The claim circulates regularly: using AI is environmentally catastrophic — so bad it cancels out lifestyle choices like being vegetarian. Meanwhile, the most significant lever most individuals have over their personal carbon footprint rarely gets a mention. This post puts the numbers alongside each other, using independent peer-reviewed research and data from sources with no commercial interest in the outcome.
The short answer: the claim about AI is wrong by a significant margin. The longer answer explains exactly why — and where AI’s environmental impact is genuinely worth taking seriously.
A note before we start: This post is not an argument for vegetarianism, and it is not a criticism of anyone’s personal choices. What you eat, how you travel, and how you use technology are decisions that belong to you. The point here is narrower than that — it is about proportion. The conversation around AI and environmental damage has become loud and confident, while other variables with a far larger measurable impact rarely get the same scrutiny. If reducing your environmental footprint is a goal, the data suggests the most effective lever is one most people don’t associate with technology at all. That is worth knowing, whatever you decide to do with it.
1. The diet saving is large and well-evidenced
The most rigorous UK-specific data comes from the EPIC-Oxford cohort study, published in Climatic Change and indexed on PubMed. It assessed the diets of over 55,000 participants across four dietary groups.[1]
“Moving from a high meat diet to a vegetarian diet would reduce the carbon footprint by 1,230 kg CO₂e per year.”
— Scarborough et al., Climatic Change, 2014 (EPIC-Oxford cohort, n=55,504)
A more recent study, published in Nature Food in 2023, linked dietary data from 55,504 participants to food-level lifecycle assessment data covering 38,000 farms in 119 countries.[2] It found that dietary emissions from high meat-eaters were roughly double those of vegetarians. These are not modelled estimates — they reflect actual consumption data from real participants.
Source: Scarborough P. et al., Climatic Change, 2014. EPIC-Oxford cohort, n=55,504. springer.com
The annual saving from switching from a high meat diet to a vegetarian diet is 1.23 tonnes CO₂e. That figure is the baseline against which everything else in this post needs to be measured.
2. Text-based AI is low-impact per query
The most credible non-industry data on per-query energy consumption comes from a peer-reviewed benchmarking study published on arXiv in May 2025, covering 30 commercial LLMs across Anthropic, OpenAI, Meta and DeepSeek.[3]
“A single short GPT-4o query consumes 0.43 Wh… even limited daily engagement with GPT-4o imposes an energy cost comparable to charging two smartphones to full capacity (approximately 10 Wh).”
— Jegham et al., arXiv:2505.09598, 2025
Text generation sits far below image and video generation in energy terms — a distinction that matters significantly when assessing heavy AI users.
AI text query figure: Jegham et al., arXiv:2505.09598, 2025. Everyday comparisons: standard electrical consumption values.
3. Image generation is meaningfully more intensive
The gap between text and image generation is substantial, and backed by independent research from Hugging Face and Carnegie Mellon University.[4]
“Generating images was by far the most energy- and carbon-intensive AI-based task. Generating 1,000 images with a powerful AI model, such as Stable Diffusion XL, is responsible for roughly as much carbon dioxide as driving the equivalent of 4.1 miles in an average gasoline-powered car.”
— Luccioni et al. (Hugging Face / Carnegie Mellon), MIT Technology Review, December 2023
A 2024 analysis from Scope3 found that image generation via GPT-4o uses up to 30 times more energy than a standard text interaction.[5] The least efficient image generation model measured consumed 11.49 Wh per image — roughly half a smartphone charge per image generated.[6]
Sources: Luccioni et al., arXiv:2311.16863, 2023 (Hugging Face / Carnegie Mellon). Average image figure from Wikipedia, Environmental impact of artificial intelligence, citing Luccioni, Jernite, Strubell (2024).
4. Video generation is the most intensive AI task
Research from Hugging Face published in 2025 found that video generation energy costs do not scale linearly with length.[7]
“The energy demands of text-to-video generators quadruple when the length of a generated video doubles — indicating that the power required for increasingly sophisticated generations doesn’t scale linearly.”
