So all is well? AI is here to save us all? No. AI itself is a complex, multipurpose technology that affects our planet in many ways. For example, in terms of climate change, AI can save CO2 emissions, but it can also cause them in many ways. For example, the training of these very large language models produces significant emissions. The emissions depend much on where it is trained, and even for the same model, how exactly it is trained (from scratch vs. only fine-tuning). In addition to CO2 emissions, training machine learning models in data centers also requires significant amounts of water for cooling. And that's not all! When you factor in the emissions from using the model, as well as changes in behavior, it gets complicated quickly! And instead of using AI to more efficiently save the planet, you can just as easily use AI to more efficiently harm the planet, for example in the oil and gas sector.

And even if it has the potential for a significant positive environmental impact, it can harm some of the other perspectives. Notice the parrot? A landmark paper warns that the trend toward extremely large and complex language models can lead to models that are biased and may appear intelligent but are nothing more than "stochastic parrots" - not justifying the environmental, social, and economic risks they pose. To make such large models useful, they require huge amounts of labeled data, which are mostly collected in labor-intensive processes by underpaid, over-exploited workers.

So there can't be a definitive verdict. While it has great potential for many use cases, it may not be needed for many use cases at all. And if it is used, it needs to be used in a fully sustainable way.