Chapter 6Impacts of AI on the environment
1.1The environmental impacts of artificial intelligence (AI) are significant and arise across the AI lifecycle, from the development and training of AI models; the deployment of AI systems for various uses in industry, business and society; and the building, decommissioning and renewal of the Information Technology (IT) infrastructure and equipment that support and comprise AI technology.
1.2This chapter sets out the evidence received by the inquiry and the committee’s views regarding:
the environmental impacts of AI in relation to energy use; greenhouse gas emissions; water use; and land and resources;
capturing, reducing and regulating the environmental impacts of AI; and
environmentally positive uses of AI.
Energy use
Data centres
1.3A number of submitters and witnesses raised concerns about the high energy use of the infrastructure that is needed to develop and operate AI technologies.
1.4In particular, the training and refining of AI models requires the running of high-performance computers and significant data storage in dedicated facilities. 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...
1.5Australia’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 1½ thousand megawatt hours…[which is] the equivalent of watching about 1½ million hours of Netflix.
1.6While acknowledging the difficulty of quantifying the global energy use of AI, DISR noted the significant global impact of data centres:
…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.
1.7Mr 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 50percent by 2030.
1.8In Australia, data centres could currently account for around 5 per cent of energy use, with some projections suggesting this could grow to between eightand 15 per cent by 2030.
1.9Commenting on predictions of ‘considerable potential growth in [energy] consumption from…data centres’, Dr Dylan McConnell, who appeared before the committee in a private capacity, questioned whether Australia’s current energy planning processes were adequately taking such predictions into account:
…going off these projections, the current planning and preparedness of the grid is insufficient...[as we] haven't been taking into account these extreme levels of growth.
1.10Mr Sullivan noted that the Australian Energy Market Operator (AEMO) is conducting a consultative process on future AI data centre energy requirements, considering upper limit and baseline cases:
Demand management is part of that future and has been built into a lot of the models. I take the evidence that AEMO is trying to improve its forecasting models with respect to the use of data centres…AEMO does that on a regular basis with respect to trying to improve both its Integrated System Plan for transmission and its energy demand models.
AI-generated outputs
1.11In addition to the energy used by high-performance computing facilities and data centres to train and refine AI models, the use of AI applications to generate content also uses significant amounts of energy. While the energy used to generate a response to a single AI query may be relatively small, the total energy used to respond to millions of AI queries a day may, on some estimates, be greater than the energy used for training and refining AI models.
1.12Google noted that the energy use of an AI application relative to other digital services such as a simple Google search depends on a number of factors:
Energy use of all…digital services varies at the time of service based on the complexity of the computation required for a given query or query chain, and over time based on the efficiency of algorithms and chips used to enable these services…[and] changes in user behaviour over time…
1.13However, 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.
1.14Dr 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’. Dr Crawford observed:
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.
1.15The ARC Centre of Excellence for Automated Decision-Making and Society (ARC centre) submission observed that the energy demands of AI would likely lead to competition for energy between AI services, businesses and public requirements.
Greenhouse gas emissions
1.16In addition to the challenges that it represents globally for energy consumption and future planning of energy systems, AI energy use derived from fossil fuel sources contributes to global greenhouse gas (GHG) emissions, and therefore to the anthropogenic warming of the planet and accelerating climate change.
1.17Some estimates suggest that the share of global GHG emissions from the entire information and communications sector is currently around 1.4 per cent, increasing to around 14 per cent by 2040.
1.18The 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. Google provided a lower estimate of the GHG emissions of ‘cloud and hyperscale data centres’, at around 0.1 to 0.2 per cent of global GHG emissions ‘based on the most recent global estimates as of 2022’.
1.19Google noted that its data centres comprise a ‘large part’ of its energy use and thus its GHG emissions, acknowledging that AI’s ‘intensive computation method’ had led to a significant increase in its GHG emissions in recent years.
1.20Similarly, Microsoft acknowledged that its AI models and services had led to increases in its energy use and therefore GHG emissions, including a 30per cent increase in its GHG emissions since 2020 due largely to its data centres.