— Hugging Face researchers, reported in Futurism, September 2025
A late 2025 analysis cited in TIME found that generating an eight-second video could consume as much electricity as charging a laptop twice.[8] Estimates for Sora 2 put individual video generation at approximately 466 grams of CO₂e per output.[9]
Source: Hugging Face researchers, reported in Futurism, September 2025. futurism.com
5. Agentic and reasoning AI uses more energy than standard chat
Standard language models predict the next token in sequence. Reasoning models and agentic systems that chain multiple calls operate differently, generating large volumes of intermediate computation before producing a final output.[10]
“Reasoning models generate thousands of hidden tokens to consider a question before producing a visible response, which dramatically multiplies energy costs.”
— Earth911, based on independent benchmarking, March 2025
The arXiv benchmarking study found that o3 and DeepSeek-R1 consume over 33 Wh per long prompt — more than 70 times the consumption of smaller optimised models.[3] For users running agentic workflows regularly, this is the category most likely to push their personal AI footprint upward.
Source: Jegham N. et al., “How Hungry is AI?” arXiv:2505.09598, May 2025. arxiv.org
6. AI water consumption — the honest picture
Water use is a legitimate part of the AI environmental argument and deserves the same honest treatment as energy. Data centres use freshwater for cooling, and the figures are real. Training GPT-3 at Microsoft’s data centres consumed an estimated 700,000 litres of water — a significant one-time cost.[15] Inference (answering user queries) adds to this continuously.
Per-query figures are contested, because methodologies differ significantly depending on what is counted. The most widely cited estimate, from a 2023 University of California study reported by the OECD, puts water consumption at approximately 500 millilitres per 10 to 50 queries — depending on the data centre, cooling system, and region.[16] A more conservative independent analysis puts the figure closer to 5 millilitres per query for modern systems with efficient cooling.[17] The gap between these estimates reflects how much the cooling technology and location of a data centre matters — not a fundamental disagreement on the underlying physics.
“Nearly half of the world’s 9,000+ data centres are already located in regions of high water stress.”
— IE Insights, citing Morgan Stanley projections, 2025
That geographic concentration is the most legitimate water concern. A data centre drawing from an already-stressed aquifer is a different problem to one in a water-abundant region using recycled cooling water. The water footprint of food production — and meat in particular — dwarfs AI at the individual level. It is worth being precise here, because the widely cited figure of 15,000 litres per kilogram of beef includes green water (rainfall absorbed by pasture), which is part of a natural cycle and not drawn from freshwater reserves. The more meaningful figure for water scarcity is blue water — surface and groundwater actually consumed. Research from Cranfield University calculated the UK national average at 67 litres of blue water per kilogram of beef carcass, rising to approximately 2,000 litres per kg in irrigated US production systems.[18] Food production accounts for approximately 70% of all freshwater withdrawn from the environment globally.[18]
AI water figures: Li P. et al., UC Riverside, 2023, via OECD.AI; Goedecke S., independent analysis, 2024. Beef blue water figure: Cranfield University, reported in The Conversation, April 2024. Total beef water footprint: Mekonnen and Hoekstra, global average 15,400 L/kg.
7. A worst-case annual estimate for a heavy AI user
The table below builds a worst-case annual estimate for someone using AI heavily — including regular image generation, video generation and agentic tasks — against the annual diet saving from being vegetarian.
| Use type | Assumption (worst case) | Est. annual CO₂e |
|---|---|---|
| Text / chat (daily use) | 20 queries/day @ 0.5 Wh, UK grid | 0.07t |
| Image generation | 30 images/month @ 3 Wh avg | 0.01t |
| Video generation | 5 videos/week @ 466g CO₂e | 0.12t |
| Agentic / reasoning AI | Heavy daily use, 10× text multiplier | 0.15t |
| Total | ~0.35t |
Note: This is a worst-case estimate for a user generating images and videos regularly and running agentic AI workflows daily. Most AI users who rely primarily on text chat would sit well below 0.1t annually.
