A study investigating the impact of artificial intelligence on health care has shown that using large language models to process thousands of patient records daily across multiple hospitals could lead to substantial resource consumption.

Published 15 November in Internal Medicine Journalresearchers from the University of Adelaide and the University of Reading highlights Ways in which hospitals can use AI responsibly—including using shorter prompts to summarize patient data.

Oliver Kleinig, who led the research from the University of Adelaide, said, “Every day you are in hospital, doctors, nurses, and other hospital professionals are documenting pages and pages about your health. By the end of a hospital stay, it is Possible to accumulate tens of thousands of words to your name. Unlike busy health care staff, private large language models similar to ChatGPT have time to read through and process this information.

“However, with great processing power comes great responsibility. A single AI query uses enough electricity to charge a smartphone 11 times and consumes 20 milliliters of freshwater in Australian data centers. ChatGPT is estimated to use 15 times as much energy as Google.

“Implementing large language models across health care could have very significant environmental consequences. Hospital bosses need to think carefully about where and when artificial intelligence should be used in their organizations.”

Questions to consider

ChatGPT’s daily carbon emissions already equal that of 400-800 US households. Health care AI systems would likely have an even larger footprint, as they require more powerful models to handle complex medical information and must be run locally for patient privacy.

Beyond energy consumption, the hardware needed for these AI systems requires extensive rare earth metal mining, potentially causing habitat destruction. The manufacturing process alone can double the carbon footprint of AI operations.

To reduce the impact of hospitals and medical centers on the environment, the researchers propose five key questions health care providers should consider before implementing AI systems, including:

• Does my organization need a large language model? Could existing technology be sufficient?
• What LLM should I choose? Use the smallest possible model to decrease resource consumption—smaller, fine-tuned LLMs can outperform larger applications.
• How can I optimize my LLM? Use smaller and specific prompts to reduce the carbon impact of applications. Succinct prompts with refined information are more energy efficient.
• What hardware should I run my LLM from? Using hardware that runs on renewable energy is preferable.
• What data should I share? Maximize LLM efficiency by sharing data where appropriate.

The study suggests AI could potentially reduce health care’s environmental impact in other ways, such as improving patient flow and reducing paper use.