Updates
Every new item, every revision of a value and every correction, dated. A corrected value is also corrected in the history of its entry.
Type
NewWhich AI use charter should you adopt?A guide to writing an organisation's AI use charter: 16 published charters compared article by article, and a model in 10 articles to adopt, including the accounting of use and its footprint.NewCan AI be decarbonised?An analysis of the levers that remove emissions from AI, ranked from a common reference: model efficiency, grid electricity, model size, hardware lifetime, and what electricity contracts change.NewWhat happens to the data you give an AI?An analysis of what happens to a conversation, a document or an image entrusted to an AI assistant: training, retention, human review, business offers and GDPR rights.NewHow do you account for AI in a Bilan Carbone®?A method for counting AI in a Bilan Carbone®: the items and scopes depending on who runs the servers, activity data, factors and their uncertainty.NewWhat happens to entry-level workers when AI does their work?An analysis of what studies on the hiring of beginners in AI-exposed occupations measure, in France and other countries, since late 2022.NewWhat is the footprint of my AI use over a year?An analysis of the weight of a year of AI use, from the occasional profile to the developer with agents, calculated with the AI Impact Database and set against the annual footprint of a French person.NewIs AI making us stupid?An analysis of the most cited studies on AI and our abilities, ranked by level of evidence, from the survey to the randomised trial.NewHow could AI amplify the climate crisis?An analysis of the routes by which AI can push emissions up, based on a 2026 study that gives it 2 futures: 11% more CO2, cumulated from 2025 to 2040, or 1.4% less. It reviews AI electricity by origin, autonomous transport and AI put to the service of oil and gas.NewDoes AI really make us more productive at work?An analysis of the gains measured at 3 levels: up to 40% less time on a task, 1 to 2 hours a week according to users, almost nothing in the accounts of companies and countries. It carries 29 sources, 3 infographics and a calculation to play with, which shows what remains of the gain on one task at the scale of a team.MethodologyThe analysis on diverging estimates, rewritten and extendedThe analysis "Why do estimates of AI's impact differ?" now opens on the 4 ways of counting the CO2e of one question, and ends each of its 6 parts with the question to ask of a figure. No value changes: the additions come from the "Where does this figure come from?" notes on Meta, on the water bottle and on Japan.NewA tool to compare the footprint of a text by AI modelThe page "Footprint of a text, by model" gives the weight in gCO2e of a text written by 16 models (GPT, Claude, Gemini, Mistral), from a tweet to a book, for 4 server locations. It carries a table of the 16 models and the 6 lengths to download, their distribution on a logarithmic scale and 5 images to share. The values are estimates from EcoLogits (model database of 14 October 2026), hardware manufacturing included: the page says what separates them from those of the AI Impact Database.CorrectionThe 84% of OpenRouter attributed to their study, the Mistral AI response reduced to half a pageThe analysis on diverging estimates attributed to Arcep the share of OpenRouter tokens processed by models with no environmental data: the figure comes from Luccioni et al. (2025), on the 20 most used models in May 2025. The 45 mL of Mistral AI apply to a response of 400 tokens, about half a page, and not a page. The paragraph on the utilisation rate of servers no longer compares the load assumed by GreenIT with the power drawn observed by RTE, which do not measure the same thing, and the final table no longer gives a gap without a calculation.CorrectionOne question to an AI, 0.012 gram of CO2e presented as a calculationThe analysis on diverging estimates presented 0.012 gram of CO2e as a published figure: it is a calculation by the hub, 0.3 Wh with French electricity (41.4 gCO2e/kWh). The text now says so. The chart thumbnail, which showed "x 70", gives the calculated gap between 0.012 and 1.14 gram: 92 times.CorrectionGoogle counts the manufacturing of chips and memoryThe analysis wrote that Google does not count hardware manufacturing. Its August 2025 study counts that of chips, processors and memory: about 30% of its 0.03 gCO2e per question (Elsworth et al., table 2).CorrectionJobs created and eliminated by AI by 2030, 11 and 9 millionThe chart thumbnail highlighted 170 million jobs created by 2030, a World Economic Forum figure for all factors combined. It now gives the share of AI and information processing: 11 million jobs created and 9 million eliminated (Future of Jobs Report 2025).CorrectionLabels of 3 figures on the home page and the Climate topicOn the home page, the graphic designer's 3 tasks (520, 370 and 110 g) are annual shares of the job and now