A single digital product can have a smaller environmental footprint than its physical equivalent (think of paper documentation replaced by digitally signed contracts). However when we use digital services, we use them at an entirely different scale, which makes that comparison far less straightforward. We might run hundreds of search queries a day, chat with bots, and prompt minor edits to generated images. New capabilities brought by technologies like artificial intelligence create new use cases, which in turn drive up the overall usage.

This is a textbook example of the so-called Jevons paradox. When a process becomes more optimized and efficient, we tend to use it far more — not less. So as AI models become easier to use, cheaper to run (in purely financial terms, more on that shortly) and faster, we reach for them more and more often, wiping out any efficiency gains and increasing, rather than decreasing, the overall cost. If money were the only concern, it would be a smaller problem. It’s important to understand that AI carries serious environmental costs that we need to learn to reduce, just as we work to reduce emissions and waste from every other activity undertaken by individuals and organizations.

Infrastructure, training, and inference: where do the environmental costs of AI actually come from?

The environmental footprint of AI models and tools is about more than energy consumption. As with any digital technology, we need to look at every stage of the so-called lifecycle and the resources required to keep a given service running. Civil society organizations (CSOs) are more often consumers than creators of digital technology and services. To make informed choices about products and services, it helps to understand how they come into being and function in a broader context than just what appears on our screens.

From raw material extraction to hardware production

Before a company can train a new AI model or launch an AI-powered service in data centers, the hardware on which training will take place — and on which the model will later run — must first be built. Manufacturing graphics processing units (GPUs) is extraordinarily resource-intensive: a semiconductor fabrication plant consumes around 38 million liters of water per day to produce the ultrapure water required in chip manufacturing. On top of that comes the extraction of rare earth metals (which frequently entails severe local environmental costs such as soil and groundwater contamination), transportation, and the physical construction of data centres themselves. This dimension of AI's environmental footprint rarely surfaces in public debate. It is harder to measure than energy consumption and forces us to confront AI's ethical costs, such as the lives and health of people who live or work near sites directly tied to mineral extraction or data centre construction.

The flip side of the same problem is e-waste. The relentless pressure for ever more powerful hardware shortens the lifespan of many devices. AI is estimated to generate an additional 1.2 to 5 million tonnes of electronic waste. Meanwhile, global e-waste production is growing five times faster than recycling capacity, and a significant share ends up in landfills in developing countries, where substances such as mercury, arsenic, and lead leach into local ecosystems.

Model training

Training a large language model (LLM) consumes an enormous amount of energy. Training GPT-4 required approximately 25,000 GPUs running for 90 to 100 days. Estimated energy consumption ranged from 51,000 to 62,000 MWh — more than 40 times the energy needed to train the previous model, GPT-3. That is equivalent to the energy consumption of 1,000 American homes over five to six years. While many models grow with each new version (as with the GPT family), numerous companies are working to optimize this process (a commercially high-profile example being the launch of the Chinese DeepSeek model at the end of 2024) or are developing far more efficient models tailored to specific tasks or less powerful hardware. These are known as small language models.

Inference (running the model)

Training commands attention because of its sheer scale, but it is inference, generating responses to millions of user queries, that represents the primary ongoing environmental cost of AI. At the scale of a billion queries per day, even a highly efficient model consuming just 0.42 Wh per short query generates annual cumulative energy consumption comparable to that of 35,000 homes in the United States.

The differences between models are vast, and most of the data, published very selectively by AI companies, concerns text-based models. Image and video generation is considerably more expensive and energy-intensive.

Human labor in the AI supply chain

If we take a broad view of "environment" that includes impacts on human health, then beyond the pollution already mentioned, we also need to talk about the people who evaluate, filter, and fine-tune AI models. This process involves labelling training data, rating model responses, and filtering harmful content — a practice known as RLHF, or reinforcement learning from human feedback. This low-paid work is deliberately kept invisible, outsourced to large subcontractors operating in countries with weak labor protections like Kenya, Uganda, India. The mental health costs are comparable to those faced by content moderators on social media platforms, who are exposed to violent and disturbing material in order to filter it before it reaches users.

