This blog is part of the Good practice in research software development in the age of AI series.
A few weeks ago, I made my kids watch “Hidden Figures”, the 2016 film based on Margot Lee Shetterly’s book, which follows three African-American mathematicians, Katherine Goble Johnson, Dorothy Vaughan and Mary Jackson, who worked as “human computers” at NASA’s Langley Research Center during the Space Race. It brings to life the race and gender barriers they faced in the early 1960s, and while it takes some liberties with the history, dramas like this make the human cost of those barriers far easier to understand (plus the kids needed a break from watching dross on YouTube).
Halfway through the film, an IBM mainframe arrives promising to do the women’s calculations in a fraction of the time, with obvious consequences for jobs. Like computers in the 1960s, the introduction of AI to almost every aspect of research feels inevitable. There is huge promise, but how it will change roles is unclear, and likely to bring both upsides and downsides. One of the largest downsides is the potential environmental cost.
Everyone knows there is an environmental cost to AI, but it’s poorly quantified, often invisible and frankly very easy to forget about.
While AI in research might be inevitable, we must minimise its environmental impact by considering carefully when and how we use it, to ensure that the benefits outweigh the costs.
What can you do? Large Language Models
Large Language Models (LLMs) such as ChatGPT, Claude and Gemini are many people’s first exposure to AI. Researchers use them to summarise literature, draft text for revision or assist with coding. These tools still feel a bit like magic, but every query is a series of statistical calculations predicting the most likely next word based on patterns in data. That process consumes real computing power and electricity for every token generated.
It is estimated that in terms of operation alone, global data centres are currently responsible for 0.5% of global CO₂ emissions (and just over 1% of global electricity consumption). This is projected to rise to 1% by 2030 under central estimates, or up to 1.4% in a faster-growth scenario (Carbon Brief / IEA).
The environmental impact of LLMs comes both from the training phase (where large amounts of data are used to create the model) and the inference phase (where the users actually use the model).
LLMs good practice tips
- Choose a model that fits the task. Small models are fine for small tasks. Keep the largest models for complex reasoning.
- Write clear, focused prompts, and batch related questions. State the format, length and audience up front to cut down on back-and-forth. One well-written prompt usually costs less than a long series of small ones.
- Start a new thread when you don’t need the history, as the full conversation is reprocessed with every message. Share only the relevant sections of a document or dataset.
- Switch off extras you don’t need. Web search, deep research and extended reasoning modes all add compute, so use them when the task calls for them. Limit image and video generation too, as these are generally much more energy-intensive than text (and, in my opinion, often of low value).
- Refine rather than regenerate, and reuse what works. Edit small errors yourself, and keep good prompts and verified answers for next time.
- Get familiar with the costs, and favour models that openly publish them. Some providers are starting to publish figures, for example Mistral’s contribution to a global environmental standard for AI. You can also try the EcoLogits calculator to estimate the footprint of LLM use.
As an aside, in the final part of Hidden Figures, the IBM computer did the critical calculation, but it was checked by a “human computer” before it was used, a step that is definitely vital for LLM output!
Beyond LLMs: AI as a research tool
Of course, AI in research is much wider than just LLMs. While LLMs may be a part of many people's workflow, AI is a research tool in itself, and one with great potential for new discoveries. One of the most famous examples of this is the AlphaFold AI system which transformed structural biology and contributed to the 2024 Nobel Prize in Chemistry. And this isn't an isolated case: AI tools hold huge potential in medicine, environmental science, astronomy and many other fields.
Considering the environmental impact of AI as a research tool is, however, tricky, and to be honest, I wonder whether we need the AI label at all, or whether we should simply consider the environmental impact of all digital research (spoiler: we definitely should).
AI can both reduce and increase the environmental cost of research
In some fields, AI models are replacing more computationally expensive traditional ones, e.g. in weather and climate research GraphCast and ECMWF’s AIFS models learn from decades of reanalysis data to produce forecasts far faster (and with less energy) than traditional numerical models. Moving experiments from the laboratory to digital systems can also reduce their environmental impact, though we seldom track the cost of either, so it is hard to quantify.
On the other hand, AI is an accelerant. It lowers the barrier to computational research, which is good news for access, but more people can now write and run code. It also makes it easy to throw large volumes of data at a model to see what happens, without the planning that used to come from code being slow to write and slow to run. Each run feels cheap to the person pressing the button, while the energy and water sit elsewhere, out of sight.
Published data on any of this is scarce, which makes good practice harder to pin down. There are still a few steps we can all take.
- Ask whether the task needs AI. A simpler, well-established method may do the job with a fraction of the compute.
- Reuse pre-trained models where you can. This avoids the cost of training another from scratch.
- Test on a small sample before a large run. A quick pilot catches mistakes before they have consumed hours of compute.
- Know what is running, and for how long. Set time limits, and switch off anything you have forgotten about.
- Measure and share. Estimate the footprint of your work and report it alongside your results. What we don’t measure we can’t improve, and published figures help the next researcher make better choices.
Some tools can help with estimation. The Green Software Foundation’s Software Carbon Intensity for AI gives a way to help make the environmental impact of AI more transparent, the Green Algorithms calculator estimates the footprint of computation before you run, and the EcoLogits calculator can be used to estimate the footprint of LLM use. The GREENER framework offers broad principles for environmentally sustainable computational science, and applies just as well to AI.
We need open emissions data and a common reporting framework
The only thing I’m sure of is that the AI landscape will look very different in just a few years. In the 1960s, few would have imagined the extent to which computers would be used in almost every job. This blog will probably date quickly, but what will stay relevant is the need to ask what our choices cost. We can use these new tools blindly, or we can be aware of the costs: ask companies to provide environmental data, use tools to estimate emissions (and publish the results), and make careful choices. We can all play a part in normalising this environmentally aware behaviour, so that we benefit from these new tools while minimising the environmental cost.