Impact of Large Language Models not fully realised

Less than one in six people in the UK (16%) are aware of the environmental impacts of the likes of ChatGPT, with some even inputting simple queries such as ‘what is one plus one’, ‘what day is it’, and ‘what is the weather’, new research from the Institution of Engineering and Technology (IET) shows.

As ChatGPT celebrates its second birthday, and avid AI users are on the up, the new study found half of the UK public surveyed (49%) report to have casually experimented with the technology, with two in five (39%) using it at work, amid concerns from experts that misuse of the technology will have a vast environmental impact.

On average, we now use it once a week for tasks including writing emails (50%), creating presentations (48%) and project planning (48%) – as well as general home tasks such as writing shopping lists (46%). Other uses of the platform reported include writing card messages, asking how to boil an egg, what’s the meaning of life, and to tell a joke.

Almost half (46%) believe the technology will make their lives easier, with 50% seeing themselves using it more in the future. But with increased usage comes a greater environmental toll – to train ChatGPT 3 alone is estimated to have used 1,300 MWh of electricity, which is the equivalent of 1,625,000 hours of watching Netflix.

In terms of cooling the servers for AI to work, it requires significant water to keep systems at optimal temperatures, with analysis estimating that ChatGPT consumes just over one 500ml bottle of water per 100-word request. On asking the public, IET stats reveal that nearly one third of us are oblivious to AI’s water use. Just think – if the population of London at over 9.7 million residents each asked ChatGPT to write a 100-word email this would total 4,874,000 litres – equivalent to filling over seven 25m swimming pools.

Positively, 60% of respondents would reduce their usage if they became aware of the significant energy usage to run a request.    

Professor Peter Bannister, Chair of the IET’s Sustainability and Net Zero Policy Centre, said: “We’ve seen a colossal uptake of AI tools in the past two years, and while LLMs can provide great support to users, there is significant cost and environmental impact, which is only going to increase as the adoption of AI rises to drive efficiency and productivity.

“For the public, there are ways to reduce the environmental impact straight away, such as asking if you really need to use these platforms – can you find this information another way? If you do use a LLM tool, think about the prompts or queries you are inputting – provide as much detail as possible to reduce inaccuracies or needing to redo the request. It’s good to remember that LLMs aren’t search engines, they’re intended to help with writing documents and large pieces of text.

“We’re aware of what we need to recycle or to switch lights off when they’re not in use, but this is the next phase of consumer awareness – choosing carefully what technology you use and how you use it more efficiently.”

The IET’s best practice tips for using LLMs, like ChatGPT, more efficiently and effectively include:

• LLMs are most effectively used for producing text so use them to assist with writing documents.
• Consider the searches you are inputting – can you find the information via a traditional search engine, which is far less intensive on the environment.
• If you need to use LLMs, think about the prompts you are inputting for your request – provide as much detail as possible into it, so that you get as much out of it. The result will only be as good as the information you initially provide.
• Consider the environmental costs of LLM searches – it’s reported that each ChatGPT query consumes an estimated 2.9 watt-hours of electricity, nearly ten times more than a standard Google search.

As world leaders move to regulate the technology, considerations need to be made to mitigate the high energy requirements and usage of its data centres and network infrastructure. To tackle this, the IET believes regulations should now look beyond the immediate risks of AI development to the much broader impact it has on the environment.

Dr Graham Herries, Chair of the IET’s Digital Futures Policy Centre added: “This could include a bronze, silver, gold standard for the approval of new data centres – which support AI servers – in the UK, based on a sustainability rating.

“Current data centres are unsustainable without greener infrastructure, so we need to see government support for green technologies by building sustainability into appropriate regulation of new and expanding technologies.

“Training new AI models is also very costly and inefficient with each new model requiring training from scratch. We need to see alternative methods for new AI models to learn more effectively and encourage greater re-use of this learning. This isn’t an easy task, but we need sustainable innovation to be the future.

“By focusing on carbon efficient AI and technologies, the UK has the potential to become a leader in the global AI economy.”