Experts Recommend Ditching AI Prompts to Reduce Energy Consumption

Researchers from the United Nations are advising people to adopt a less polite approach towards artificial intelligence. A new report indicates that trimming excess words from prompts can reduce ChatGPT’s energy consumption by as much as 25%.

According to the report by the United Nations University Institute for Water, Environment and Health (UNU-INWEH), omitting unnecessary terms like “please” and “thank you” from AI prompts could save between 87 and 98 gigawatt hours of electricity each year. This amount could power approximately 760,000 households in sub-Saharan Africa.

To combat energy use and lower carbon emissions, users are encouraged to craft concise prompts, avoid conversational loops, and refrain from forming unnecessary relationships with AI, the study suggests.

“While I’m not advocating disrespect, it’s important to steer clear of interaction traps and not to get absorbed in extended conversations,” states Kave Madani from UNU-INWEH.

The language model driving AI chatbots processes information in small units called tokens. Madani explains that shorter prompts save energy because they decrease both token processing and generation. In many cases, more concise prompts simplify the task at hand, further cutting down power requirements.

This comprehensive assessment of AI’s environmental impact warns that increased usage is leading to higher energy, land, and water consumption.

Currently, ChatGPT processes around 2.5 billion daily queries, while Google handles approximately 16 billion, many of which entail integrated AI summaries.

Due to a lack of transparency regarding energy usage among tech companies, researchers relied on data from their own data centers for this analysis.

AI is estimated to account for about 20% of energy used in data centers, a figure expected to double to around 40% in the coming years. By 2030, AI alone could consume about 378 TWh annually, with data centers using a total of 945 TWh, nearly 3% of expected global electricity demands.

The projected need of 9.3 trillion liters of water for data centers by 2030 could meet the essential domestic water needs of 1.3 billion people in sub-Saharan Africa.

“We’re seeing technologies being adopted on a global scale quicker than ever before, leading to a rapid rise in energy consumption,” says Miriam Axel from UNU-INWEH.

The researchers emphasize the need for AI companies to disclose their energy usage and urge governments to set energy limits for corporations and individuals alike. Additionally, educating the public on efficient AI use is crucial.

Madani encourages users to minimize unnecessary AI interactions and use shorter prompts with less complex models when required.

It’s also notable that generating images requires 60 times more energy than text queries, enough to power a 10-watt LED bulb for about 17 minutes. Complex videos demand up to 8000 times more energy than text, capable of sustaining the same light bulb for nearly 1.7 days.

“We’re not suggesting that AI is harmful,” Madani remarked. “We’re advocating for responsible use. It’s akin to a knife: it can save lives in surgery, but it can also cause harm.”