More data on AI’s climate impact won’t save us from inaction

The environmental costs of the development and use of AI have recently been described as ‘out of control’. Both Microsoft and Google are now cautious about their chances of meeting the net-zero targets they set for themselves. Google, who admitted to a 48 per cent increase in its greenhouse gas emissions since 2019, blame much of its loss of confidence on ‘the uncertainty around the future environmental impact of AI, which is complex and difficult to predict’.

When those who hold the most power over the AI ecosystem – and therefore have control over its negative environmental impact – give up on the net-zero challenge, what can be done and by whom?

In a recent article, economist Mariana Mazzucato argues that we should not let big tech off the hook so easily. She criticises the lack of transparency over energy consumption by tech companies and calls for governmental action to hold them accountable. Mazzucato aligns with others who argue that lack of transparency and clear data standards prevents regulators from taking action on big tech’s environmental impact.

Europe’s recently approved Energy Efficiency Directive addresses this gap by requiring data centres with at least 500 kW of information technology energy demand to report on their water and energy consumption and renewable-energy use. Although this is a step in the right direction, it assumes that we can hold the bigger polluters accountable by collecting more data, establishing data standards and comparing the impact of different companies. Ultimately, the argument goes, carbon-accounting measures will lead big tech to reduce its energy consumption. However, generating more accurate and comparable data is neither easy nor enough to address the emergency of climate change.

The metrics challenge

Measuring the environmental impact of digital technologies is a complex operation. The data collected and shared varies widely as it depends on the methodology used. Digital carbon calculators, such as Green Software Foundation tools, Impact Framework, Green Algorithms and Cloud Carbon Footprint, have proliferated, with each approaching the measurement of carbon emissions in a different way. One calculator may estimate energy consumption for data storage and compute, and use that as a proxy for carbon emissions. Another may consider the degree to which a data centre’s energy supply comes from renewable sources. A third tool might also consider the carbon present in the hardware constituting the digital infrastructure.

Sharing energy-consumption and carbon-emission data and making it truly accessible requires that we agree on the standards, criteria, models and methodologies used to calculate and evaluate it. Choosing among different available approaches is a value-laden action. Different research communities make sense of numbers according to their discipline’s purview and the notions that orientate its image of the world. Data about energy consumption, like any other form of data, does not exist in a vacuum and depends on social context.

Efficiency isn’t everything

Even if regulators and tech developers agreed on a fair and accurate way to calculate and present energy consumption and carbon emissions, these metrics would be woefully inadequate to convey the full extent of a company’s environmental impact. Well-marketed net-zero strategies and approaches to ‘sustainable’ or ‘green’ AI have focused on improving the efficiency of digital technologies, but we should not conflate efficiency improvements and optimised energy consumption with reduced pollution.

Such conflation is a problem for two reasons. Firstly, by emphasising energy efficiency, companies risk neglecting the broader environmental impact of digital infrastructures – such as the large amounts of electronic waste produced by data centres, much of it containing mercury, lead and other hazardous substances.

Secondly, prioritising efficiency risks ignoring so-called ‘rebound effects’ – the social behaviours that develop in response to a regulatory or technical intervention but backfire against its intended purpose.

For example, a company may succeed in reducing a device’s energy consumption, and therefore its carbon emissions, by improving its efficiency. This may make the technology cheaper and lead to an expansion of its market. Broader market distribution and adoption means that the overall energy use grows and greenhouse gas emissions increase.

Rebound effects can be very difficult to predict and are rarely incorporated into digital technology innovation processes and the policies surrounding them. Big tech’s tendency to conflate environmental sustainability with efficiency diverts attention away from the impact of societal behaviours on digital consumption and the environment.

Not just a ‘numbers game’

Framing the negative effects of AI on the climate as purely an issue of data transparency may validate the attitude that societal and ecological issues can be addressed by getting the numbers right. Such a view bolsters the tendency to ‘improve the numbers’ through technological solutions without paying enough attention to the wider context.

For example, AI corporations have been trying to reduce carbon emissions by shifting their AI processing across the globe and transferring it to remote data centres where local grids derive a large proportion of their energy from renewable sources. However, this workaround ignores how energy grids balance supply and demand. While big tech may claim that the real-time shifting of AI processing reduces emissions, what often happens is that the local grid, feeding the remote data centre, fires up its fossil fuel supply to satisfy an unanticipated spike in demand, as fossil fuel responds faster than renewable or nuclear alternatives. Meanwhile, the original grid doesn’t ramp down its fossil fuel energy supply – that energy is simply wasted.

Delay tactics

Solely calling for more data on AI’s impact on climate aligns with an embedded policy approach that has justified climate inaction for years: ‘let’s wait until we have more evidence’. This approach points to a double standard in the way those in decision-making positions deal with uncertainty in the context of innovation.

Projections of how AI contributes to environmental goals are often speculative and sometimes misleading. ‘Vague promises and vaporware’ are set against the demonstrable scale of AI’s adverse environmental effects. Yet such uncertainties do not lead to procrastination in the same way that lack of standardised data about AI’s environmental harms does. We take action, by investing in AI, despite doubtful projections around future benefits. Meanwhile, we do nothing about AI’s present environmental impact – only wait until the day we will supposedly be able to better quantify it.

The importance of a holistic approach

As big tech companies retreat from their net-zero commitments because of the pace of technological innovation, we must remind them that they currently occupy the driver’s seat and are responsible for the positive or negative contribution of AI to the environmental crisis. At the same time, regulators must own up to the role they have in holding industry accountable. While transparency around carbon emissions is an important step in this direction, collecting data and making it accessible is not a fix. Not only are the numbers that would account for the environmental impact of the whole apparatus of AI production and use impossible to calculate, but even if we could calculate them, what matters is how we act (or don’t) in response to what we know.

Mazzucato calls for the adoption of a holistic perspective to mitigate the impact of the tech industry on the environment and to approach the climate crisis in a systematic way. We agree with this and argue that such a perspective should question whether asking for more data evidence is really addressing the climate crisis or just serving the interests of big tech. To approach the climate crisis in a holistic way, we should instead focus on establishing adaptable frameworks that can operate based on what we already know and integrate new data over time as well as bring together different research communities to challenge each other’s underlying assumptions.