Water use
1.21Another significant environmental impact of AI technology is due to the significant amounts of water used by high performance computing facilities and data centres for cooling the energy-intensive graphics processing units (GPUs) that provide the massive computing power required by AI. In some cases, water might also be used in connection with on-site generation of the electricity needed to power the facilities.
1.22In 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.
1.23The ARC Centre observed that the increasing speed of development and frequency of use of AI are likely to ‘significantly heighten the demand for and use of…water for computing for the purposes of AI’. As with energy, the increasing water demands of AI could lead to competition between AI services, businesses and public requirements in places where water resources are scarce.
1.24In this regard, Dr Kate Crawford noted that Google and Microsoft had reported ‘spikes in their water usage during the deployment of their chatbots’, leading to populations near data centres expressing concern about impacts on residential water supplies:
For example, in West Des Moines, Iowa, a giant AI hyperscale data centre was built to serve OpenAI’s most advanced model, GPT-4. During peak times in summer, the custom-built facility of over 10,000 GPUs draws about 6% of all the water used in the district, which also supplies drinking water to the city’s residents.
1.25Dr Crawford observed that, given the occurrence and frequency of drought in Australia, it is essential that Australia’s AI policy approach accounts for risks associated with its water usage.
1.26Mr Mark Stickells AM, Chief Executive Officer, Pawsey Supercomputing Research Centre, provided an example of a sustainable approach to water use by the Pawsey supercomputing centre. Mr Stickells noted that the centre had been designed with reference to principles of environmental sustainability, including in relation to water use. The centre uses water drawn from an aquifer for cooling the computing systems, which is then recharged back to the aquifer, amounting to the recycling of several million litres of water per year.
Impacts on land use and resources
1.27As set out above, AI has significant environmental impacts that arise directly from the energy use (and associated GHG emissions) and water required by the high-performance computing facilities and data centres that support the development of AI models and the and operations of AI applications.
1.28In addition to these impacts, inquiry participants identified a range of other environmental impacts on land use and resources arising from the associated processes and industries that are critical not just to AI technology but also to the technology industry more broadly.
1.29The ARC Centre submission noted that, in addition to the land requirements for high-performance computing facilities and data centres, land is also required for the mining and processing of the key resources needed by the technology industry; the manufacturing plants that produce the computing and other equipment for developing and platforming digital technology applications; and the infrastructure, such as communications equipment and undersea cables, that provide the connectivity for digital technologies.
1.30The ARC Centre also noted that the AI industry would be an ‘increasing contributor to Australia’s existing waste and recycling challenge’. It observed:
The underlying logic of AI uptake also almost necessarily involves increasing proliferation and regular upgrading and replacement (and hence production and waste) of hardware from graphics processing units in data centres to the proliferation of business, home and personal mobile devices that incorporate AI applications.
1.31Dr Kate Crawford highlighted the mining of critical minerals as a particularly concerning aspect of AI’s environmental impacts. In addition to contributing to land use and waste management pressures, the mining of critical minerals such as lithium has direct environmental impacts, including on habitat and species.
Capturing, reducing and regulating the environmental impacts of AI
Capturing the environmental impacts of AI
1.32A number of inquiry participants pointed to the need to more fully capture the impacts of AI on the environment, including through greater transparency and the development of standards to support better measurement and reporting of AI’s environmental impacts.
1.33The UNSW AI Institute submission noted that the impacts of AI are currently difficult to quantify due to there being ‘few standards for reporting’.
1.34The submission of Dr Kate Crawford observed that the available data on AI’s environmental impacts is incomplete, and that it is ‘very hard to get accurate and complete data on [its] environmental impacts’. Dr Crawford ascribed the paucity of data in part to commercial sensitivities, with the ‘full planetary costs of generative AI’, for example, being ‘closely guarded corporate secrets’. This view was echoed by Dr Dylan McConnell, who commented that the publicly available information in relation to the energy consumption, operating profiles and expansion plans of data centres is ‘somewhat opaque’ due to ‘commercial sensitivity’.
1.35Dr Crawford suggested that a multifaceted approach involving the AI industry, researchers and legislators is required to better capture the environmental impacts of AI, suggesting in particular the need for measuring and public reporting of energy and water use by the AI industry as well as ‘regular environmental audits by independent bodies’ to ‘support transparency and adherence to standards’.
