Why AI is a disaster for the climate
The carbon footprint of server farms used to power generative AI could be problematic. Amid all the hysteria about ChatGPT and co, one thing is being missed: how energy-intensive the technology is. John Naughton discusses the environmental impact of AI.
The Environmental Impact of AI
The current hype around AI has diverted attention from the downsides of the technology, such as bias, inscrutability, unaccountability, and its tendency to 'hallucinate'. However, one pressing question that is missing from public discourse is the environmental impact of a world suffused with AI. The technology's tremendous energy consumption and CO2 emissions pose a significant challenge.
AI relies heavily on computing power, which requires electricity and cooling systems. The GPUs used in AI systems run at high temperatures, further increasing their electricity consumption. This translates into large-scale CO2 emissions. While the industry claims carbon neutrality through offsets, the reality is that AI everywhere would create a technology-dependent world with a substantial environmental footprint.
The Carbon Footprint of Generative AI
A study conducted in 2019 estimated that training a single large language model like GPT-2 emitted around 300,000kg of CO2, equivalent to 125 round-trip flights between New York and Beijing. Since then, AI models have grown exponentially, leading to proportionately larger carbon footprints. However, training is just one phase of the generative AI life cycle.
The inference phase, when AI systems are in service and interact with millions or billions of users, contributes to the technology's carbon footprint. Tasks involving text generation, summarization, image generation, and captioning are particularly energy- and carbon-intensive. Surprisingly, training AI models remains even more carbon-intensive than using them for inference. The carbon cost of training can only be offset by a significant number of inference interactions.
The Urgent Need to Address AI's Impact
The environmental impact of AI cannot be ignored. The carbon footprint of AI systems continues to grow as their usage expands. The recent attempt to estimate the carbon cost of the inference phase highlights the energy and carbon intensity of generative AI tasks. The growth of ubiquitous AI raises concerns about its impact on the planet and the necessity to find sustainable alternatives.
To mitigate the environmental impact of AI, it is crucial to explore energy-efficient alternatives and solutions. A world where generative AI falls into the 'trough of disillusionment' would give us an opportunity to reassess the technology's environmental impact and prioritize sustainable practices. It is essential to address these concerns to ensure the responsible and sustainable development of AI.
