AI's carbon footprint appears likely to be alarming

Judging from Wall Street's recent tech boom, artificial intelligence systems are likely to become increasingly ubiquitous and profitable. AI systems are also said to be on course to improve labor productivity, supply chains, and medical care; boost the economy; and even help to solve the environmental challenges facing the planet, including global warming. But what if using AI contributes to more carbon emissions that accelerate climate change? Experts have begun debating that question—and the answers are far from reassuring.

AI tools can be designed to detect greenhouse gas (GHG) leaks from pipelines, monitor deforestation, contribute to the design of novel materials with lower carbon footprints, among many other climate-friendly endeavors. But just how much energy does AI actually consume in achieving these benefits? Is the environmental cost of AI energy consumption currently greater or smaller than the benefits it can deliver to combat climate change?

Unfortunately, the data documenting the carbon footprint of AI are scarce or nonexistent. In fact, AI companies such as OpenAI are not fully transparent about the costs of developing their systems, processing deep learning algorithms, and training their large language models (LLMs). As different countries take on the challenge of regulating the AI sector, full transparency, particularly regarding carbon emissions associated with the industry, needs to be given much greater priority.

Billions of devices connected to the internet could be responsible for up to 3.5 percent of global carbon emissions by next year. Energy consumption in data centers is associated both with electricity use, as well as with the continuous, 24/7 use of hefty air conditioners, without which computers and servers would overheat. The AI sector is heavily dependent on data centers, and its increasing use and dissemination will inevitably raise data center carbon emissions over the next several years.

A recent study conducted by Cornell University scientists found that training LLMs like GPT-3 consumed an amount of electricity equivalent to 500 metric tons of carbon, which amounts to 1.1 million pounds. A typical coal-fueled power plant working continuously for 24 hours burns about 2.7 million pounds of coal.¹ Training LLMs is therefore equivalent to burning coal for 10 straight hours, or nearly half a day. Importantly, these models are not trained only once: LLMs need to be systematically trained on the most up-to-date data not only to remain relevant and accurate but also to deliver improvements over time. It is not an exaggeration to say that many policymakers are counting on the promised improvements of AI for the desired increase in productivity they envisage, an outcome that may prove difficult to achieve.

As the authors of the study note, evaluating the environmental impact of LLMs is no easy task, since the companies responsible for developing these models are frequently reluctant to openly discuss how they are trained and retrained. Notwithstanding their opacity, researchers have measured the amount of energy consumption by analyzing the volume of specialized hardware needed to feed GPT-3, GPT-4, and other similar deep learning algorithms. By measuring the energy consumption of specialized hardware alone—thus leaving aside the energy needed to power up servers and other necessary machines—scientists have concluded that the carbon footprint of training an LLM is equivalent to some 626,000 pounds of carbon dioxide.

Once the models have been trained and are ready for use, they likely consume even more energy than what they use up for training alone. The act of asking GPT-4 a question—or typing in a prompt to generate a response—is called inference. Some researchers and industry experts have concluded that inference probably consumes more energy than training, a finding that, if confirmed, is alarming.

After all, anyone who has ever used an LLM like GPT-4 knows that rarely does a query end with one response generated by a single prompt. In fact, one either has to submit further queries to get a more relevant response, or multiple prompts might be needed to arrive at an adequate answer. Also, one is often tempted to continue probing the capabilities of the model, often unnecessarily—the temptation to test the limits of LLMs, or how long will it take for the model to come up with an absurd response, is almost irresistible. Unknowingly, therefore, users are frequently generating large amounts of carbon emissions, contributing to the carbon footprint of AI.

In a recent paper, Google identifies four practices that would reduce the energy and carbon footprints of AI systems. They are:

1. Using sparse models, with fewer parameters.
2. Using processors that are specifically designed for machine learning training, rather than general-purpose processors.
3. Using cloud-based data centers, which are more energy efficient than local data centers.
4. Optimizing location through cloud-based data centers: The cloud allows selection of locations with cleaner energy, thus minimizing the carbon footprint.

Although the practices sound sensible, governments and users alike have no means of knowing whether the AI sector is adopting them. Take GPT-4, OpenAI's latest version of its LLM. Based on recent leaks, some industry insiders estimate that it has 1.76 trillion parameters, which makes it ten times bigger than GPT-3 with its 175 billion parameters. That's hardly a move towards greater sparseness. In fact, the trend reveals that there are currently no economies of scale when it comes to LLMs: The larger they are, the more energy they use.

What about the use of specialized processors and cloud-based data centers? Again, we are in the dark since AI companies don't provide that information.

The little that one can surmise about AI's carbon footprint indicates that it is big, and it is likely getting bigger as models become more ambitious and companies spend a growing amount of time training and retraining them.

The hype over AI has also led many to use these models and systems without questioning whether they run counter to the environmental goals they purportedly will help achieve. There are many reasons to regulate the AI sector. The most pressing among them is to bring transparency to AI's environmental impact. Failing that, any elusive productivity gains from this new technology could be swamped by irreparable damage to the environment and to the efforts to combat climate change.

Note

1. A typical coal-fueled power plant uses 1.14 pounds per kilowatt hour (KWh), according to the US Energy Information Administration.