The Senate Select Committee Report on Adopting Artificial Intelligence
When Australia’s Senate Select Committee on Adopting Artificial Intelligence published its final 222-page report raising 13 key recommendations on the domestic adoption of AI, a lot of attention was focused on the definition and treatment of “high-risk use” of AI. The committee went as far as to ask the government to include general-purpose AI models, such as large language models (LLMs), in a list of “high-risk AI uses”. This could, if adopted, lead to further legislative protections – something the tech giants are not keen on.
However, for data centre operators, the report also presented the “significant” environmental impacts of AI in relation to energy use; greenhouse gas emissions; water use; and land and resources. The evidence was a sobering reminder the challenge data centres will face given the concerns around their high energy use needed to deliver AI.
The Department of Industry, Science and Resources (DISR) noted: “AI is inextricably linked with data, which is the building block that powers machine learning and large language models. Training and using AI systems depends on massive amounts of computational resourcing, physical hardware and infrastructure. This means AI can be responsible for consuming large amounts of energy.”
DISR added that data centres currently represent 1-1.5% of electricity use globally…with estimates suggesting a single data centre may consume energy equivalent to heating 50,000 homes for a year. In Australia, data centres could currently account for around 5 per cent of energy use, with some projections suggesting this could grow to between eight and 15 percent by 2030. Ireland is an example of how things could develop. Sean Sullivan, deputy secretary, Department of Climate Change, Energy, the Environment and Water, cited the example of International Energy Agency estimates that data centres accounted for approximately 13 to 14 per cent of Ireland’s total electricity use in 2020, and was projected to grow to 40 to 50 percent by 2030.
Australia’s Chief Scientist, Dr Catherine Foley, commented on the large amount of energy needed to train generative AI models: “…[training] a model like GPT-3…[is estimated] to use about 1.5 thousand megawatt hours…[which is] the equivalent of watching about 1.5 million hours of Netflix.”
Generated outputs requires energy
The energy used by AI applications, particularly for generative AI, is clearly significantly greater relative to other technologies. Dr Ascelin Gordon, senior lecturer at RMIT University, noted estimates that a single ChatGPT query generating text could use between ten and 90 times as much energy to process as a simple Google search, with a query generating an image being probably 20 times more energy intensive. The energy intensity of generating video was likely to be ‘orders of magnitude higher’ again.
US-based Australian expert, Dr Kate Crawford cited estimates that ChatGPT alone uses the energy equivalent of 33,000 US households per day, with ‘future generative AI models potentially using the energy equivalent of entire nation-states. She added that recently OpenAI CEO Sam Altman admitted that the AI economy is heading for an energy crisis…[warning] that the next wave of generative AI systems will consume vastly more power than expected, and that energy systems will struggle to cope.
Green house gas (GHG) emissions and water use painted a similar challenging picture. The DISR submission cited estimates that the share of GHG emissions specifically from the operation of data centres is currently 0.6 per cent of annual global GHG emissions, while Science and Technology Australia put this figure at one percent, potentially increasing to 14 percent of annual global GHG emissions by 2040.
In terms of the water use attributable to AI applications, the Monash University submission cited estimates that in Europe a single ChatGPT-3 query uses a tablespoon of water. The submission of the ARC Centre cited a slightly higher estimate in Australia of every 26 ChatGPT-3 queries using approximately 500ml of water (roughly four times higher). In global terms, Dr Kate Crawford cited studies suggesting that annual AI demand for water could be half that of the United Kingdom by 2027. Given the occurrence and frequency of drought in Australia, this will pose key challenges fordata centre operators and regulators alike.
Given the absence of consistent and widely applicable standards, a number of witnesses and submitters indicated their support for the development of standards to more effectively and comprehensively measure and report the environmental impacts of AI. The UNSW AI Institute, for example, recommended that the government ‘support the development of standards for measuring the full environmental cost of AI’ along with committing to best practice “for AI projects developed in the public sector”.
Reducing impact
Dr Kate Crawford observed that designing data centres to be more energy efficient would contribute to reducing their energy use and improving the sustainability of such facilities more generally. She observed that government regulation may be needed to achieve such efficiencies across the wider industry: “At the outset…’government] could set benchmarks for energy and water use, incentivise the adoption of renewable energy and mandate comprehensive environmental reporting and impact assessments. Over time, laws and regulations could require adherence to strict environmental approaches that prioritise sustainability, especially for energy and water usage.”
Australia has made some good progress here withs NABERS and complementary work by the Uptime Institute. In addition to increasing the efficiency of the facilities and buildings that house computing and data centres, a number of inquiry participants noted the benefits of improving the efficiency of the computing methods used to develop, train and deploy AI models and systems.
However, the final report noted that increased energy efficiency of AI may not lead to reductions in AI’s total energy use, where there is increased demand overall for AI services. Research conducted by Goldman Sachs, for example, found that between 2015 and 2019 the energy demands of data centres remained relatively stable despite a tripling of their workload, partly due to gains in energy efficiency, but since 2020 the benefits of these efficiency gains ‘appear to have dwindled’. It concluded that ‘the widening use of AI’ implies an increase in AI’s energy consumption overall notwithstanding improvements to the efficiency of AI data centres and systems.
Dr Crawford noted the potential for increasing the energy efficiency of developing AI models as well as designing AI systems to operate using less energy. The BigScience project in France, for example, had developed the BLOOM AI model that is a similar size to OpenAI’s ChatGPT-3 but has a significantly lower carbon footprint.
More regulation?
Many industry submissions provided large amounts of detail around how each company was managing and mitigating the environmental impacts of AI. AI technologies can also be applied to a wide variety of environmentally positive uses. Such uses include the use of AI to avoid, reduce or mitigate the negative environmental impacts of human industry and economic activity generally, and more directly to further the understanding and management of specific environmental challenges such as climate change and species extinction.
However, while the committee appreciates that greater efficiencies in the development and operation of AI models may provide corresponding reductions in energy and water use, it notes that the pursuit of computing efficiency by large AI companies is driven by commercial rather than environmental imperatives.
In the context of the continuing rapid growth of the AI industry, the committee considers it very likely that any related environmental gains are therefore likely to be insignificant, amounting to merely slower rates of growth in energy and water use overall.
The Committee’s conclusion hinted that further regulation may be on the way. It noted that by 2023, the world’s data centres consumed more energy than India, the world’s most populous nation, driven primarily by a massive extension of AI infrastructure. “This, in addition to the significant water use, land use and other environmental concerns associated with this infrastructure, necessitates a coordinated and holistic government approach to ensuring the growth of this sector in Australia is sustainable,” opined the Committee.
[Author: Simon Dux]
