|
TECHNOLOGY & POWER
AI and Power
by Andrew Cox
One could argue that the arrival of generative AI (gen AI) has been almost as helpful for highlighting existing problems as through its direct benefits. The intense controversy swirling around AI has reactivated important debates within society about the growing social impacts of digital technologies.
For example, the information failures of ChatGPT have reminded us about the limits of reliability of information that many digital tools have. I know librarians and information professionals are well aware that there are biases in search engine results. Now, the publicity around biased images from gen AI has made people outside of our profession aware that, without careful use, digital tools can promote damaging stereotypes.
Another issue that ChatGPT and its ilk have brought to the fore is the problems arising from the unregulated power of technology giants. We already knew their power was an issue for our privacy and freedom, but the disruptive impact of AI, particularly on education, has revealed it very starkly. Hopefully, the public (and regulators) will take notice of this.
ENVIRONMENTAL IMPACT OF AI
There has also been a surge of concern around the environmental impact of digital technologies, provoked by evidence of the intense power demands of AI. (Another form of power!) The news headlines during 2024 revealing that Google and Microsoft had failed their carbon emission targets because of AI development were part of increasing coverage in the mainstream media that there is a sustainability issue with AI.
Again, librarians and information professionals know very well that digital content and tools are not somehow magically created without a material infrastructure to support it. Our everyday digital activities, especially streaming, create carbon emissions (Griffiths 2020). Driving to work uses resources, but so does a Zoom call. Hopefully, the heated debate around gen AI is prompting a wider audience to start asking questions about digital sustainability.
The environmental impact of computing, including AI, occurs at many points throughout its hardware and software lifecycles (Estampa 2024; UNEP 2024). The manufacture of computing devices requires mining for metals and minerals, such as gold, copper, nickel, and cobalt. Some of these mining operations are very exploitative in character, and the raw materials seem to be used wastefully (Estampa 2024).
But that’s not the only problem. IT manufacture is also highly polluting. The IT industry generates carbon emissions from the transportation of devices around the world. There is also the power required to manufacture devices and then use them to perform tasks. Another form of digital’s environmental impact centers on the use of water to cool data centers: Sometimes, these demands are leading to drinking water shortages. At the other end of the lifecycle, recycling of materials from technology equipment remains poor (Estampa 2024).
Microsoft plans to alleviate the problem of using fossil fuels to power data centers by turning to nuclear energy. It has received government permission to reactivate the Three Mile Island power plant.
These issues apply to all digital technologies. But gen AI has a particularly strong environmental impact. Training and using gen AI are demanding in terms of power consumption. Just the storage of data to train AI has its own impact. Generating images is particularly demanding. And, unfortunately, it seems that the current growth in demand is being met by less clean energy generation in the form of power stations using fossil fuels. Higher levels of demand are also triggering the construction of new data centers, again with huge environmental impacts. The high-performance graphics processing units (GPUs) often used to run AI processes are particularly environmentally impactful due to the raw materials needed to manufacture them and the power required to operate them.
Worryingly, critics suggest that such “infrastructural harms” fall disproportionately on the majority world (Valdivia 2024). For example, extractive mining is having its worst effects in the Global South, in countries such as the Congo and Peru (Estampa 2024). We also know that much e-waste continues to be dumped in African countries. So, it’s the social justice as well as environmental dimensions of sustainability that are impacted by AI. The wonderful Cartography of AI by Estampa (2024), echoing Crawford’s work (2019, 2021), brilliantly illustrates some of the exploitative and extractive elements across the AI infrastructure.
AI FIGHTING CLIMATE CHANGE
Counterbalancing such points, of course, there is also the potential for AI to be used to fight climate change. For example, AI can be used in agriculture to improve crop yields and make water use more efficient. AI has the potential to help save energy by enabling people to work from home, improving travel efficiency, predicting power needs to make generation more efficient, and helping consumers monitor their energy use more effectively (Royal Society 2020). AI can also be used to model climate systems and environmental change to make more effective interventions.
We also must think about what is a fair benchmark for the environmental impact of AI. Tomlinson et al. (2024) suggest that using gen AI to write text or produce an illustration actually has significantly less impact than a human performing the same task, when you consider the consumption of energy from their computer and their personal daily environmental impact. The difference is much greater when that human is based in the Global North, where per-capita environmental impact is greater. This calculation does not take into account the whole lifecycle impact of AI or indeed the human impacts of job displacement. But it is thought-provoking.
We also know that there are efforts being made to create greener AI. Smaller language models seem to be able to match the performance of LLMs with much less environmental impact. Not everyone agrees that the power demands from AI will actually be significant (Ritchie 2024).
Returning to the question of the unregulated power of Big Tech, it would be helpful if the current gen AI providers were far more open about (among other things) the environmental impact of their AI (Smith and Adams 2024). The failure to declare this in open and standard ways is one important aspect of the problem. Having that information would at least give us a chance as institutional customers and individual consumers to opt for green AI.
LIBRARIANS SHOULD SPEAK OUT
As with many issues around AI, we have been thinking about digital’s environmental impacts for a while. Gen AI seems to have intensified the debate. One could argue that this is a useful opportunity. Hopefully, we as librarians are speaking up in these conversations not only to keep asking about the sustainability of AI but also to educate people about the informational issues around AI.
I also think it is something that students need to be made more aware of. They are not, surprisingly, the most passionate about climate action: It will affect them directly throughout their lives. Would they view using gen AI in the same light if they realized its climate impact? This needs to be part of our AI literacy offerings.
I certainly don’t count myself as any sort of expert on the environmental impact of technology. This column is an attempt to share a few thoughts based on my own desk research. As a reader, I hope you will take time to look at some of the resources I have cited, share them with others, and continue asking questions. And I am sure it’s a topic that will be increasingly recognized as central for the Technology & Power column.
|