1.36Given 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’. The Computing Research and Education Association submission also called for the development of ‘tools for measuring the full environmental cost of AI’ and for government to ‘lead by example with AI projects developed in the public sector’.
1.37The Salesforce submission called for ‘standardised metrics for measuring and reporting the impact of AI systems’, along with requirements for the public disclosure of the energy efficiency and carbon footprint of the development and operation of AI systems.
1.38The ARC Centre pointed to ‘strong moves’ towards the development of standards for measuring and reporting in Europe and the US, calling on Australia to:
…encourage and promote if not mandate the development of environmental impact logging and transparency standards to support environmental reporting and transparency across the whole AI supply chain.
1.39The ARC Centre noted that the work of multiple stakeholders could contribute to the development of such environmental standards for AI:
To support such standards, multiple groups in academia, advocacy, software engineering and industry are working on the quantification of carbon impacts of AI in application software, as well as more holistic assessments of the environment impacts of AI systems.
1.40Dr Crawford cited efforts in the US to improve the understanding of AI’s environmental impacts, as well as establish standards and a voluntary framework for measuring and assessing those impacts, via the introduction of the Artificial Intelligence Environmental Impacts Act of 2024 (AIEI Act). The AIEI Act would:
…[require] the [US] Environmental Protection Agency to lead a study on the environmental impacts of AI…[and direct] the National Institute for Standards and Technology to collaborate with academia, industry and civil society to establish standards for assessing AI’s environmental impact, and to create a voluntary reporting framework for AI developers and operators.
1.41The ARC Centre submission noted that the AIEI Act framework is intended to measure and report the ‘full range of…[AI’s] environmental impacts including energy consumption and pollution across the full AI lifecycle’, and potentially be implemented as a mandatory rather than voluntary reporting framework.
1.42Further, the ARC Centre noted that an ‘earlier version’ of the EU Artificial Intelligence Act (EU AI Act) had included similar provisions to the AIEI Act for environmental auditing and reporting, and that these elements remained implicit in the approach taken by the EU AI Act. It noted that the EU AI Act ‘encourages the creation and implementation of voluntary codes of conduct for assessing and minimising environmental impact for all AI developers and providers’, while high-risk AI systems are expected to be subject to risk and technical assessments that address the impact of AI development, operation and deployment on ‘environmental protection’. The ARC Centre considered that Australia should also adopt an approach in which ‘environmental impact and sustainability…[are] defined as included within the broad concept of AI safety for the purposes of risk assessment and mitigation’.
1.43The Salesforce submission also supported including establishing efficiency standards for high-risk AI systems, calling for environmental impacts to be included in ‘assessing the risk of AI systems and classifying high-risk models’.
Reducing the environmental impacts of AI
Data centres
1.44Noting the significant environmental impacts of the infrastructure that supports the development and deployment of AI (discussed above), a number of inquiry participants noted the benefits of increasing the use of renewable energy in, and energy efficiency of, AI facilities and AI systems, as well as improving the computing efficiency of the high-performance computing facilities used to develop and deploy AI systems.
Increasing the use of renewable energy
1.45The submission of the UTS Faculty of Engineering and Information Technology, for example, called for government to increase investment in ‘mitigating the environmental impact of AI data centres’ given their significant consumption of energy. It noted increasing efforts over the last decade to use renewable energy sources for datacentres, thereby reducing their GHG emissions and leading to the concept of a ‘green data centre’.
Increasing energy efficiency
1.46Dr 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. Dr Crawford 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, incentivize 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 prioritize sustainability, especially for energy and water usage.
1.47In 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.
1.48However, it should be 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.
1.49The AI and Cyber Futures Institute recommended encouraging the adoption and creation of environmentally sustainable AI technologies by investing in research into and development of ‘green AI’.
1.50Dr 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. Dr Crawford further suggested that AI researchers could collaborate with social and environmental scientists to ‘optimize [AI] neural network architectures…[and] guide technical designs towards greater ecological sustainability’.
