Researchers are warning that the way users interact with artificial intelligence systems could have a measurable environmental impact. A new United Nations-supported study suggests that shorter and more direct prompts may significantly reduce the energy consumed by models like ChatGPT.
Shorter prompts lead to lower energy use
The study, published by researchers affiliated with the United Nations University Institute for Water, Environment and Health (UNU-INWEH), argues that unnecessary wording in prompts increases computational load. Even polite expressions such as “please” or “thank you” add extra tokens that the model must process.
According to the findings, removing these phrases and using more concise instructions could save between 87 and 98 gigawatt-hours of electricity per year, based on ChatGPT-level usage alone.
Large language models process input in units called tokens. The longer the prompt, the more tokens are required. This not only increases processing time but also raises the energy required to generate a response. In many cases, longer prompts also lead to longer outputs, further increasing total energy consumption.
Every interaction adds up at scale
Researchers highlight that the scale of daily usage is what turns small inefficiencies into a larger environmental issue. ChatGPT alone is estimated to handle around 2.5 billion queries per day, while AI-generated summaries integrated into search engines contribute to an even larger volume of interactions.
The study also points to behavioral patterns, such as extended conversational exchanges with chatbots, as another factor increasing energy demand. Continuous back-and-forth interactions require repeated model inference cycles, which multiply total compute usage.
“We are not saying users should be rude to AI,” said UNU-INWEH researcher Kaveh Madani. “But avoiding unnecessary conversational loops can make a meaningful difference in energy consumption.”
Image and video generation carries a much higher cost
The report also emphasizes the environmental difference between text and multimedia generation. Creating an AI-generated image can consume up to 60 times more energy than a standard text query, while complex video generation can require thousands of times more computational power.
Researchers warn that as generative tools become more widely used, total energy demand is likely to increase further unless efficiency improves.
While the study highlights environmental concerns, it also notes that AI remains a valuable technology when used efficiently and purposefully. The key issue, according to researchers, is not usage itself but unnecessary computational load created by inefficient interaction patterns.
