People who know me or have read my blog posts before likely know that I’m a bit of an environmentalist. I litter pick in my spare time and recycle anything I can. I’ve been interested in digital sustainability for a while now too. However, some of you might be thinking… then why do you spend so much of your time going on about AI which is not good for the environment at all. Well, I genuinely think AI can be used for good (as well as bad) and I think it’s not going anywhere, so I’m pretty conflicted about it all. I think it’ll be a core part of many people’s lives going forward and so knowing about it and being able to use it will be a necessity but I’m also scared about the upscaling of data centres and the use of power and water to run them.
I’ve been learning a lot about AI recently. Although some models are getting smaller (and that means more efficient to run) and I think we’ll find those sort of efficiencies will come with time, there are some things we as users can do to make our own use of AI more efficient/environmentally friendly – just as there is with IT in general. I’ve written blog posts about keeping files small when sharing to reduce the amount of cloud storage they need or bandwidth used to share, for example, or considering if your audience really needs that information passed to them in video format when a simple text version would work just as well. It’s the same for AI, there are things we can do ourselves to limit our impact. Here is a quick summary of my findings.
1. Consider your options
Do you really need to use AI to solve a problem or complete a task? If you do, consider which AI model to use – do you really need the largest or would a smaller, greener model do? AI models or machine learning tools designed for specific tasks are generally more efficient – for example, if you have a structured dataset that you want to analyse or classify or use for predictions. This is particularly important if you have a large dataset or a lot of analysis to do. For example, if you were doing object recognition in an image database, it would be much quicker/cheaper and greener to use something like YOLO than an LLM to do it. The LLM could do the job but would be vastly more expensive both in terms of money and impact on the environment. (YOLO – You Only Look Once – is an open source object classifier). Something like YOLO is harder to get going with – so there’s a bit of extra effort required.
One thing to consider is that many search engines now give you an AI response by default as a result for your search. Do you use it? If not, switch it off. If it’s not possible to do this in your search engine (yes, Google, I’m looking at you…) considering switching search engine. A quick search for green search engines show’s the choice of search engines available (like Ecosia for example). I use DuckDuckGo and have switched the AI search results off.
2. If you are using an LLM, be as efficient as you can
Use greener/smaller models where you have the option to change model.
But also think about how you are using the model – set up projects with sources so you upload files once for the LLM rather than many times. Use the same chat conversation for similarly themed problems so you can make use of the model’s memory. For any LLM which allows you to give it rules or guidance – do that so it’s more likely to do what you want first time. For anything bigger such as image generation or document creation, get it to summarise what it’ll do or provide the prompt first so you can check it looks right/makes sense and then get it to go ahead when you are happy with it, rather than generating 10 versions of an image that’s always just a little bit wrong. Be clear in your prompts – so give your prompts a bit of thought before you add them and save yourself time as well as making your use more efficient.
3. Keep an eye on your agents
AI agents can work autonomously without you supervising so it’s critical you keep an eye on what they are doing – for security as well as environmental purposes. Make sure you know what they are doing and that it is sensible, safe and required.