1.51The submission from the Pawsey Supercomputing Research Centre and Curtin Institute for Data Science noted a number of specific opportunities or strategies it employed to ‘limit, reduce and mitigate’ the environmental impacts of the Pawsey supercomputing facility, which could be applied to the ‘uptake of AI technologies throughout Australia’. These include:
working with the research community to ensure that ‘code is efficient’ and AI algorithms are optimised to reduce the computational resources they require. Improved computer code efficiency is achieved, for example, through ‘compressing and pruning AI models to reduce [their] size and computational complexity, transfer learning (leveraging pre-trained models and transfer learning techniques)…[and] supporting and advising’ on developing more energy efficient algorithms and techniques;
considering energy efficiency as a key factor in hardware selection and performing ongoing monitoring, such as conducting lifecycle assessments of AI technologies to identify opportunities for efficiency improvements throughout their lifecycle; and
investing in and developing training and education programs to inform AI developers, researchers and users about the environmental impacts of AI technologies and promote sustainable practices. An example of this is a current collaboration with the Commonwealth Scientific and Industrial Research Organisation (CSIRO) to provide researchers with a report on the energy consumption and GHG emissions of the work they undertake using the Pawsey supercomputing facility.
1.52To assist and ‘guide companies in minimising their ecological footprint’, the UTS Faculty of Engineering and Information Technology submission called for the establishment of ‘clear regulatory guidelines that mandate the environmental assessment and optimization of AI systems’.
AI industry
1.53The submissions and evidence received from some significant AI industry participants described a range of ways to reduce the environmental impacts of AI, particularly in relation to the issues of energy use, GHG emissions and water use described above.
1.54A number of these submitters pointed to their overarching environmental commitments to, for example, carbon neutral operations, as well as practical efforts to achieve more efficient computing and methods used for AI development and deployment.
1.55Google, for example, advised that, since 2017, it has matched 100 per cent of the energy use of its global operations with annual purchases of renewable energy and is pursuing a target to be carbon neutral by 2030. It also seeks to minimise its carbon footprint through optimal management of building temperature and lighting settings.
1.56Google also noted rapid improvements it has made in the efficiency of AI development and refinement, ‘reducing the energy required to train an AI model by up to 100 times and…associated emissions by up to 1000 times’. Faster and more efficient techniques for AI and machine learning were ‘enabling [AI] models that are ‘higher quality, faster, and less compute-intensive to serve’.
1.57Similarly, the submission of Amazon noted its ‘sustainability commitment’ to ‘match 100 percent electricity with renewable energy by 2030’ and reach ‘net-zero carbon’ by 2040. It noted the potential for significant energy efficiency gains, and therefore GHG emissions reductions, by running AI systems in cloud-based computing systems, and pointed to its efforts to improve the energy efficiency of its data centres as well as power them by renewable energy sources.
1.58Amazon is also working on the development of more efficient chips for use in developing and deploying AI systems, such as ‘high-performance machine learning’ chips ‘designed to reduce the time and cost of training generative AI models’ that could achieve energy-consumption reductions of up to 29 per cent.
1.59Microsoft considered that it remained ‘on track’ to achieve its commitment to be carbon neutral by 2030, despite increased emissions due in part to AI in recent years, and noted commitments to its operations being ‘water positive’ and ‘zero waste’ by 2030. It detailed a number of initiatives being undertaken as part of a ‘sustainability-by-design approach’, including optimising energy consumption efficiency in Microsoft datacentres through renewables projects and zero-carbon electricity; and the use of low-carbon building materials for constructing its datacentres.
1.60As with the other large AI technology firms, Microsoft indicated that it was exploring ways to develop and deploy AI models with fewer resources while achieving similar or better performance of current approaches—for example, through its release of ‘Phi’, a suite of small language models whose performance matches and outperforms models up to 25 times larger.
Regulating the environmental impacts of AI
Existing government regulation and policy
1.61As discussed below, regulation of the environmental impacts of AI falls under existing Commonwealth and state and territory legislative schemes for environmental protection and management.
1.62In addition, a number of other policies and initiatives developed in recent years are relevant to managing and mitigating the environmental impacts of AI. These include AI-specific policy frameworks and guides on the development and use of AI and general policies, such as procurement policies that could capture the purchase of AI systems and construction of AI facilities by government.
