Is generative AI bad for the environment?

Generative AI has a real environmental cost: training and running models uses electricity, cooling may consume water, and the hardware has a manufacturing footprint. It is often more computationally demanding than a specialised system doing one narrow task. But the size varies enormously. A short text reply from a small model is not equivalent to a long reasoning run, a batch of images or generated video.

So generative AI is not environmentally harmless, but neither is every use exceptionally damaging. Its impact becomes most concerning when heavy models generate low-value or disposable output at huge volume, particularly on carbon-intensive grids or in water-stressed places. The right comparison is with a suitable alternative and the value of the result.

Generation does more work than simple retrieval

A traditional database can return stored information. A task-specific classifier might select one label. A generative model calculates a new sequence of tokens, image pixels or video frames from learned patterns. Larger outputs and more complex modalities generally mean more computation.

In Power Hungry Processing, researchers Luccioni, Jernite and Strubell compared 1,000 inferences across representative tasks. They found multi-purpose generative architectures substantially more energy-intensive than task-specific models, even after considering model size. Their measured values should not be pasted onto every commercial service: hardware, utilisation and software differ. The study’s more durable lesson is that generality carries a cost.

Sometimes that cost is worthwhile. One capable model can replace several tools or help a person finish valuable work. Environmental judgement requires the whole task, not just the label “generative”.

Training is visible, but inference accumulates

Training a large model coordinates many processors over an extended period. It can be repeated for experiments, fine-tuning and failed runs. Public model cards rarely provide enough energy, location and hardware data for an independent life-cycle total.

Inference begins after deployment. Each request is smaller than a training run, but a popular service may handle an enormous number. Long context, hidden reasoning tokens, retrieval, web browsing and tool calls can multiply the work behind one visible answer. Whether training or inference dominates over a model’s life depends on traffic and how long it remains useful.

Allocating training equally to each query can be informative for accounting, but it is not electricity drawn when the user presses send. A young model with uncertain future traffic makes that allocation especially unstable. Good reporting keeps training, inference and embodied hardware separate before offering any combined figure.

Text, images and video are not interchangeable

“One generation” is a poor unit. Text is produced token by token. A concise answer with limited context may use a fraction of a watt-hour under recent estimates, while a long document or reasoning task can use much more. Our explainer on AI’s environmental impact shows why boundary and workload matter.

Images involve iterative computation over a large array of values. Users often create several candidates and upscale a favourite. Video adds many frames, temporal consistency and usually greater computational demand. Commercial providers publish little comparable, product-level energy data, so precise ratios between modes would be guesswork.

This is a place where behaviour helps. Specify the result you need, generate fewer variants and avoid upscaling drafts. For text, request an appropriate length and begin a fresh conversation when old context is irrelevant.

Electricity, carbon and water tell different stories

Electricity measures energy. Carbon depends on how that electricity was generated. Water can be consumed on-site for cooling and indirectly by power generation. A service can improve one dimension while worsening another—for example, dry cooling may save local water but require more electricity in some conditions.

The IEA estimates all data centres used about 415 TWh in 2024, not AI alone. It expects AI-driven accelerated servers to be a major source of growth towards 2030. System totals show why aggregate demand deserves attention even when an ordinary text prompt is small.

Location matters at least as much for water. A litre consumed in a stressed catchment can carry greater consequences than the same volume in a water-abundant region. Global averages hide those local decisions. Read how AI affects the environment for the complete chain.

Efficiency is improving, but demand can outrun it

New accelerators can perform more calculations per unit of electricity. Quantisation reduces numerical precision and work. Mixture-of-experts models activate only part of their parameters for each token. Better batching spreads server power across multiple users, and smaller models can handle ordinary tasks.

These are genuine improvements. They do not guarantee lower total consumption. Cheaper, faster generation can attract more users and enable more demanding products. This is the rebound problem: energy per output falls while the number and size of outputs grow faster.

Both measures belong in the same report. Providers should disclose per-workload efficiency and absolute energy, water and emissions. A falling prompt average is not proof that total impact is falling; a rising data-centre total is not proof that every prompt has become less efficient.

A proportionate way to use generative AI

First decide whether generation is needed. Search is often better for finding a known page; a spreadsheet is better for transparent arithmetic; a template may beat regenerating routine copy. When AI adds value, select the smallest capable model rather than defaulting to the most powerful.

Keep prompts and attachments relevant. Ask for one strong direction before requesting variations. Reuse good outputs instead of starting over. Save image, audio and video generation for results you intend to use. These choices will not solve infrastructure impacts, but they reduce avoidable demand.

Ask vendors what their numbers include and whether they are measured or modelled. Treechat’s methodology describes an estimate based on model choice, tokens, data-centre overhead and average grid carbon intensity. It cannot see a provider’s live power meter, water system or full hardware life cycle.

Generative AI is environmentally costly enough to use deliberately, not so uniquely catastrophic that every valuable use should stop. The deeper question in how bad AI really is is whether the tool, scale and outcome justify the resources. That answer can change from one prompt to the next.