AI has become an ubiquitous part of our lives, whether we explicitly engage with it or not. The websites we visit, the apps we use, and even the chatbots we communicate with to locate undelivered parcels are all utilising or beginning to utilise AI and its chatbot capabilities. We may ask a chatbot to suggest ideas and outlines for a conference, to write simple media posts, or to discover a quick and simple answer to a question we’re currently pondering. Being aware of the limitations – of false news, biases, and often a lack of appropriate referencing – we can intelligently work with AI in a responsible manner. But what of the environmental costs?
Emerging research indicates vast amounts of energy and water are being consumed by the data centres powering these large language models. However, discussions around these concerns are often obscured by the ongoing promotion of AI as a progressive too – especially when related to human ingenuity and originality. It seems an opportune time to ask one of these language models, in this case Bing Copilot, powered by GPT-4, about its own perceived environmental impact.
However, in recognition that a singular query contributes approximately 4.32g of CO2, compared to Google queries which generate approximately 0.2g of CO2, these questions were kept to a fair minimum.
Q & A with Bing Copilot – a dialogue about energy consumption
Hi Bing, can you tell me how much energy you consume?
As an AI, I don’t consume energy in the way humans do. However, the servers that host me and other digital services do use energy. The exact energy consumption would depend on many factors, including the energy efficiency of the data centres and the methods used to generate the electricity they use.
It’s been reported that the training of language models requires a large amount of energy consumption. How much energy does it take to train a language model like ChatGPT and others powered by OpenAI ?
Training ChatGPT requires a lot of computational power and electricity. According to a study by researchers at the University of Washington, training a single large language model like ChatGPT-3 can consume up to 10 gigawatt-hour (GWh) of power. This is on average roughly equivalent to the yearly electricity consumption of over 1,000 U.S. households. The study also estimated that the carbon footprint of training ChatGPT-3 is between 55 and 284 tons of CO2, depending on the source of electricity.
How much energy does it take to run a model like Bing Copilot or ChatGPT?
Running Bing Copilot or ChatGPT also consumes energy, but not as much as training it. The exact amount of energy depends on the size of the model, the number of tokens processed, and the efficiency of the hardware and software. One estimate by TechHQ suggested that the energy consumption of a ChatGPT-4 query is between 0.001 and 0.01 kWh (3.6 and 36 kJ), depending on the model size and number of tokens processed. That means a single ChatGPT query consumes 1,567%, or 15 times more energy than a Google search query, which is estimated to be 0.0003 kWh (1.08 kJ)2.
What are the environmental impacts of using Bing Copilot and other AI systems?
The environmental impacts of using ChatGPT and other AI systems are not trivial. As the demand for AI services increases, so does the need for more data centres, which are facilities that house servers and other equipment that store and process data. Data centres consume a lot of energy, not only for running the hardware, but also for cooling and power management. Data centres are responsible for about 1 – 1.5% of the global electricity consumption and 0.3% of the global CO2 emissions. Moreover, data centres use a lot of water, both for cooling and for generating electricity from hydropower plants. According to a report by Livemint, ChatGPT-3 alone consumes about 800,000 litres of water per hour, which is enough to meet the daily water needs of 40,000 people.
How do you think we can reduce the energy consumption and environmental impact of Bing Copilot, ChatGPT and other AI systems?
There are several ways to reduce the energy consumption and environmental impact of Bing Copilot, ChatGPT and other AI systems. One way is to improve the efficiency and design of the hardware and software that run the AI models. For example, using liquid immersion cooling instead of air cooling can lower the hardware heat and reduce the carbon emissions and water usage of data centres. Another way is to use renewable energy sources, such as wind, solar, and hydro, to power the data centres. Some countries, such as Norway and Iceland, have abundant natural resources that make their electricity cheap and green. Many large organisations have built their data centres in these countries to take advantage of their low-carbon energy. A third way is to limit the use of AI models to essential and meaningful applications, and avoid using them for trivial or harmful purposes. For example, using ChatGPT to generate fake news or spam messages can waste energy and cause social problems. On the other hand, using ChatGPT to create educational or artistic content can have positive impacts on society and culture.
So what do you think about the future of energy consumption in AI?
The future of energy consumption in AI looks promising. With advancements in technology, we can expect more energy-efficient AI models and data centres. The use of renewable energy sources is also expected to increase. However, as AI becomes more prevalent, it’s important to continue focusing on reducing energy consumption and moving towards more sustainable practices.
A lack of accurate ChatGPTs figures
It is perhaps telling that whilst Bing Copilot does state or reference certain figures, these figures are preceded by language that indicates a lack of definitive data. Indeed, existing and emerging research on the environmental impacts of AI chat models are all based on estimated data sets and projections. Unlike other industries, the true energy inputs and outputs of AI processes have not been publicly disclosed. This includes both their carbon emissions and water footprints.
Despite not having the full details, research can still provide us with an indication of what we might expect in terms of environmental impact. A study by Li et al (2023) has observed that global AI demand could be accountable for 4.2 – 6.6 billion cubic metres of water being withdrawn in the year 2027 – more than the total annual water withdrawal of half the United Kingdom. They note that issues regarding water consumption and wider potential environmental impacts of AI are often overlooked, and that this isn’t helped by the lack of data provided by developers. In other studies, it’s also been estimated that by 2027, worldwide AI electricity consumption may compare to the annual electricity consumption of Argentina and Sweden.
A need for greater accountability and transparency?
Greater transparency is required, related to both the operational and developmental emissions resulting from AI processes. This would include the disclosure of water efficiency data and comparisons of different energy inputs. A data centre powered by coal is likely to be less energy efficient than one powered by solar or wind energy. But there has to be open data in order to make these comparisons and to assess the true environmental impacts of these language models – and the differences between them. An interplay of economic, social and environmental factors needs to come under close and careful assessment. Community engagement, local knowledge and individual understanding could play a key part in influencing and persuading developers to share this data. Being aware that there are potential environmental impacts, alongside the more widely discussed ethical concerns of AI, could make individual calls for a more accountable and responsible AI industry just that bit stronger.
More on AI
