AI models keep getting more efficient. According to the study by Wu et al. (2022) on Meta's AI systems, hardware and software improvements cut the energy consumption of models by around 20% every six months. And yet total AI consumption keeps growing.
The explanation has a name: the Jevons paradox. When a technology becomes more efficient and cheaper to use, it gets used more, and that increase in use offsets part of the savings. That is exactly what is happening with AI: more servers, more models and more queries.
It is one of the conclusions of the Master's thesis by Lorena Enajas, from our team, on the impact of AI on sustainability. Technical efficiency helps, but it is not enough. Companies that use AI need to manage it with judgement. That is what Green AI means: designing and using AI to get the same result with fewer resources.
These are five practical ways to apply it.
1. Choose the model for the task
Not every task needs the most powerful model. Classifying documents, extracting simple data or summarising a short text can be done with light, fast models that consume much less. Higher-capacity models make sense for tasks that require complex reasoning or a lot of context.
Before adding AI to a process, ask yourself what level of capability it really needs. Always using the largest model by default is one of the most common ways to overconsume.
2. Monitor use and consumption
What is not measured cannot be improved. Tracking how many queries are made, to which models, at what cost and with what response time helps spot inefficient use: processes that run more often than needed, duplicate queries or tasks that a smaller model could handle.
Although the main goal of this tracking is usually cost control, it has a direct effect on energy consumption. Less unnecessary processing means less electricity.
At infrastructure level, matching capacity to real demand also counts. In the case analysed by Wu et al., releasing unused capacity during low-demand periods cut equipment use by up to 25%.
3. Use data centres powered by renewables
The source of electricity changes the result a lot. When we measured the AI Manglai uses, emissions calculated with the market-based method, which accounts for the provider's renewable purchases, were just over 40% lower than with the location-based method.
If you choose a provider or region for your cloud services, consider the electricity mix and renewable energy contracting. This has to be combined with data protection requirements, which also determine where data can be stored and processed.
4. Demand transparency from providers
Most companies use AI as a service, so their impact depends on other people's infrastructure. Managing it requires data: emissions allocated to each customer, calculation methodology, water consumption and which part of the life cycle is included. In this article we detail what to ask for and which scope each figure belongs to.
Make sustainability a criterion when selecting technology providers, alongside performance, cost and service quality. Asking for this data also pushes the sector to publish it.
5. Keep human oversight and look after data quality
AI works from patterns. If input data is incomplete or ambiguous, the result will be worse and the process will have to be repeated, with the consumption that implies. Well-structured data from the source helps AI get it right sooner.
Human oversight is still essential to validate results, especially for information that ends up in a report or an audit. AI is a support tool, and judgement still belongs to people.
How we apply it at Manglai
At Manglai we apply several of these practices on the platform: we use models of different capacity depending on the task, we continuously monitor model use, we run on cloud infrastructure with renewable energy purchases and we combine AI with validation rules and human review. And we measure the result.
AI is neither sustainable nor unsustainable on its own. It depends on how it is designed and used. To go deeper, the summary of Lorena's thesis is available in our resources section.



