How bad is AI for the environment, really?
A myth audit of AI energy, water and carbon claims: what is supported, what is outdated and what remains genuinely concerning.
AI’s environmental impact is real and growing, but many viral comparisons are too certain. Ordinary text prompts appear to use a fraction of a watt-hour under recent estimates; long reasoning, large context, images and video can use far more. Data-centre electricity and water demand matter at scale, especially in particular regions.
The most honest verdict is neither panic nor dismissal. Some popular numbers are outdated, stripped of their assumptions or applied to the wrong unit. Meanwhile, the system-level trend—rapid infrastructure growth led partly by AI—is well supported. This myth audit separates what should be retired from what deserves serious attention.
Myth: every AI query uses 3 Wh
The 3 watt-hour figure was an early estimate built around older A100 hardware, a GPT-3.5-scale model and an assumed 2,000 generated tokens. It was not a universal measurement.
Newer public estimates for an ordinary text query are lower. OpenAI chief executive Sam Altman stated an average ChatGPT query used about 0.34 Wh. Google reported 0.24 Wh for a median Gemini Apps text prompt in May 2025 using a methodology that included supporting infrastructure.
Neither number covers every mode, and Altman did not publish the calculation behind his average. Long inputs and outputs can still use many times more. Retire “3 Wh for every query”; keep asking which workload and boundary.
Myth: one message drinks a bottle of water
The academic estimate behind the bottle claim did not say one message. Li and colleagues estimated that a 500 ml bottle could correspond to roughly 20–50 medium-length responses, depending on where and when the model ran. That works out to about 10–25 ml per response under their assumptions.
Later operator figures were much lower: Altman gave about 0.000085 US gallons—roughly 0.32 ml—for an average ChatGPT query, while Google reported 0.26 ml for its median Gemini text prompt. These are not clean refutations because the systems and boundaries differ.
The conclusion is uncertainty, not zero. Water varies with cooling, weather and electricity. See does AI really use a bottle per message? for the full trail.
Myth: AI always uses ten times a web search
This ratio usually places an old Google search estimate beside a newer AI estimate. Search infrastructure has changed since Google published energy figures in 2009, and many search products now include generated summaries. Meanwhile, “AI query” could mean a short completion or a multi-step research job.
A ratio assembled from different providers, years and system boundaries is not a controlled comparison. It may once have been a useful illustration, but it should not be treated as a current physical constant.
For an individual task, the more relevant question is whether search, a direct source or a generated answer is the right tool. At system level, compare transparent, contemporary totals. Precision without comparability is not accuracy.
Myth: renewable power makes AI impact-free
Low-carbon electricity can sharply reduce operational emissions, and technology-company contracts can help build new generation. The IEA expects renewables to meet nearly half the growth in data-centre electricity supply to 2030 in its base case.
It also expects natural gas and coal together to meet more than 40% of additional supply over that period. Annual renewable matching may not align generation with the data centre’s place and hour of demand. Cleaner electricity also does not remove cooling water, hardware manufacture, buildings or land.
“Lower-carbon operation” is a defensible claim with evidence. “No footprint” is not. The distinction is central to whether AI is bad for the environment.
Myth: a small prompt cannot matter
One short prompt has a small operational footprint. That is a reason for proportion, not denial. Billions of uses, growing output lengths and new image, video and agentic workloads add up. Providers build capacity for expected aggregate demand, not one person’s message.
The IEA estimated all data centres—not AI alone—used around 415 TWh in 2024 and projects roughly 945 TWh in 2030 in its base case. AI-led accelerated servers account for almost half the net increase.
Individual guilt is still the wrong policy. Operators choose models, hardware, siting and energy; governments oversee grids and water. Users can avoid waste, while structural decisions determine most of the system’s trajectory.
Myth: AI is the only reason data centres are growing
AI is an important driver, not the entire sector. Data centres also provide storage, video, financial systems, enterprise software and conventional cloud computing. Repeating the full data-centre total as “AI electricity” overstates current attribution.
The reverse error is to ignore AI because precise allocation is difficult. The IEA models accelerated-server electricity growing much faster than conventional-server demand and identifies AI as the leading source of expansion.
Good reporting keeps both facts together: observed totals are broader than AI, and AI materially changes the outlook. Our article on AI data centres explains why local grid and water impacts may be more informative than one global percentage.
What is genuinely concerning
Rapid absolute demand can outrun efficiency and clean-energy construction. Facilities cluster, so local grids and water systems may face pressure that global averages disguise. Supply-chain emissions and hardware manufacture remain poorly allocated. Public model-level operational data is still sparse.
Generative media and agentic systems could make each user request more computationally complex. Rebound may turn cheaper inference into more total demand. Those are planning risks rather than proof of the worst scenario, which is why scenario ranges matter.
Users can choose smaller capable models, keep context focused and avoid disposable generations. Providers should publish workload-specific and absolute totals. Treechat explains its limited, model-based estimates on its methodology page. A receipt can encourage better choices, but it is not a full life-cycle audit.
AI is bad enough to require transparency, cleaner infrastructure and restraint. The evidence does not support treating every prompt as an environmental emergency. It supports building a culture in which expensive computation has to earn its keep.