1.63The submission of the ARC Centre commented on the need to review broadly the effectiveness of Australian law in regulating the environmental impacts of AI:
There is…a need to review and update existing Australian laws and policies to ensure they are fit for the purpose of ensuring environmentally responsible and sustainable AI, as AI applications are taken up across the whole of the public and private sector. Different legal and policy frameworks will touch on AI’s environmental impact across different parts of the whole AI lifecycle including: environmental planning laws and licensing regimes for the siting and running of facilities such as mines and processing facilities for critical minerals, and data centres and undersea cables for data storage and compute power; energy grids including the creation and use of renewable energy facilities; carbon reporting and ESG frameworks; policies to incentivise and obligate product stewardship and e-waste reduction and re-use.
General scheme of environmental protection in Australia
1.64The environmental impacts of the AI industry in the broadest sense are, as with any business or industry activity, subject to Australian Commonwealth, state and territory environmental laws, regulations and polices.
1.65At the Commonwealth level, the main environmental legislation is the Environmental Protection and Biodiversity Conservation Act 1999 (EPBC Act), which is primarily concerned with the environmental assessment and regulation of development activities that could have a significant impact on matters of national environmental significance.
1.66Outside of this, the states and territories are responsible for a range of environmental laws, policies and administration, in relation to, for example, development assessments (other than those falling under the EPBC Act) and the management, use or protection of living and non-living resources.
AI Ethics Framework (2019)
1.67The AI Ethics Framework (the ethics framework) was released by the Department of Industry, Science and Resources (DISR) in November 2019 with the aim of guiding businesses and government to design, develop and implement AI responsibly.
1.68The ethics framework includes eight voluntary AI Ethics Principles, intended to:
achieve safer, more reliable and fairer outcomes for all Australians;
reduce the risk of negative impact on those affected by AI applications; and
help businesses and governments to practice the highest ethical standards when designing, developing and implementing AI.
1.69The first principle listed in the AI Ethics Principles relates to ‘human, societal and environmental wellbeing’, and states that ‘AI systems should benefit individuals, society and the environment’. This principle is explained as encouraging the assessment of AI’s environmental impacts throughout the lifecycle of AI system, as well as the use of AI to help address areas of global concern such as the United Nations’ sustainable development goals.
Net Zero Plan (2022)
1.70The Net Zero Plan, outlined in the government’s 2022 Annual Climate Statement to Parliament, is being developed to guide Australia’s transition to the legislated target of net zero GHG emissions by 2050. The submission of the Department of Industry, Science and Resources notes that the plan:
…will be supported by six sector plans that are being developed for electricity and energy, agriculture and land, infrastructure and transport, industry, resources, and the built environment sectors. The plans will consider ways to reduce emissions in each sector and between them the plans cover all major components of the economy.
1.71In terms of AI environmental impacts specifically, the submission advises that:
…data centres as a commercial building are captured under the Built Environment Sector Plan, which will provide an emissions reduction pathway to 2050 for the built environment sector.
National framework for the assurance of artificial intelligence in government (2024)
1.72The national framework for the assurance of artificial intelligence in government (the assurance framework) was agreed to by relevant Commonwealth, state and territory ministers in June 2024. The assurance framework draws on the 2019 ethics framework and is intended to establish ‘cornerstones and practices of AI assurance’ as part of the broader governance of the use of AI by governments.
1.73The assurance framework cornerstones describe key assurance practices to assist governments to ‘effectively apply’ the ethics principles underpinning the ethics framework. The cornerstones relate to AI governance and data governance; risk based assessment and management of AI; AI standards; and procurement.
1.74The assurance framework practices are intended to demonstrate how governments can practically apply the ethics principles to the assurance of AI. The practices specify the following three actions implementing the ethics principle of ensuring AI systems achieve ‘human, societal and environmental wellbeing’:
Document intentions: governments should define and document the purpose, objectives and expected outcomes of AI use cases for people, society and the environment; and consider whether there is a clear public benefit from the use of AI, and whether the use of AI is preferable compared to non-AI alternatives.
