefashion-approved recovery operatorEU · ESPR 2024/178112.05.2026

DPP checker

Is your product covered by the Digital Product Passport?

1 minStart
Check an obligation
COSTENVIRONMENTAL386POINTS257PER 100 GOfficial methodEcobalyse, v7.0.0CHECKER · ENVIRONMENTAL LABELLINGELIGIBLETextiles · voluntary labellingENVIRONMENTAL COST386 ptsT-shirt 150 g · 257 pts per 100 gExample: Ministry for the Ecological Transition FAQ

Textile environmental labelling 2026

Voluntary labelling, but third parties can publish it from October 2026: where do you stand?

1 minStart