Why Researchers Recommend Ditching AI Prompts to Reduce Energy Consumption

United Nations researchers are advocating for a reduction in politeness towards artificial intelligence (AI) to lower energy consumption.

A study from the United Nations University Institute for Water, Environment and Health (UNU-INWEH) reveals that omitting polite phrases like “please” and “thank you” from AI prompts can decrease ChatGPT’s energy usage by 25%, saving between 87 and 98 gigawatt-hours of electricity annually. This is comparable to the yearly electricity consumption of up to 760,000 households in sub-Saharan Africa.

To curb energy consumption and minimize carbon emissions, it’s essential to craft concise prompts, steer clear of conversational loops, and avoid emotional attachments to AI, according to the researchers.

Kave Madani from UNU-INWEH states, “We’re not advocating disrespect towards AI, but it’s crucial to evade the interaction trap and not become overly engaged.” Learn more about Kave Madani.

The AI model behind chatbots processes text in small units known as tokens. Concise prompts save energy by decreasing the number of tokens the model needs to analyze and generate for responses. Shorter prompts can simplify tasks and further reduce power requirements.

The recent United Nations study is one of the most thorough evaluations of AI’s environmental costs. It warns that the rapid rise in AI technology is significantly escalating energy, land, and water use.

Currently, ChatGPT handles around 2.5 billion queries each day, while Google processes 16 billion, most of which include AI-generated summaries.

Due to the lack of disclosures from tech companies regarding energy use, researchers utilized data available from their data centers.

AI accounts for approximately 20% of energy consumption in data centers, with projections indicating this could rise to about 40% in the coming years. By 2030, AI might consume around 378 TWh annually, while total data center consumption could reach 945 TWh, representing nearly 3% of anticipated global electricity demand.

The projected 9.3 trillion liters of water required for data centers by 2030 is sufficient to meet the annual basic domestic water needs of all 1.3 billion residents in sub-Saharan Africa.

As Miriam Axel from UNU-INWEH points out, “Technologies are being adopted globally at an unprecedented rate, leading to a swift rise in energy consumption.” Discover more about Miriam Axel.

The researchers assert that while AI companies should be mandated to disclose their energy usage, and governments should impose energy restrictions on businesses and individuals, public education on efficient AI use is equally critical.

Madani emphasizes the importance of minimizing unnecessary AI usage and suggests using fewer words and less powerful models when engaging with AI.

It’s also noteworthy that generating an image consumes 60 times more energy than a text query, enough to power a 10-watt LED light bulb for about 17 minutes. In contrast, a complex video could use up to 8000 times more energy than text, powering that same light bulb for approximately 1.7 days.

“AI isn’t inherently bad,” Madani concludes. “We should just use it wisely. It’s like a knife: it can save someone’s life in surgery, or it can be misused.”