Consult with stakeholders: governments should identify and consult with stakeholders, including subject matter and legal experts and impacted groups, to allow for the early identification and mitigation of risks.
Assess impacts: governments should assess the likely impacts of an AI use case on people, society and the environment to determine if the benefits outweigh risks and manage any such impacts appropriately.
Environmentally Sustainable Procurement policy (2024)
1.75The Environmentally Sustainable Procurement policy (ESP policy) and associated reporting framework are intended to reduce the environmental impact of Australian Government procurement, and thereby support ‘Australia’s transition to a net zero, circular economy’, by preferencing the purchase of products that minimise GHG emissions, are safer for the environment and retain their value for longer.
1.76The ESP policy has applied from 1 July 2024 to construction services procurements at or above $7.5 million; and from 1 July 2025 will also apply to procurements for information and communications technology (ICT) goods; uniforms and textiles; and fit outs of builsing and office interiors (furniture, fittings and equipment) at or above $1 million. The Sustainable Procurement Guide, which informs the ESP policy, states that these procurement areas represent the government’s ‘highest impact purchases’.
Environmentally positive uses of AI
1.77As set out above, AI infrastructure and systems can have significant negative environmental impacts in terms of energy use; GHG emissions; and water, land and resource use.
1.78However, AI technologies can 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 our understanding and management of specific environmental challenges such as climate change and species extinction.
1.79A number of inquiry participants pointed to the increasing potential for use of AI by industry in ways that reduce the impact of industry on the environment. For example, the submission of the DISR cited the finding of a 2022 IBM Global AI Adoption Index that ‘two-thirds of companies either use or plan to use AI to pursue their [environmental] sustainability objectives’.
1.80Submissions cited numerous examples of potential uses of AI to improve the efficiency of industries in ways that have corresponding environmental benefits, usually through more efficient use of energy and resources thereby reducing environmental impacts. These examples include the use of AI to:
optimise manufacturing processes;
optimise retail supply chains by, for example, using AI to determine the most efficient packaging options and detect damaged goods before shipping;
operate mines with greater energy efficiency through, for example, automated ore sorting; water and power monitoring; supply chain monitoring; and environmental monitoring;
monitor agricultural crops to provide precise data on crop health and improve agricultural practices; as well as to automate management of water and pest control for improved environmental sustainability; and
enhance transport planning and traffic management to reduce transport GHG emissions through more fuel-efficient routing and optimised traffic flows.
1.81AI can also be used to discover and develop solutions to specific environmental impacts caused by industrial activities or processes. For example, DISR cited the use of AI to develop a new concrete formula that reduces the highly GHG emissions-intensive process of concrete manufacturing by 40 per cent.
1.82In addition to industrial applications, AI can significantly improve environmental and natural resource management. Examples of beneficial uses for these purposes include the use of AI to:
optimise the use of natural resources such as water and energy to reduce waste and environmental impact;
monitor and track wildlife populations and ecosystem health to aid in the preservation of biodiversity; and
monitor and predict environmental changes, such as air quality and deforestation rates, to facilitate and inform conservation efforts.
1.83AI is also able to assist with understanding and addressing the significant environmental challenges and impacts of climate change. Examples provided by a number of inquiry participants included the use of AI to:
analyse environmental data and provide more accurate climate forecasts to aid in planning and executing more effective environmental policies and climate change mitigation efforts;
forecast renewable energy production from sources like solar and wind power, and balance energy supply and demand to ensure efficient utilisation of renewables and reduce reliance on fossil fuels;
develop more efficient solar cells and improve production processes for manufacturing solar panels; and
predict and respond to more frequent and intense catastrophic weather events such as flooding and bushfires through, for example, riverine flood modelling and forecasting, and the prediction, identification and tracking of bushfires in real time.
Committee view
1.84The evidence received by the inquiry demonstrates that the development and use of AI technologies have a range of significant environmental impacts, most notably in respect of energy and water use; GHG emissions; and land and resource use.
