From Polite Prompts To Power Grids: The Growing Debate Over AI's Energy Use

At the NDTV AI Summit, two prominent voices - former UK Prime Minister Rishi Sunak and AI researcher Stuart Russell - addressed growing public questions around artificial intelligence, ranging from electricity consumption to long-term existential risks. Sunak drew attention to the energy demands of AI through a personal anecdote about his daughters' interactions with chatbots. He said they often include polite phrases such as "please" and "thank you" in conversations with AI systems, which he advised against.

"It's polite, but it's not a person. And by the way, it takes up a lot of compute power, so better if you don't," he said, recounting the exchange during a fireside discussion hosted by NDTV.

The comment reflects a broader industry discussion about the computational resources behind generative AI systems. Large language models operate on specialised processors inside data centres, and each interaction requires multiple layers of neural-network computation.

Russell said adding polite language would not meaningfully influence how AI systems behave. He also suggested that the overall electricity use of AI should be viewed in context relative to other sectors of the economy.

Evidence from industry research and academic studies provides insight into both points.

Studies comparing digital services have estimated that a single generative AI query can consume several times more electricity than a conventional web search, largely because search engines retrieve indexed information while AI systems generate responses through intensive computation. Academic research has also found that factors such as prompt length, output size, hardware configuration and model architecture significantly affect energy consumption per query.

At a system level, the growth in AI usage is contributing to rising electricity demand from data centres. News agency Reuters, citing research from the Electric Power Research Institute, said data centres could account for up to 9% of US electricity consumption by 2030, more than double current levels. Separate analysis supported by the US Department of Energy has projected that data-centre power demand could nearly triple within a few years, driven partly by AI workloads.

Bloomberg has reported similar trends globally, noting that AI infrastructure expansion is accelerating electricity demand growth and could require substantial new power generation capacity in multiple regions.

At the same time, researchers emphasise that the energy footprint varies widely depending on efficiency improvements, hardware upgrades and deployment methods. Optimisation techniques can significantly reduce power consumption without reducing performance.

Beyond energy use, Russell addressed concerns about advanced AI systems surpassing human capabilities. He said the stated goal of several major technology companies is to build machines more intelligent than humans, which raises questions about long-term control and safety.

He referenced warnings dating back to Alan Turing, who suggested that machines could eventually exceed human intellectual abilities.

Russell also said some AI industry leaders have acknowledged potential large-scale risks, including severe societal disruption. Researchers often describe such scenarios using terms like artificial general intelligence or superintelligence - systems that do not yet exist but are the subject of ongoing debate in policy and scientific communities.

Current AI systems, however, operate within defined computational frameworks and do not possess consciousness or independent intent.

The discussions at the summit reflected two parallel realities: AI systems require significant computing infrastructure that is expanding rapidly, while questions about the long-term trajectory of increasingly capable AI remain under active examination by scientists, governments and industry.