The diet saving from being vegetarian rather than a high meat-eater is 1.23 tonnes per year, backed by peer-reviewed research on 55,000 UK participants. The worst-case annual AI footprint for a heavy user is approximately 0.35 tonnes. The diet saving is roughly 3.5 times larger.
8. How AI usage compares to my full personal carbon footprint
The chart below compares my personal carbon footprint against the UK average of approximately 11 tonnes CO₂e per year, using a worst-case AI figure of 0.35t. My profile: vegetarian diet, renewable home energy (Ecotricity), London Underground and walking as primary transport, petrol car used infrequently for longer distances, one return flight to the US and one to Greece annually, low clothing consumption, and heavy AI use at worst-case estimate.
My profile: vegetarian diet, Ecotricity renewable energy, London Underground / walking, infrequent petrol car use, one return flight to the US + one to Greece annually, low clothing consumption, heavy AI use at worst-case estimate. UK average: ~11t CO₂e. Sources: DEFRA UK average; EPIC-Oxford cohort (Scarborough et al., 2014) for diet figures; DEFRA GHG conversion factors for flights and transport; AI estimates from Jegham et al. (2025) and Luccioni et al. (2023).
9. The diet saving vs worst-case AI footprint, head to head
Diet saving source: Scarborough et al., Climatic Change (2014), EPIC-Oxford cohort, n=55,504. AI figure: worst-case estimate for heavy user (image, video, agentic), based on Jegham et al. (2025) and Luccioni et al. (2023).
10. The legitimate concern about AI — and why it’s a different argument
None of the above means AI’s environmental impact is trivial at a systems level. The IEA’s 2025 Energy and AI report is the most authoritative macro picture available.[11]
“Data centres accounted for around 1.5% of the world’s electricity consumption in 2024, or 415 terawatt-hours (TWh)… global electricity consumption for data centres is projected to double to reach around 945 TWh by 2030.”
— International Energy Agency, Energy and AI Report, 2025
That is a systemic infrastructure concern — one that sits with technology companies, data centre operators, and energy regulators. It is not the same argument as “your personal AI usage is cancelling out your diet choices.” Those are separate claims, and conflating them obscures both.
A note on data quality: Transparency in AI energy reporting is poor. Most major model providers do not disclose sufficient data to accurately estimate their total energy use or carbon footprint. The figures used in this post are from independent researchers working with publicly available data. Real-world usage — including overhead, failed requests, idle compute, and infrastructure inefficiency — may run higher than controlled lab conditions. The AI figures here are best available estimates, not precise measurements.
Source: International Energy Agency, Energy and AI Report, 2025. iea.org. Data centre figure represents projected growth 2024–2030 in IEA base case.
11. The variable nobody mentions: flights
In any honest personal carbon accounting, aviation deserves more scrutiny than AI. Applying DEFRA’s GHG conversion factors to a return economy flight from London to New York produces approximately 2.24 tonnes CO₂e per passenger before radiative forcing adjustments.[12] With radiative forcing — the additional warming effect of emissions released at altitude — some methodologies put the figure closer to 1.59 tonnes one-way, or over 3 tonnes return.[13]
“A single return flight from London to New York emits nearly 2 tonnes of CO₂ per passenger — more than the total annual emissions of an average person in over 85 countries.”
— Greenly, citing aviation emissions data, 2025
A single transatlantic return flight produces more emissions than five to six years of worst-case AI usage from a heavy user.
Flight figure: DEFRA GHG conversion factors 2024, economy class, London–New York return. AI figure: 6 × 0.35t worst-case annual estimate (Jegham et al., 2025; Luccioni et al., 2023).
Summary
The claim that AI usage cancels out the environmental benefit of being vegetarian is not supported by the data. A worst-case annual AI footprint for a heavy user — including image generation, video generation and agentic workflows — sits at approximately 0.35 tonnes CO₂e. The annual saving from a vegetarian diet compared to a high meat diet is approximately 1.23 tonnes, based on peer-reviewed research on 55,000 UK participants published in Climatic Change and replicated in Nature Food. The diet saving is roughly 3.5 times larger.