How to choose AI tools and models from an environmental cost perspective

We can reduce the environmental costs of using any digital technology. But we need to be clear about which actions have systemic impact, which are individual choices, and which risk becoming greenwashing — superficial or overstated gestures toward environmental responsibility.

As organizations, we have no say in how models are built, but we can choose which ones we use. Access to data on their environmental costs and energy efficiency should be the foundation of those choices. For now, however, that information remains limited. Beyond relying on data, we should also be learning to use AI deliberately and sparingly. At the moment, many individuals and organizations are using AI (generative AI in particular) for tasks where it is not necessarily the best-fit solution. It is worth distinguishing between a few core AI technologies that matter from an environmental cost perspective:

• LLMs (large language models) are model families like GPT, Claude, Gemini, and DeepSeek. Trained on hundreds of billions of parameters, they run on the servers of large technology companies. The cost of training is astronomical, but it is spread across millions of users. Access happens through an online interface (websites and mobile apps) or via API for custom tools (such as connecting an AI model to a CRM system), which means every query passes through an external data center. These models are highly capable, but for simple tasks they can consume more resources than makes sense — for example, asking a chatbot a question that a basic search engine query would answer just as well.

• SLMs (small language models) are models such as Mistral Small 24B, Phi-4 14B, or Gemma 2B. They have fewer parameters, run faster, and can be deployed locally — on a laptop, an organization's own server, and in some cases even a smartphone. For many common tasks such as document classification, summarization, and answering questions from an organization's own knowledge base, they are entirely sufficient. You can try these models using tools such as LM Studio or Ollama.

• Open-weight models, such as Llama, Qwen, and Mistral, release their parameters and can be run locally. Beyond giving organizations control over where and when the model operates, those needing more tailored solutions can fine-tune such a model to their specific needs.

A few practical examples of matching AI tools to the task at hand:

• Repetitive queries and tasks: such as searching documentation and answering questions based on an organization's own materials, can be handled by a local model running on the organization's server, drawing on a defined knowledge base. These are known as RAG systems (retrieval-augmented generation).

• Transcribing a recording: this does not require firing up a large cloud-based language model like ChatGPT or Gemini, which will consume far more resources than, say, a locally running tool like OpenWhispr.

• For voice-command computer control: something the makers of large commercial AI models actively encourage, you can use a local, open-source tool like Handy.

• For editing and proofreading text: small locally run language models work well and require no internet connection.

Choosing a smaller or specialized model for a specific task can meaningfully reduce the environmental costs of AI use within an organization. Local models are environmentally beneficial when they run on adequately capable hardware, or when the alternative is intensive use of a large commercial model for tasks a small model could handle just as well. An added advantage of local models is greater control over when they actually run — many AI services, especially those bundled into office suites, run in the background even when you don't need them.

In many situations, however, the most effective way to reduce AI's environmental costs is simply not to use it, or to complete the task in a traditional way. To produce a multilingual video, for instance, standard subtitles usually do the job or, at most, recording and dubbing a new audio track. Generating entirely new AI videos for each language, in which both the video and voice are first cloned and then regenerated, will in many cases be unnecessary and wasteful.

Preparing for the future

The energy consumed by the data centres that support AI training and inference represents only part of the energy used by the digital sector as a whole. Demand from both is growing, and even if AI ends up accounting for just a few percent of total CO₂ emissions, NGOs should be demanding more, from the companies that provide these technologies and from themselves, than simply watching the electricity bill. We can take advantage of what AI models have to offer, but both from an environmental standpoint and in the interest of organizational independence, it is worth using them deliberately and intentionally. That sometimes means stepping back from whatever is most popular on the market, or from whatever we reach for by default because an advertisement told us to.

Your Feedback Matters

What did you think of this text? Take 30 seconds to share your feedback and help us create meaningful content for civil society!

Disclaimers

This piece of resources has been created as part of the AI for Social Change project within TechSoup's Digital Activism Program, with support from Google.org.

AI tools are evolving rapidly, and while we do our best to ensure the validity of the content we provide, sometimes some elements may no longer be up to date. If you notice that a piece of information is outdated, please let us know at content@techsoup.org.