Energy use and GHG emissions
1.85The committee heard that the development and training of generative AI models, and the subsequent deployment and use of AI-powered applications, such as ChatGPT-3, require vast amounts of computing power and data storage. This computing and data infrastructure is housed in specialist buildings that require very large inputs of energy and water to power and cool the computing facilities.
1.86The growing development of generative AI models in recent years, and their rapidly increasing application to the provision of consumer products and services, has seen notable increases in the energy use of computing facilities and data centres. This is due not only to the increasing number of queries submitted daily to platforms like ChatGPT-3—which on some estimates receives as many as 10million queries a day—but also to the higher amount of energy required to produce AI-generated responses compared to earlier technologies, such as a response to a simple text query generated by Google.
1.87While estimates vary, the committee heard that the energy use of data centres is currently around 1.5 per cent of total global energy use. In Australia, the energy use of data centres may be around 5 per cent of total energy use, with some projections suggesting this could grow to as much as 15 per cent by 2030. While such estimates are uncertain, the committee considers that the rate of growth in the energy use of data centres will undoubtedly be considerable, driven strongly by commercial incentives to develop new AI models and AI-powered products and services. The committee acknowledges, accordingly, the importance of governments ensuring that the energy demands of the AI industry are factored into future energy system planning.
1.88The high energy use of the AI industry also gives rise to environmental concerns regarding its associated GHG emissions. To the extent that the energy used by the AI industry is derived from fossil fuels, the development, training and deployment of AI models and applications make a corresponding contribution to GHG emissions and therefore to the problem of climate change.
1.89In this regard, the committee acknowledges that, as with all emissions-intensive industrial and economic activities, there is significant environmental benefit to be gained wherever high-performance computing facilities and data centres can derive their energy from emissions-free renewable sources. Given this, the committee notes the importance of siting such facilities in locations that have access to renewable energy.
1.90Further, the committee notes that the AI industry’s GHG emissions represent a proportion of Australia’s total emissions and, as such, fall under the government’s existing emissions reduction targets and climate policies, including:
Australia’s commitment, as a party to the Paris Agreement, to the goal of limiting the increase in global average temperatures to well below 2°C of warming and pursuing efforts to keep warming to less than 1.5 °C;
Australia’s legislated and formal emissions reduction targets under the Paris Agreement of 43 per cent by 2030 and net zero by 2050;
the government’s Net Zero plan, which involves six sectoral emissions plans, including the built environment sector plan which will capture AI facilities;
the government’s Rewiring the Nation plan to modernise Australia’s electricity grid and upgrade transmission infrastructure to support the transition from fossil fuel-based energy sources to renewable energy sources; and
the government’s Capacity Investment Scheme, which provides a national framework for investment in renewable energy capacity, such as wind, solar and battery storage.
1.91The committee considers that, subject to ensuring that AI’s energy needs are factored into future energy system planning, Australia’s current GHG emissions reduction policy settings, and progress toward a substantially renewables-based energy system, provide a robust policy framework to address the significant energy use and associated GHG emissions of the AI industry.
Water use
1.92In relation to AI’s water use, the committee heard that, despite varying estimates, AI facilities use very significant quantities of water, thereby potentially competing with social and environmental water uses and placing pressure on water resources management, infrastructure and planning.
1.93As demand for AI-related water use will only grow as the AI industry continues to expand, the committee acknowledges the need for governments to ensure that future water resources management, infrastructure and planning take account of the water needs of the AI industry. Given the varying availability of water and cycles of drought across Australia, the committee notes the importance of siting such facilities in locations that can service the AI industry’s water requirements without impacting critical social and environmental water uses. In this regard, the committee notes examples of the strategic siting of such facilities near abundant water sources, such as aquifers and hydroelectric stations, to allow for the use and recycling of water to cool computing infrastructure.
Reducing the environmental impacts of AI
1.94In light of the significant environmental impacts of AI’s energy and water use, a number of inquiry participants pointed to the environmental benefits of increasing the efficiency of the computational methods used to develop and train AI models, as well as to operate the AI models themselves, insofar as these gains translate to more efficient use of energy and water resources. In this regard, the submissions of the large AI tech companies detailed a number of efficiency gains in recent years, achieved through improvements to computing hardware and the computing methods used to develop and operate AI models.