AI’s environmental impact is a real issue — at infrastructure scale, and particularly for energy-intensive tasks like video generation. That conversation is worth having. But it is a different conversation from individual usage, and neither one touches the most significant lever available to most people: what they eat, three times a day.
Sources
- Scarborough P. et al. “Dietary greenhouse gas emissions of meat-eaters, fish-eaters, vegetarians and vegans in the UK.” Climatic Change, 2014. EPIC-Oxford cohort, n=55,504. springer.com / PubMed
- Scarborough P. et al. “Vegans, vegetarians, fish-eaters and meat-eaters in the UK show discrepant environmental impacts.” Nature Food, 2023. nature.com
- Jegham N. et al. “How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference.” arXiv:2505.09598, May 2025. arxiv.org
- Luccioni A. et al. (Hugging Face / Carnegie Mellon University). “Power Hungry Processing: Watts Driving the Cost of AI Deployment?” arXiv:2311.16863, 2023. Reported in MIT Technology Review, December 2023. technologyreview.com
- Scope3. “How energy intensive are AI-generated images?” Reported in Sustainable AI, May 2025. sustainableai.substack.com
- Luccioni A., Jernite Y., Strubell E. “Power Hungry Processing” (2024). Per-task energy benchmarking across 88 models. Cited in Wikipedia, Environmental impact of artificial intelligence. wikipedia.org
- Hugging Face researchers. Text-to-video energy scaling study. Reported in Futurism, September 2025. futurism.com
- TIME / IEA data. “The Climate Impact of Different AI Prompts.” TIME, July 2025. time.com
- Reclaimed Systems / Forbes analysis on Sora 2, late 2025. Cited in The Sustainable Agency. thesustainableagency.com
- Earth911. “Your AI Carbon Footprint: What Every Query Really Costs.” March 2025. earth911.com
- International Energy Agency. Energy and AI report, 2025. iea.org
- DEFRA / DESNZ GHG Conversion Factors 2024. Applied via Sustainable Travel International calculator. sustainabletravel.org
- TravelNav emissions calculator, London–New York with radiative forcing. travelnav.com
- Greenly. “Could Aviation Be Decarbonised and Become Sustainable?” 2025. greenly.earth
- Li P. et al. “Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models.” University of California, Riverside, 2023. Reported by EESI and OECD.AI. eesi.org / oecd.ai
- Goedecke S. “Talking to ChatGPT costs 5ml of water, not 500ml.” Independent analysis, October 2024. seangoedecke.com
- Cranfield University / The Conversation. “Here’s how much water it takes to make a serving of beef.” April 2024. Includes UK national average blue water figures. theconversation.com
A note on the production of this post
This post was researched, written and edited in collaboration with Claude (Anthropic) over a single extended session. That session involved approximately 100 exchanges, including web searches, source verification, SVG chart generation, file editing and iterative redrafting.
Applying the independent benchmarks cited in this post — Jegham et al. (2025) for text queries and Luccioni et al. (2023) for energy-intensive tasks — the estimated energy consumption for the entire session is in the range of 15–40 Wh. The equivalent of running an LED bulb for two to five hours. Water consumption, using the UC Riverside / OECD estimate, is approximately 500ml to 2 litres depending on methodology.
These are estimates, not measured figures. Anthropic does not publish per-session telemetry, and real-world overhead — idle compute, failed requests, infrastructure — may push the true figure higher. The uncertainty range mirrors the same transparency problem discussed in section 10.
For comparison: the carbon cost of this session is estimated at under 10 grams of CO₂e. A single beef burger produces approximately 2.5kg.
![Kate Baxter and Emma Jane Palin on Blogging Design [Podcast]](/wp-content/uploads/xportrait-kate-baxter-fabric-of-my-life-emma-jane-palin-003.jpg.pagespeed.ic.rIbdHDEWa6.webp)
Leave a Reply