UNESCO and UCL Report Reveals Energy-Saving Strategies for AI Models

A recent report published by the United Nations Educational, Scientific and Cultural Organization (UNESCO) and University College London (UCL) highlights that minor adjustments in the design and utilization of Large Language Models (LLMs) can lead to significant reductions in energy consumption—potentially up to 90%—without sacrificing performance. The findings underscore a crucial pivot towards more sustainable AI practices amidst growing concerns over the environmental impact of artificial intelligence technologies.

The report, released on July 8, 2025, emphasizes that generative AI's energy consumption has reached alarming levels, comparable to that of a low-income country. Tawfik Jelassi, the Assistant Director-General for Communication and Information at UNESCO, stated, "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's mandate, in collaboration with its 194 Member States, aims to support digital transformations that prioritize ethical and energy-efficient AI policies. In 2021, these Member States adopted the UNESCO Recommendation on the Ethics of AI, which encompasses guidelines on mitigating AI's environmental footprint. The report advocates for increased investments in sustainable AI research and development and emphasizes the importance of AI literacy to empower users to make informed decisions about their technology use.

The research team at UCL conducted original experiments on various open-source LLMs, identifying three key innovations that can lead to substantial energy savings:

1. **Utilization of Smaller Models**: The report found that smaller, task-specific models are just as effective as larger, general-purpose models, potentially reducing energy consumption by up to 90%. Current practices often rely on expansive models for diverse tasks, but tailored smaller models can enhance efficiency.

2. **Mixture of Experts Model**: This approach activates specific smaller models only when required to complete a designated task, thereby conserving energy by avoiding unnecessary computations.

3. **Concise Prompts and Responses**: By optimizing the length of prompts and responses, users can decrease energy consumption by over 50%. Additionally, model-compression techniques can yield energy savings of up to 44% while retaining accuracy.

The implications of these findings extend beyond just energy savings. The accessibility of AI technologies varies dramatically across different regions, particularly in low-resource settings where energy and internet connectivity are limited. According to the International Telecommunication Union (ITU), only 5% of Africa's AI talent has access to the necessary computing power for developing or utilizing generative AI. This disparity deepens global inequalities; thus, the proposed energy-efficient strategies are crucial for democratizing access to AI technologies.

In conclusion, the UNESCO and UCL report presents a vital call to action for governments, industries, and consumers to advocate for sustainable AI practices. By implementing these minor changes, stakeholders can contribute to a more environmentally conscious approach to AI, ensuring that advancements in technology do not come at the expense of our planet. As the landscape of AI continues to evolve, the integration of sustainable practices will be essential for fostering equitable development in the digital age.