AI is not literally destroying the environment, but the infrastructure used to develop and run it has real environmental costs. Data centers need electricity and cooling; their hardware requires raw materials; and discarded equipment adds to electronic waste. The scale of these impacts depends on how AI is built and used, where its infrastructure is located, and how that infrastructure is powered.
What are the main negative environmental impacts of AI?
The environmental cost of AI spans its lifecycle—not just the electricity used to train a model or answer a prompt. It includes making and powering servers, cooling data centers, extracting materials for electronics, and eventually replacing and disposing of equipment. These effects can grow as AI services and the facilities that support them expand.
- Electricity and greenhouse-gas emissions: AI systems run on computing equipment in data centers. When the electricity comes from fossil fuels, generating it releases greenhouse gases. The impact varies with the amount of computing required and the electricity mix supplying the facility.
- Water demand: Data centers may use water directly for cooling, while the power plants supplying their electricity can also use water. The amount and local consequences vary by facility and location.
- Raw materials and mining: Servers and chips require materials to manufacture. The UN Environment Programme’s overview of AI’s environmental footprint notes that data-center hardware relies on raw materials and that microchips require rare earth elements, which can be mined in environmentally damaging ways.
- Electronic waste: Replacing and disposing of computing equipment creates e-waste. UNEP notes that discarded data-center electronics can contain hazardous substances such as mercury and lead.
These are infrastructure-related impacts, not a fixed environmental cost for every AI tool or request. The footprint of a particular service depends on factors such as its model, hardware, usage, electricity supply, cooling system, and equipment lifespan. The U.S. Government Accountability Office (GAO) also cautions that companies generally do not report full details of generative AI’s resource use, making its specific footprint difficult to measure.
How does AI use water, and why can it matter locally?
Data centers can use water directly to cool computing equipment. There is also an indirect water footprint: the electricity powering a data center may come from power plants that use water to generate electricity. As Southern New Hampshire University’s overview of AI’s environmental impact explains, both data-center cooling and the energy supply can contribute to water demand.
The impact is not the same everywhere. It depends on a facility’s cooling method, its power source, and local water availability. A data center drawing on water in a place with limited supplies may add pressure to competing local needs; the same amount of use can have different consequences in a water-rich region. It is therefore misleading to assume every data center uses the same amount or draws from the same type of water supply.
Precise estimates for generative AI’s water use remain limited. GAO says available estimates are sparse and that companies generally do not disclose detailed resource-use information. For that reason, a single water figure for one prompt or model should not be treated as a universal, settled measure.
Does an AI request use more electricity than a Google Search?
Available comparisons suggest that some generative-AI text requests use more electricity than a typical search-engine request, but they do not establish one ratio that applies to every AI model and every search. UNEP reports an International Energy Agency comparison in which a ChatGPT request used 10 times the electricity of a Google Search. Separately, the National Education Association’s summary of AI environmental research cites an estimate of four or five times the energy of a typical search-engine request.
Those estimates differ and should be understood as source-attributed comparisons, not a definitive, like-for-like measurement for all services. Electricity per request is also only one part of the picture: it does not by itself account for the hardware, cooling, electricity generation, or disposal associated with AI infrastructure.
What do the available figures say about AI’s environmental footprint?
One useful figure describes data centers overall, not AI alone. GAO reports that U.S. data centers used approximately 4% of U.S. electricity demand in 2022 and cites an International Energy Agency projection that they could reach 6% in 2026. The 6% figure is a projection, not a verified 2026 result, and neither figure measures generative AI’s share specifically.
GAO says the portion of data-center electricity consumption attributable to generative AI is unclear. More broadly, incomplete company disclosures and limited or inconsistent measurement make it difficult to calculate a complete AI-specific total for electricity, emissions, water, and hardware impacts. These figures show why data-center growth matters, but they should not be presented as a precise measure of AI’s environmental footprint.
Can AI also help the environment?
Yes. UNEP describes AI being used to map destructive sand dredging and to detect methane emissions from oil and gas installations. AI can also support environmental monitoring and help people make decisions using patterns in large datasets.
Those potential uses do not automatically cancel out the resources needed to build and operate AI systems. Whether AI’s benefits outweigh its environmental costs depends on the application and the infrastructure behind it. The evidence supports neither the claim that AI is always harmful nor the assumption that its environmental benefits necessarily make its overall impact positive.
Is ChatGPT destroying the environment?
That wording overstates what the evidence shows. ChatGPT and other generative-AI services rely on computing infrastructure that uses electricity, may require water for cooling, and depends on physical hardware with material and disposal impacts. Those costs are real, but the available evidence does not show that ChatGPT alone is destroying the environment. The more useful question is how much infrastructure is being built and used, what resources it consumes, and whether that impact is being measured and managed transparently.