AI has real environmental costs. Training and running models increases electricity demand, data centres consume water for cooling, and specialised hardware depends on minerals with mining impacts on land and at sea. At the same time, AI can accelerate research in health, materials and climate science. The outcome depends on choices: model size, clean energy use, storage, transparency, and how we curb wasteful usage.
When people ask “is AI bad for the environment”, they are usually referring to four things:
Electricity for data centres
Water used for cooling
Mining of minerals and rare earth elements for chips and servers
Wasteful usage patterns that drive needless compute and storage
Each area is summarised below, with links so you can go deeper.
Modern AI runs on high performance data centres. The International Energy Agency estimates data centres used roughly 415 TWh of electricity in 2024 and that AI is a growing share of this total. See the IEA’s overview and demand page:
In the UK, the Department for Energy Security and Net Zero has published research on how data centre growth could affect national electricity demand, plus where digital services can replace higher impact activities:
Report hub: https://www.gov.uk/government/publications/impact-of-growth-of-data-centres-on-energy-consumption
The House of Commons Library provides an accessible briefing on UK data centres:
Yes, increasingly. Operators sign long term power purchase agreements for wind and solar, and some now add firm clean sources such as geothermal. For example, Google’s 24/7 carbon free energy programme explains hourly matching and why annual certificates are not enough:
24/7 overview: https://sustainability.google/stories/24×7/
Technical paper: https://sustainability.google/reports/24×7-carbon-free-energy-data-centers/
Real world geothermal expansion for data centres has also been reported:
Wind and solar are variable. Until grid scale storage is deployed at higher levels, grids backfill with flexible fossil generation, often gas. Good primers on storage and system stability:
IEA storage explainer: https://www.iea.org/energy-system/electricity/grid-scale-storage
IEA batteries report: https://www.iea.org/reports/batteries-and-secure-energy-transitions
For UK policy on long duration storage, see Parliament’s Science and Technology Committee materials:
Bottom line: green energy can power AI, but today most grids still rely on some fossil generation during calm or dark periods. More renewables plus more storage is the route to cleaner AI.
Large cloud campuses often use evaporative cooling, which consumes water directly, and they also use water indirectly through power stations that need cooling. A clear explainer and a UK-focused summary:
EESI primer: https://www.eesi.org/articles/view/data-centers-and-water-consumption
UK executive summary: https://assets.publishing.service.gov.uk/media/688cb407dc6688ed50878367/Water_use_in_data_centre_and_AI_report.pdf
Water risk is local. Ask for site level water data and seasonal variation, not only global averages.
GPUs, memory and servers depend on complex supply chains.
Rare earth processing can produce significant waste streams if not well managed. See a life cycle review and health impacts perspective:
Lithium from brines in Chile’s Atacama raises concerns over water balances and subsidence:
Cobalt and copper are concentrated in the DRC, where human rights organisations have reported abuses around industrial sites. Due diligence is essential:
These materials also serve other sectors like vehicles, phones and grid storage, but rapid AI growth adds additional demand for high end chips.
As demand for critical minerals grows, some companies are exploring mining in the deep ocean. Three deposit types are usually discussed:
Polymetallic nodules on abyssal plains such as the Clarion Clipperton Zone, containing manganese, nickel, cobalt and copper, with trace rare earth elements
Cobalt rich crusts on underwater mountains
Seafloor massive sulphides near hydrothermal vents
Accessible primers:
IUCN issues brief: https://iucn.org/resources/issues-brief/deep-sea-mining
IUCN short PDF: https://iucn.org/sites/default/files/2022-07/iucn-issues-brief_dsm_update_final.pdf
IUCN Netherlands explainer: https://www.iucn.nl/en/story/the-impact-of-deep-sea-mining-on-biodiversity-climate-and-human-cultures/
Key concerns in plain English
Habitat removal: Nodules are hard surfaces where some animals live. Removing them can mean very slow recovery.
Sediment plumes: Collecting nodules stirs up fine particles that can drift widely and smother marine life.
Noise and light: Continuous industrial activity in a dark, quiet ecosystem has uncertain effects.
Are nodules a major source of rare earths? Not primarily. They are mainly targeted for manganese, nickel, cobalt and copper. Rare earth content is typically low. Crusts and some sulphide deposits may have more rare earths, but extraction and processing at scale remain open questions.
How this links to AI: AI does not directly cause deep sea mining, but demand for chips, servers and backup systems contributes to global demand for these metals. Policy choices on sourcing standards and recycling will influence whether seabed mining proceeds and under what safeguards.
There is a growing wastage factor when generative tools are free or very low cost. When the price per query falls, people tend to generate more content than needed. Economists call this a rebound effect. In practice, this leads to:
Lots of disposable prompts and images that never deliver value
Extra cycles in inference and storage
A surge of low quality automated content, often called Ai slop.
