TLDR: A new report by UNESCO and University College London (UCL) reveals that artificial intelligence’s escalating energy demands can be drastically reduced by up to 90% without compromising performance. The key strategies involve using shorter, more concise prompts and shifting towards smaller, specialized AI models, alongside techniques like model compression. This comes as generative AI’s energy footprint is rapidly expanding, placing significant strain on global resources.
Geneva, Switzerland – A groundbreaking study released by UNESCO in collaboration with University College London (UCL) at the AI for Good global summit in Geneva has identified practical methods to cut the energy consumption of artificial intelligence by as much as 90%. The report, published on Tuesday, July 9, 2025, highlights that adopting concise prompts and utilizing specialized, lightweight AI models can lead to substantial energy savings without any degradation in the quality of AI responses.
The findings emerge amidst escalating concerns over the environmental impact of generative AI tools such as OpenAI’s ChatGPT, Google’s Gemini, and Microsoft’s Copilot, all of which demand immense computational power. OpenAI CEO Sam Altman recently disclosed that each ChatGPT request consumes an average of 0.34 watt-hours of electricity, a figure 10 to 70 times higher than a single Google search. With ChatGPT alone processing approximately one billion prompts daily, its annual electricity usage reaches an staggering 310 gigawatt-hours, equivalent to the yearly consumption of three million people in Ethiopia.
UNESCO has issued a stark warning, stating that ‘The exponential growth in computational power needed to run these models is placing increasing strain on global energy systems, water resources, and critical minerals.’ The organization further noted that the energy demand from generative AI is doubling roughly every 100 days, underscoring the urgency of sustainable practices.
The study, based on experiments conducted by UCL computer scientists on various open-source Large Language Models (LLMs), presents several effective remedies. One primary strategy involves the use of shorter, more efficient prompts. Tests demonstrated that reducing a prompt from 300 to 150 words, combined with switching from a large general-purpose AI model to a smaller, domain-specific one, resulted in nearly 90% energy savings without any loss in performance. The report also advocates for ‘mixture of experts’ systems, where only necessary specialist models are activated for a given task.
Another significant technique identified is model compression, including methods like quantisation, which can reduce energy consumption by up to 44% while maintaining performance. These combined approaches offer a pathway to significantly mitigate AI’s environmental footprint.
Encouragingly, major tech companies appear to be responding to these concerns. Firms like Google have introduced leaner models such as ‘Gemma,’ Microsoft offers its Phi-3 series, and OpenAI recently launched GPT-4o mini, all miniaturized versions of their language models designed for greater efficiency.
Tawfik Jelassi, UNESCO’s Assistant Director-General for Communication and Information, emphasized the need for a fundamental shift: ‘Generative AI’s annual energy footprint is already equivalent to that of a low-income country, and it is growing exponentially. To make AI more sustainable, we need a paradigm shift in how we use it, and we must educate consumers about what they can do to reduce their environmental impact.’ UNESCO, which saw its 194 Member States adopt the Recommendation on the Ethics of AI in 2021 (including guidance on environmental impact), is now urging governments and industry to invest in sustainable AI research and development, alongside promoting AI literacy to empower users to make informed decisions about their AI usage.
Also Read:
- MIT Researchers Uncover Significant Environmental Footprint of Generative AI
- Generative AI: Empowering Consumers to Reduce Their Carbon Footprint in the Telecom Sector and Beyond
The report also highlights broader implications, particularly for equitable access to AI. Currently, most AI infrastructure is concentrated in high-income countries, with only 5% of Africa’s AI workforce having access to the computing power required for generative AI tools. Promoting energy-efficient AI could also help bridge this digital divide.
