A greener future: Why AI development must prioritise sustainability
5 minute read
As AI is more widely adopted globally, what are we learning about its impact on the planet? And what can developers do to make their AI more sustainable?
Recently I virtually attended the Global AI conference. I heard lots of interesting talks on the newest and most exciting tech innovations across a whole variety of services and sectors. What stood out to me most were talks centred around sustainability, and the role Artificial Intelligence (AI) could play in reducing digital carbon footprints.
In particular, I'd highly recommend checking out a talk on green AI by Yelle Lieder as a standout session from the day. Alongside my own recent learnings about the subject, it inspired me to bring all these ideas together into a handy “How to make AI more sustainable” guide.
But... why care?
A big question though is, why care? Why is it so important that AI is used responsibly and sustainably?
AI leaves a huge carbon footprint
For starters, training and using AI requires vast amounts of computation power, contributing to a high carbon footprint. Analysis from OpenAI researchers has shown that since 2012, AI’s computational power requirements have doubled every 3.4 months. And this number is only getting larger; the carbon footprint of the latest AI models is only increasing. A recent paper by the University of Massachusetts assessed the training process for several of the most prominent large AI models. They found that the process can emit more than 626,000 pounds of CO2— that’s almost nearly five times the lifetime emissions of the average car. Training GPT-4 alone had a carbon footprint equivalent to powering more than 1300 homes for an entire year.
Why does AI produce so much carbon?
AI is usually hosted in data centres. These data centres produce incredible amounts of carbon and currently contribute around 0.6% of all greenhouse gasses . Gerry McGovern has written both about the environmental impact of AI, and the overall impact of our data consumption on the planet. Gerry also joined us at Camp Digital 2022, where he gave a fantastic talk on the subject which I'd highly recommend checking out.
Steps to regulation
Organisations should also be mindful of existing regulatory structures, and the likelihood more will be introduced very soon. While there are no green AI-specific regulations published by the UK government yet, in their 93-page white paper ‘A pro-innovation approach to AI regulation’ it is argued that, “Human rights and environmental sustainability are not explicitly named in the revised principles as we expect regulators to adhere to existing law when implementing the principles.”.
The government also offers internal guidance for the public sector on making their technology sustainable, which emphasises that orgs should focus on reducing carbon emissions, and hardware and software waste. It’s also likely that with an increasing focus on sustainability and ‘anti-greenwashing’ corporate responsibility by the UK government in 2024, we may see more specific guidance on AI appearing soon.
Aligning with the UK’s commitment to achieve net-zero emissions by 2050, the EU also has new legislation coming into place which includes the Corporate Sustainability Due Diligence Directive. This directive puts sustainability first and means companies will have to publicly share their sustainability efforts and carbon footprint. The directive is also amongst this new legislation which mandates more detailed and transparent reporting of sustainability efforts to “increase a company’s accountability, prevent divergent sustainability standards, and ease the transition to a sustainable economy.”
Customer demands for sustainability
Customers are more aware than ever of how their choices affect the planet. Brands who make a commitment to addressing the problem can build stronger trust within the market.
Gordan Ellis-Brown recently wrote a blog post about navigating the environmental, social, and corporate governance landscape where he highlights how having good sustainability practices can make or break a brand. Notable statistics include “84% of customers say that poor environmental practices will alienate them from a brand or company” and “79% of people who hear brands communicating about sustainability are likely to trust that messaging”.
And most important of all, there is an ethical responsibility for anyone involved in AI to mitigate the environmental impact of such tools, aligning with broader goals of reducing climate change and promoting a sustainable future.
Suggestions to reduce your carbon output
There are many different things we can do, some easier than others but all with the same goal.
- With data being the key to all this, storage should be a high priority. Opting to store your data in cold storage reduces the energy needed. Another important storage consideration is that 73% of data stored currently never gets used again, so adding a ‘Time to Live’ on data can free up space and reduce the amount of energy needed.
- With AI, thousands of pre-trained models are only a Google search away. It's important to ask “Do I need to train a new model?” because more than likely there is already a model ready to use, which you just need to fine-tune to suit your purposes.
- Using hardware-optimised solutions is also a key factor in how much energy is consumed, so using a data centre service which is optimised for machine learning and huge data processing is an effective way to reduce unnecessary energy consumption.
- Time shifting is the concept of only using energy when it is cheap and green, outside of peak hours. This can mean highly-intensity operations don’t have such high carbon figures as they’re powered completely renewably.
- When developing a model, consider how often you need to retrain it and if it's necessary at all. Is there a big enough benefit?
- Implementing caching on models. This means that if the AI has answered a question before or created an image with the same prompt before, it doesn’t need to create something new, just give the same answer that has been saved before.
Measurability and benchmarking
It’s difficult to understand how good or bad using AI can be without some form of benchmarking to show consumption and create a comparison that we as consumers can understand. That’s where emissions calculators come into play. By using a machine learning emissions calculator like this one you can estimate with your given hardware how much carbon in kg CO2 you’re going to produce. The calculator also puts the emissions into somewhat of an understandable metric, like how far you’d have to drive an average car to produce the same emissions.
To summarise, AI is a brilliant tool that I believe will help us create a brighter future, and I'm excited to see its potential. But in the course of these new developments, we need to be consistently mindful of the technology’s impact on people and the planet, to enhance and not compromise that future.
There’s no point in AI if no one's there to use it, right?