Readable introductions and studies on rebound effects in digital systems and AI:
Tech policy explainer: https://techpolicy.press/jevons-paradox-makes-regulating-ai-sustainability-imperative
Peer reviewed work on AI rebound effects: https://dl.acm.org/doi/10.1145/3715275.3732007
Review article on energy rebound in AI systems: https://www.frontiersin.org/journals/energy-research/articles/10.3389/fenrg.2025.1460586/full
Examples of the quality problem and moderation responses:
AI-generated spam coverage: https://www.forbes.com/sites/daveywinder/2025/06/20/ai-is-behind-50-of-spam—and-now-its-hacking-your-accounts/
News on malicious email generation: https://www.infosecurity-magazine.com/news/ai-generates-spam-malicious-emails/
“Slop” entering search and web results: https://www.ft.com/content/27cf6dcb-354a-4456-8216-23ba02112103
arXiv actions on low value survey floods: https://decrypt.co/347196/arxiv-blocks-ai-generated-survey-papers-flood-trashy-submissions
Music platforms removing automated spam uploads: https://www.theguardian.com/music/2025/sep/25/spotify-removes-75m-spam-tracks-past-year-ai-increases-ability-make-fake-music
Practical fixes
Set sensible limits on free tiers and nudge users to right size models
Default to smaller, efficient models for routine tasks
Require human review for public publishing
Apply provenance and deduplication, and clear low value outputs quickly
These steps can cut waste without blocking useful access.
Balanced answers to “is AI bad for the environment” should include what AI enables for science and conservation.
Protein science and drug discovery
AlphaFold overview in Nucleic Acids Research: https://academic.oup.com/nar/article/52/D1/D368/7337620
AlphaFold Protein Structure Database: https://alphafold.ebi.ac.uk/
Materials discovery for batteries, solar and semiconductors
Nature paper on millions of candidate crystals: https://www.nature.com/articles/s41586-023-06735-9
Plain summary: https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/
Faster weather forecasts that aid grids and resilience
GraphCast in Science: https://www.science.org/doi/10.1126/science.adi2336
Accessible explainer: https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/
Applied conservation via the Earthshot Prize
Winners and finalists: https://earthshotprize.org/winners-finalists/
NatureMetrics finalist page: https://earthshotprize.org/winners-finalists/naturemetrics/
These examples do not erase AI’s footprint, but they show where AI can deliver public benefit if used carefully.
If your aim is to reduce impact while keeping value, start here.
Right size your models
Use small or task-specific models wherever accuracy allows. Save large models for cases with clear lift.
Pick cleaner regions and ask for water data
Choose cloud regions with lower hourly grid intensity and ask for site level water metrics.
Schedule flexible work when the grid is greener
Batch inference and training can run when wind and solar output is higher.
Storage and flexibility context: https://www.iea.org/energy-system/electricity/grid-scale-storage
Tighten content quality controls
Limit wasteful use in free tiers, introduce human checks for public content, and automate removal of low value outputs.
Push for responsible minerals and circularity
Ask vendors for proof of due diligence and third party audits. Plan for refurbishment and parts harvesting.
Cobalt and copper context: https://www.amnesty.org/en/latest/news/2023/09/drc-cobalt-and-copper-mining-for-batteries-leading-to-human-rights-abuses/
Look at the whole service, not just servers
Digital workflows can replace higher impact physical processes if well designed.
Does training or day-to-day use consume more energy?
Training very large models is heavy, but over a model’s lifetime, millions of inferences can exceed training energy. Right sizing helps in both phases. See the IEA’s baselines: https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
Is water use a problem everywhere?
No. It is local and seasonal. Ask for site level figures and cooling type.
Is deep sea mining mainly about rare earth elements?
Not at present. Nodules are mainly targeted for manganese, nickel, cobalt and copper.
Can renewables fully power AI today?
They can cover large shares, but variability means grids still rely on some fossil backup until storage is scaled.
Storage explainer: https://www.iea.org/energy-system/electricity/grid-scale-storage
AI can be harmful if built and used without care. Electricity and water footprints are material, mining impacts are real on land and potentially at sea, and wasteful use inflates demand. Yet AI can also speed up research that supports cleaner technologies, better weather resilience and practical conservation. The most useful answer to “is AI bad for the environment” is this: it depends on choices. Choose smaller models where possible, power systems with more renewables and storage, demand transparent water and carbon reporting, enforce quality controls to avoid AI slop, and insist on responsible mineral sourcing and circularity. That is how to keep the benefits while managing the costs.
You must be logged in to post a comment.