Why generative AI is not so smart

The substantial water and energy consumption of generative artificial intelligence (GenAI) usage is acknowledged by many in the scientific community, yet public awareness of the ecological impact of AI usage is remarkably small.

While certain social stigmas surround its usage - namely how it facilitates laziness or lack of creativity by the user - discussions around how AI damages the environment are rare outside of scientific debates.

GenAI works by using large-scale machine learning models (usually ‘neural networks’) which have been trained on massive datasets to learn patterns and structures within that data. This enables AI to form a sentence which looks like a human has written it. Many believe that AI can solve problems by ‘thinking’ about them, but it cannot  understand the human-written prompt or its own generated text. It acts as if it ‘understands’, but it doesn’t.

The computational power needed to train and enable the decision-making capabilities of these machines is immense. All logic is performed by supercomputers, which are running 24/7 to train them on datasets. This training can consume 1,287 megawatt-hours of electricity - equivalent to powering 100 US homes for a year. Additionally, every time a user sends a query to the machine, the entire model activates, using power hungry graphics processing units (GPUs) that consume nearly 4 times as much as traditional computer chips.

The water consumption required to support these computers is a large concern of GenAI. Computers are like radiators, where most of their energy is converted to heat. Water cooling systems are used to cool the servers by evaporating the heat. This water must be drinking quality because impurities can damage the servers.The servers built for AI process huge amounts of data, all the time, with significant  cooling demands. . For scale, 500 milliliters of water is used upon every request with ChatGPT.

Therefore, while GenAI is a transformative technology, the importance of conscientious use cannot be understated. Making small changes in how we interact with these systems can collectively make a big difference.