Environmental Impacts of AI

A common question we’re asked is how we're thinking about the environmental impacts of AI. At Fundsorter we’re all about equity, and any robust concept of equity has a safe and livable planet at its centre. We’ve been digging into the data and, TLDR: AI has negative environmental impacts. BUT it’s our belief that these can be managed and mitigated through good governance and regulation - which we don't currently have. 

How do we manage AI use in this environment? Our basic principle: only use AI where it adds genuine value through efficiency, quality, and/or capability. Fewer AI-generated action hero images, more AI-enhanced community funding. 

Governance and regulation are required to reduce the environmental impacts of AI

Electricity

AI uses electricity in two ways: processing data to answer queries, and training AI models.

According to a 2021¹ paper², training ChatGPT-3 consumed around the same amount of electricity as powering 173 average New Zealand homes for a year. The latest models (like ChatGPT-4) likely use 2-3 times more. Data centers consume about 1.5% of the world's electricity³, but most of that power is being used on streaming, social media and online shopping, not on AI. AI accounts for 15% of overall data centre electricity, or about 0.2% of the global total. 

That’s not too bad. But without additional efficiency improvements, global data centre electricity consumption could increase by over 75% between 2022 and 2026 (rising from around 460 terawatt-hours in 2022 to more than 800 Terawatt-hours in 2026⁴). AI is likely to take up a bigger and bigger slice of that energy consumption. How big that slice gets depends on a lot of things, like monetisation of AI, the efficiency of AI models and the speed of transition to AI driven work. The carbon impact of that electricity use also has dependencies, like the percentage of renewable versus fossil fuels, clever optimisation & load sharing across data centres and (my favourite) good governance and regulation.

Water Consumption

For me a greater concern is the amount of freshwater AI uses. Specific annual figures for AI's current water use are hard to isolate. But we do know that just under half the world’s population (3.6 billion people⁵,⁶) experience water scarcity each year​. In some places, data centres compete for the same water supplies. 

Google alone used 21 billion litres of water for cooling in 2022, and 23 billion in 2023. Most of this was used on data centres, though not specifically for AI purposes⁷. Training a large model like GPT-3 reportedly consumed 700,000 liters of water directly on-site (enough to fill a small swimming pool)⁸.

By 2027, global water withdrawal for AI is projected to reach around 6 billion cubic meters (B m³) annually. This is more water than Denmark uses annually, or enough to fill up to 2.6 million Olympic-sized swimming pools⁹. This increasing demand could worsen water stress in some regions by 2030.

Like electricity usage, this is a solvable problem, but only if there is motivation and investment. Evaporative cooling techniques can be replaced by air cooling or closed-loop systems and data centres can be built or relocated to cooler climates. None of this is likely to happen without good governance and regulation.

Can AI have a positive climate impact?

It’s not all bad news though - AI can have positive climate impacts. These tools excel at identifying complex patterns, analysing large datasets, and optimising systems, which can have positive environmental impacts:

• Energy optimization: Electricity grid management and traffic flow improvements

• Water conservation: Leak detection, infrastructure monitoring, and smart irrigation (critical as agriculture uses 70% of global fresh water)

• Climate research: Prediction modeling and monitoring to inform policy

All of these problems are solvable, but to solve them we need better governance and regulation. Recent moves, like the UK’s AI Energy Council and the new EU Artificial Intelligence Act (AI Act) are a big step in the right direction. As a community, and a community sector, we can advocate for stronger AI regulation in Aotearoa. New Zealand currently has no dedicated AI law yet but is actively consulting and likely to regulate high-risk AI uses by 2026–2027. In the meantime, we can limit our AI use to the useful, not the frivolous. 

References:

1. Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L. M., Rothchild, D., So, D. R., Texier, M., & Barham, P.(2021). "Carbon Emissions and Large Neural Network Training." arXiv preprint arXiv:2104.10350.↩︎
2. We couldn't find any robust recent data ↩︎
3. International Energy Agency reports ↩︎
4. International Energy Agency (IEA). (2024). Tracking Clean Energy Progress 2024: Data Centres and Data Transmission Networks. Paris: IEA. ↩︎
5. Mekonnen, M. M., & Hoekstra, A. Y. (2016). Four billion people facing severe water scarcity. Science Advances, 2(2), e1500323. DOI: 10.1126/sciadv.1500323 ↩︎
6. UNESCO (2024). United Nations World Water Development Report 2024: Water for Prosperity and Peace. ↩︎
7. Google 2024 Environmental Report (PDF) ↩︎
8. Research paper "Making AI Less 'Thirsty'" (arXiv:2304.03271). arXiv Link. ↩︎
9. Source: Research paper "Making AI Less 'Thirsty'" (arXiv:2304.03271). arXiv Link. ↩︎