What is green AI?

Green AI is the practice of designing, training, deploying and using artificial intelligence with environmental efficiency as an explicit goal. It asks how much useful performance an AI system delivers for the energy, carbon, water and hardware it requires—not only how accurate or capable it is.

The term does not mean an AI product has zero impact. Nor does renewable electricity, an offset or a tree-planting programme automatically make a model “green”. A credible green-AI claim identifies the task and life-cycle boundary, measures or estimates resource use, reduces unnecessary computation, uses lower-impact infrastructure and reports what remains uncertain. It also considers whether the AI produces enough social or environmental value to justify its own footprint.

Green AI began as a change in incentives

The influential 2019 paper Green AI contrasted efficiency-focused work with “Red AI”: research that pursues better results mainly by increasing computation. Its authors argued that efficiency should become a standard evaluation criterion alongside accuracy, and that researchers should report the computational or financial cost of their results.

That proposal was partly about climate and partly about access. If progress requires ever-larger compute budgets, fewer universities, small companies and independent researchers can participate. Efficient methods can lower environmental pressure and make experimentation more inclusive.

The label has broadened since then. It now commonly covers model architecture, data, hardware, data-centre operation, electricity sourcing and the way people use a service. The original insight remains useful: capability without a resource denominator gives an incomplete picture of progress.

It also changes what counts as an improvement. Matching a benchmark with less compute can be meaningful progress even when the headline accuracy does not move.

Green AI and AI for green are different

Green AI reduces the footprint of AI itself. “AI for green” uses AI to pursue an environmental outcome, such as forecasting renewable generation, detecting methane leaks or optimising building controls. A project can belong to either category, both or neither.

An efficient language model used for advertising is still an example of greener computation, even if its purpose is not environmental. A large model used for climate research may be AI for green, but still have a significant footprint. Its net value depends on whether the claimed benefit is real, additional and greater than the resources used.

UNESCO’s AI and environment framework uses both axes: mitigate AI’s negative socio-environmental effects and use AI ethically, where needed, to support sustainability and resilience. Keeping them separate prevents a possible benefit from becoming a blanket excuse for unmeasured consumption.

For an AI-for-green project, the comparison should include a realistic non-AI baseline and any rebound effect. A possible application is not the same as a measured environmental saving.

The boundary extends beyond electricity

Operational energy is the most visible part of green AI, but it is not the whole life cycle. Electricity has associated emissions that vary by grid and time. Data centres may consume water directly for cooling and indirectly through electricity generation. Chips, servers and buildings require raw materials, manufacturing and transport, then eventually become waste.

Training and inference should also be distinguished. Training is a concentrated process that creates a model. Inference occurs whenever the trained model answers, classifies or generates something. A heavily used service may accumulate large inference demand even if each response is efficient.

A narrow per-query estimate can help compare day-to-day choices. It cannot establish the complete environmental impact of a service. A practical guide to sustainable AI shows how to define a boundary without pretending unavailable data is known.

Teams can report these layers separately when a full inventory is impossible. Separating operational, embodied and water estimates is more informative than merging incompatible evidence into one precise-looking score.

Efficiency has several layers

At model level, developers can use better data, smaller architectures, pruning, quantisation, distillation and sparsity. A small language model may be sufficient for a constrained task. A mixture-of-experts model can activate only selected parts of a larger network for each token.

At serving level, modern hardware, efficient software kernels, batching and high utilisation can reduce energy per completed task. At application level, routing, caching, limited context and sensible output caps prevent unnecessary calls. At infrastructure level, low power overhead and lower-carbon electricity reduce operational impact.

None of these levers guarantees lower total use. If efficient, cheap AI is placed in every product interaction, growth in volume can exceed savings per answer. Green AI therefore measures absolute consumption as well as intensity.

Product design belongs in this list. Making an expensive mode deliberate, caching a stable answer and removing an unnecessary generation can save more than optimising a model call that should never have occurred.

Measurement should follow useful work

Comparing models only by parameter count or a laboratory power reading misses whether they solve the task. The denominator should represent useful output: a successful classification, a resolved support case or an accepted draft. Count retries and tool calls, and hold quality and safety requirements constant.

The Green Software Foundation’s Software Carbon Intensity specification expresses software emissions per functional unit. Its framework includes energy, the carbon intensity of electricity and an allocation for embodied hardware emissions. Crucially, offsets do not lower the score; only changes that use less energy, less hardware or lower-carbon electricity do.

Public AI data is often incomplete. Providers may not disclose exact hardware, utilisation, facility location or model power. Estimates remain useful when assumptions and exclusions travel with the result. False precision is less informative than an honest range.

Measurements also age. Model versions, hardware and traffic change, so a responsible assessment records its date and is recalculated rather than becoming a permanent marketing claim.

A green claim should survive basic questions

Ask what became more efficient, compared with which baseline and over what period. Check whether the figure covers training, inference, data-centre overhead and hardware. For electricity claims, distinguish the physical grid from annual renewable matching. For carbon claims, ask whether reductions and offsets are shown separately.

Also look for trade-offs. Dry cooling can reduce direct water consumption but sometimes increase electricity demand. New hardware can improve operational efficiency while adding manufacturing impacts. A small model can save energy but cause more retries if it is not capable enough.

Treechat uses “greener” as a direction, not a zero-impact label. Everyday questions are routed towards smaller models, and each answer receives a model-based estimate. Tree funding is reported separately rather than subtracted from the receipt. The assumptions, limitations and current factors are on our methodology page.

Green AI is ultimately disciplined engineering and honest accounting. It asks teams to deliver the required benefit with less computation, lower-impact infrastructure and fewer unsupported claims—and to keep measuring as the product and its use grow. Individuals can apply the same idea through the steps in how to reduce your AI carbon footprint.