1.95However, 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.
Land use and resources
1.96In relation to the AI industry’s environmental impacts on land use and resources, the committee notes evidence that the associated industries and activities that underpin AI and the technology industry more broadly also have significant environmental impacts. In particular, the mining and processing of key minerals; manufacturing of computing equipment; and management and recycling of waste from these activities competes with environmental land uses and can impact significantly on the quality and range of natural habitats.
1.97While acknowledging these potentially significant impacts, the committee notes that responsibility for environmental regulation and resource management in Australia is shared between the Commonwealth and state and territory governments. Given this, the committee considers that these land use and resource impacts are most appropriately considered and dealt with through existing legislative schemes for environmental protection and land-use and planning, rather than by AI-specific measures or policies.
Capturing the environmental impacts of AI
1.98A significant theme to emerge from the evidence received by the inquiry is the need for the environmental impacts of AI to be more fully captured, with data currently lacking due to a lack of consistent measuring and reporting standards, as well as commercial sensitivities in relation to data that might reveal the relative costs and efficiencies of different AI models and approaches to their development.
1.99Accordingly, various inquiry participants called for the development of comprehensive standards for measuring and reporting AI’s environmental impacts, to be applied widely across the AI industry whether as part of a voluntary or mandatory reporting framework. A number of submitters and witnesses also noted the potential for the reporting and assessment of AI’s environmental impacts to be mandated by AI-specific legislative schemes, such as the EU AI Act, or as an element of an audit or assessment framework for high-risk uses of AI.
1.100In this regard, and with reference to the discussion above, the committee notes that complete and accurate information about the full range of AI’s environmental impacts is critical for the purposes of energy system and water resource planning and management, as well as to ensuring that AI’s impacts are effectively regulated under Australia’s intersecting federal, state and territory legislative regimes for environmental protection, resource management and land use and planning.
Environmentally positive uses of AI
1.101Finally, the committee notes the evidence received in relation to the numerous potentially environmentally positive uses of AI. At a broad level, the use of AI to achieve greater efficiencies in significant areas of the economy, such as manufacturing, agriculture and the transport sector, can result in more efficient use of natural resources and reduced environmental impact.
1.102More directly, AI has many potential environmental applications, including in relation to natural resource management; environmental monitoring and conservation; and understanding and finding solutions to the environmental challenges of our time, including climate change and species extinction.
1.103The committee notes that, as with any potential uses of AI, realising AI’s potential environmental benefits requires the creation of a regulatory and policy environment that fosters the development of the AI industry more broadly, while effectively mitigating the significant potential risks of AI technology. The evidence received by the inquiry and the committee’s views on these issues are discussed in Chapters 2 and 3.
Regulating the environmental impacts of AI
1.104In relation to regulating the environmental impacts of AI, in addition to being subject to existing general schemes of environmental protection and land use and planning, the committee notes, as set out at paragraph 7.61, the government’s implementation of AI-specific and general policy frameworks that are relevant to managing and mitigating the environmental impacts of AI. These include, for example, policies providing guidance on the development and use of AI, and government procurement policies applicable to the purchase of AI systems and construction of AI facilities.
1.105The committee notes also the government’s continuing consultation on the development of mandatory guardrails for the use of AI in high-risk settings, as outlined in its Safe and responsible AI in Australia proposals paper. The government’s proposed principles for guiding the assessment of potentially high-risk uses of AI include consideration of ‘the risk of adverse impacts to the broader Australian economy, society, environment and rule of law’, and the committee notes the potential, as suggested by some inquiry participants, for the assessment of high-risk AI uses to include their impact on matters of environmental protection and sustainability.
1.106By 2023, the world’s data centres are forecast to consume 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. While onshore data infrastructure is important for data security and sovereign capability purposes, any growth in the AI infrastructure industry in Australia should be managed to ensure it is delivering value for Australians and, more broadly, is in accordance with the national interest.
1.107That the Australian Government take a coordinated, holistic approach to managing the growth of AI infrastructure in Australia to ensure that growth is sustainable, delivers value for Australians and is in the national interest.
|
Senator Tony Sheldon Chair Labor Senator for New South Wales
|