Summary |
|

Rapid industrialisation over the past centuries has progressed the world beyond what could ever be imagined. However, amidst this innovation, the environmental toll of unchecked industrial growth has become increasingly evident. The latest innovation that has come under scrutiny for its environmental impact is Artificial Intelligence (AI).
The Environmental Impact of AI
Although AI is often presented as clean tech, exact calculations of emissions are often lacking due to a lack of transparency and the absence of a predefined way of measuring AI emissions. What we do know, is that AI consumes large quantities of electricity. A study from the University of Massachusetts calculated that training an AI model potentially emitted more than 283.000 KG of carbon dioxide – which would be the same level of emissions as driving 62.6 non-electric vehicles a year. AI language models have only grown exponentially since this study was conducted.
A more recent study found that ChatGPT consumes about 500ml of water for every 20-50 simple questions and answers. This water is needed to cool the data centres. It is difficult to reuse this water because chemicals are added. Another study also showed that generating one image with a generative AI, such as OpenAI’s DALL-E 3, costs the same amount of energy as fully charging your phone. Unfortunately, AI is using electricity at a rate faster than renewable energy would be able to supply it. There is a huge energy gap in AI.
Finally, AI is employed at a large scale by the fossil fuel industry. Since 2019, Microsoft has partnered with oil giant ExxonMobil. At the same, ExxonMobil claimed that by using Microsoft’s AI solution, it could optimise its mining operations and, by 2025, increase its production by 50,000 barrels per day. Ironically, Microsoft aims to be carbon-negative by 2030, and more transparency is needed on whether the use of its software in incredibly polluting practices is included in its calculations.
However, it must be highlighted that AI is also an excellent tool to mitigate climate change, as entrepreneurs continuously find new ways to employ AI to make our responses to climate change more accurate and efficient. AI can help us to identify trends and monitor climate change.
Towards a Green AI
According to a Nature study, the emissions of an AI model depend on the location of the training server, the energy grid it uses, the length of the training procedure, and the hardware used. The authors of this study advocate for ‘green AI’, which they defined as “AI research that yields novel results without increasing computational costs, and ideally reducing it”. In 2019, Roel Dobbe and Meredith Whittaker put forward seven policy recommendations that would help stimulate green AI. These recommendations include (i) mandate transparency; (ii) account for the full-stack supply chain; (iii) watch out for rebound effects; (iv) ensure energy is considered in policies that are not energy-related; (v) integrate tech and climate policy; (vi) curb the use of AI to make fossil fuel extraction more efficient; and (vii) address AI’s impact on climate refugees.
Current Practices: Google’ Green AI
Contrary to Microsoft, Google is much more transparent when it comes to the environmental impact of its AI and the steps it is taking to mitigate environmental impact. Google has committed to building AI responsibly and to manage its environmental impact. The practices employed by Google to train its AI models can reduce the required energy by up to 100 times and reduce emissions by up to 1,000 times. Moreover, Google’s data centres are 1.5 times as energy efficient as average enterprise data centres. A final aspect of Google’s green AI strategy is its efforts to cool its data centres through responsible water use practices.
Regulating a Sustainable AI Revolution
Regulation is clearly needed. This has been a challenge for several reasons. Regulation is notorious for lagging behind innovation. Jurisdictions are just now beginning to catch up to regulating AI in general. On a global level, both the Organisation for Economic Co-operation and Development (OECD) and the Global Partnership on Artificial Intelligence (GPAI) are working on developing policy recommendations. In individual jurisdictions, many countries have published rules and guidelines for regulating AI, but there is a glaring lack of focus on environmental factors.
The European Union (EU) is currently in the final stages of implementing its AI Act. In its current form, the EU AI Act states that “[p]roviders should also be encouraged to apply on a voluntary basis additional requirements [such as] environmental sustainability”. There are no further specific laws or guidelines on the topic. The EU AI Act would be the only
However, even though there is a lack of regulation that directly addresses green AI, there is plenty of regulation on data centre energy efficiency in different jurisdictions. Data centres host the computing infrastructure needed to run AI and have previously attracted regulation due to their large energy consumption. The most comprehensive legislations are again found in the EU, with four regulations being relevant:
The EU’s Energy Efficiency Directive requires owners and operators of data centres with a minimum capacity of 500 kW in the EU to disclose their energy consumption; the proportion of renewable energy used; water consumption; and the amount of waste heat reuse. The first relevant deadline will be 15 May 2024 for 2023 data.
The EU’s Taxonomy Environmental Delegated Act applies to investors and requires them to determine whether a data centre they are investing in will comply with the best sustainability practice descriptions. Failure to do so may apply a cost premium to the capital.
The Corporate Sustainability Reporting Directive requires commercial data centres to provide energy and sustainability reporting data to their customers and authorities.
The Corporate Sustainability Due Diligence Directive applies to the responsible behaviour of EU companies and ensures that social and environmental considerations are taken into account in the company governance (this is only a provisional agreement, the Directive is not in force yet).
In the United States, legislators are also working on comprehensive regulations to address data centre emissions. For example, Senator Sheldon Whitehouse is stated to be “developing draft legislation to address both crypto-asset and conventional data centres, using the [European Energy Directive] as a blueprint”.
In summary, the environmental impact of AI, from energy consumption to its role in polluting industries, necessitates sustainable solutions. Global regulations are evolving slowly, although existing data centre regulations offer a partial solution. But if we refer back to the 7 policy recommendations by Roel Dobbe and Meredith Whittaker, it seems current legislation only addresses one of seven (mandating transparency). Transparency alone is not sufficient to ensure sustainability, so ongoing efforts are needed to harmonise the AI revolution with environmental well-being.
Comments