When does it come to the point where a technology needs to be made widely available? There is no question today that generative AI (GAI) is a tool that is available to everyone, but the question is what price we pay for it. The environmental burden of GAI could soon rival the energy consumption of entire countries. Without dismissing the benefits of the technology, the question arises: wasn’t it a hasty step to immediately turn something into a market product, the effects of which we can only estimate even in the short term?

Source: DALL E 3.

There has been a lot of buzz around AI lately. Many of the related articles focus on the potential of the technology, while others discuss its drawbacks. In many industries, AI is only just beginning to be exploited. Examples include the pharmaceutical industry, where the market for AI-assisted developments was estimated to be worth close to half a billion dollars by 2022, and the healthcare sector in general, where health records summarization and AI-assisted search are expected to increase efficiency.

The other side of the coin is problems such as malicious exploitation of vulnerabilities in existing machine learning models. One example is the vulnerability discovered by a group of researchers in relation to ChatGPT. This vulnerability allowed a user to bypass the language model’s content moderation and usage policy and generate malicious code that could be used to hack into online commercial databases – a problem that OpenAI has since fixed.

In the mainstream media, there is generally less talk about the impact of AI developments on the environment. This is particularly true in the case of GAI, which has become dominant since the last year.

It should be added that the problem was not so pronounced until recently. What has changed the most is the energy requirements of machine learning models, which have increased significantly with the spread of generative models. All machine learning models essentially require computing power in two phases (which also has a significant impact on their energy requirements). One is the training of the models, and the other is the cost of predictions made when using the completed models. Traditional machine learning models, which were dominant until the 2010s, worked on completely different principles than those emerging today. In terms of the energy used, they were highly asymmetric, requiring only a negligible fraction of the cost of their training during subsequent use (inference time).

The situation started to change with the so-called transformer-based models. This type of architecture is still the basis for the state-of-the-art GAI solutions. The application of such neural network architectures started to require a significant energy investment also at inference time. This was mainly due to the increasing number of parameters in these models. More parameters meant more computations, for which from this point onwards the graphics cards (GPUs), which are also well known in the games industry, had to be used.

ChatGPT is estimated to have had around 590 million user interactions in January 2023. Calculated at 5 questions per user, this consumed as much energy as 175,000 people would have used in the same period. It is worth noting that this is before the official release of OpenAI’s GPT-4 model. This is also thought-provoking, as it is known that the latter is exponentially more energy-demanding than its predecessor.

Given the increasing reliance of large tech companies on GAI (mainly to enhance the user experience), it is reasonable to assume that we are only at the beginning of a period of rapid growth. Google’s parent company Alphabet, for example, is not secretly looking at how it can support GAI-based query answering in its search engine. Considering an earlier statement by the head of the company that a search with LLM in the background uses about 10 times more energy, it is easy to see how quickly the situation escalates.

Of course, the situation is much more complex than this, since the rapid obsolescence of the GPUs used to operate the GAI significantly increases the amount of e-waste, puts even more pressure on the supply chains, and the extraction of the (often rare) raw materials used for GPU production entails additional environmental burden.

The problem, in my view, is mainly that a technology that can have such a big impact on our lives in the long term has been made available to everyone by industry. Of course, this is logical on the one hand since the investment required for development must somehow pay off. But on the other hand, it is worth considering whether such an environmentally damaging technology really needs to be available to literally everyone. Would it not be more appropriate to exploit such models for research and development purposes and wait until such devices are available to the public in a sustainable way before commercializing them?

There is already a lot of research on reducing the size of large language models and increasing their efficiency. They promise that without any significant performance degradation, the reduced sized, more efficient models can still offer about the same capabilities as the original counterparts. This will clearly have a positive impact on the technology’s environmental footprint.

The fact that everyone can benefit from developments in a particular industry (e.g., artificial intelligence research) is a fundamental value to be defended. However, in practice, it is worth considering when the point is reached where a particular solution can be made available to all. Of course, in a globalized world, there is fierce competition: whoever offers the latest, best, most modern solution will make the most profit, and the key to this is to get the solution to as many people as possible in the form of a product. But this should not override our responsibility for the environment in any way!

Perhaps precisely because of the above, it is important that the ethical guidelines related to the ethical development of AI also take this aspect into account, and that these guidelines can be observed and enforced much more effectively than today. Thoughtful planning in connection with the development of a technology or its introduction as a product is, after all, is the basic interest of all of us.

István ÜVEGES is a researcher in Computer Linguistics at MONTANA Knowledge Management Ltd. and a researcher at the Centre for Social Sciences, Political and Legal Text Mining and Artificial Intelligence Laboratory (poltextLAB). His main interests include practical applications of Automation, Artificial Intelligence (Machine Learning), Legal Language (legalese) studies and the Plain Language Movement.