say so ("g/yr"); the frequent question on image and video now compares correctly (one video equals 31 images, or 300 questions). On the Climate topic, the increases of 81% (Google) and 58% (Microsoft) are now described as total emissions of the groups.CorrectionWater consumed by the world's data centres, 560 billion litres in 2023The IEA figure was presented as today's: it covers 2023 and also counts the water used in manufacturing chips, which the entry said was excluded. The 1,200 billion litres for 2030 are stated for the IEA base case (Energy and AI, April 2025, box 5.4).CorrectionColours of the markers on the carbon intensity mapThe marker bars under the map now take the colour of the class of their value: France (41 gCO2e/kWh) was coral, the colour of the "600 and above" class.Correction5-second video, 31 images instead of 30With 90 Wh for a 5-second video and 2.9 Wh for an image, a video equals 31 images. The analysis on video and the "Estimate the footprint of your uses" tool stated 30.DataCarbon footprint of a French person, 8.3 tCO2e in 2025The benchmark moves to 8.3 tCO2e per person, the 2025 value of the SDES and Insee series published in October 2026; it used the 2024 value of the previous edition. The usage tool and the analyses read the same value.CorrectionLife cycle assessment of Mistral Large 2, the stages outside the calculation detailedThe "Hardware manufacturing and other" share becomes "Other stages", now detailed: server hardware (11% of climate, 61% of materials), users' devices (3% and 7%) and data centre buildings (less than 1% and 1.5%).CorrectionWater per question, the gap named by its 2 boundsThe chart gives the gap between published measurements with its 2 bounds: 0.26 mL (Google, cooling only) and 50 mL (UC Riverside), a factor of about 190.CorrectionMistral AI assistant, Vibe instead of Le ChatThe viewpoint "Surface AI and deep AI" names the Mistral AI assistant by its current name, Vibe (formerly Le Chat). Charts that cite the life cycle assessment of July 2025 keep the name of the time.MethodologyA single figure for the world's electricity: 458 gCO2e per kWh, across the hubThe hub, the calculator and the blog now read the AI Impact Database: 458 gCO2e/kWh for the world (Ember, life cycle, instead of 435 gCO2/kWh according to the IEA, power plant CO2 only) and 41.4 g for France (instead of 51.9 g according to ADEME), a single source for all countries. Conversions move by 5%; a year of regular use goes from 2.8 to 3 kgCO2e.CorrectionGenerated video, 90 Wh instead of 944 WhThe first version of the chart used the 944 Wh from MIT Technology Review. The hub now uses 90 Wh, the most downloaded open generator measured by Delavande et al., the reference value of the article series; the 944 Wh are explained in a note.DataGoogle and Microsoft emissions: 2026 editions of the reportsThe 2 series use the restated values of the reports published in June and July 2026: +81% since 2019 for Google, +58% since fiscal year 2020 for Microsoft.DataIreland: 23% of electricity in 2025The share of data centres in Ireland moves to the 2025 value from the Central Statistics Office, electricity measured at the meter.New3 world maps: electricity and waterCarbon intensity of electricity, share of low-carbon electricity and water stress, country by country, based on Ember and the FAO.New3 tools: uses, server countries, orders of magnitudeThe tools calculate on the same benchmarks as the entries and the blog articles.NewThe "Where does this figure come from?" series5 analyses: the image and the phone charge, the video at 1 kWh, the email and the water bottle, AI and Japan, Meta's 1,358 tonnes.NewWhy do estimates of AI's impact diverge?The analysis of the 6 methodological choices that explain the gaps between published figures.NewViewpoint: surface AI and deep AIGuillaume Pakula's essay on the part of AI that we do not see.New29 charts and 13 reference figuresElectricity, water, climate, hardware, uses and society: each chart has its entry, its downloadable data, its reading and its limitations.MethodologyLaunch of the hub and its methodologyThe methodology for selecting sources, choosing reference values and updating is public.
An error? Report it : each report is checked.
Across the hub
AI Impact DatabaseAll the figures, their methodology and their sources.
AI Ranking10 assistants rated out of 100: performance, ethics and environment.
AnalysesThe signed pieces: analyses, viewpoints, "Where does this figure come from?".
Charts and mapsThe reference visuals, with their data and sources.
ToolsOpen calculations, your assumptions on the left, the result on the right.
A question about a figure?
An error, a source to add, a methodology question, a press request: write to us. Guillaume Pakula, Clément Reynaud, Sébastien Pierfederici and Alexandre Veilly will reply.
Write to us →
Projet Celsius, which runs this hub, also helps companies measure and reduce the footprint of their AI use.
Accounting for AI in your Bilan Carbone® ↗
