UNESCO-UCL Report Urges Global Shift Toward Smaller, Energy-Efficient AI Models

A groundbreaking new report jointly released by UNESCO and University College London (UCL) has revealed that simple adjustments in how large language models (LLMs) are designed and deployed can lead to dramatic reductions in energy consumption—up to 90%—without sacrificing performance. The findings call into question the continued use of resource-intensive general-purpose AI models and advocate for a transition to leaner, task-specific systems that are more efficient, ethical, and sustainable.

Titled “Smarter, Smaller, Stronger: Resource-Efficient AI and the Future of Digital Transformation”, the report offers concrete, accessible strategies to reduce AI’s environmental footprint, enhance energy equity, and foster more inclusive digital transformation—especially in low-resource regions.

A Call to Action for Governments and Industry

UNESCO, whose mandate is to support the digital transformation of its 194 Member States, has long warned about the environmental and ethical risks associated with unchecked AI growth. This latest research reinforces those concerns and offers solutions that are not only technically viable but also urgently needed.

“This is a pivotal moment,” the report states. “We must prioritize energy-efficient AI models to protect ecosystems, address digital inequality, and build AI systems that serve everyone—not just the wealthy and well-resourced.”

The report calls on governments, private companies, and AI developers to invest in sustainable AI research and development, develop AI literacy programs, and adopt ethical design frameworks like the 2021 UNESCO Recommendation on the Ethics of AI, which includes environmental accountability as a core principle.

The Hidden Energy Toll of AI Interactions

With over 1 billion people using generative AI tools every day, each seemingly small prompt carries a significant environmental cost. The report estimates:

• Each prompt consumes 0.34 watt-hours of electricity

• This amounts to 310 gigawatt-hours per year globally

• Equivalent to the annual electricity use of over 3 million people in a low-income African country

Such figures underscore the importance of making AI energy use more efficient—especially as adoption continues to accelerate.

Three Breakthrough Techniques for Efficiency Without Sacrifice

UCL researchers conducted a series of original experiments on open-source LLMs, and identified three key innovations that drastically reduce energy consumption without compromising model accuracy:

1. Task-Specific Smaller Models

Instead of using a massive general-purpose model for every task, developers can deploy smaller models fine-tuned for specific functions—such as summarization, translation, or question-answering. This tailored approach can cut energy use by up to 90% and reduce computational overhead.

2. Mixture of Experts (MoE) Design

The “mixture of experts” model is a dynamic framework in which only the relevant smaller expert model is activated on-demand. For example, a translation expert is called for only when translation is needed. This minimizes the energy costs of loading large, unnecessary systems for each query.

3. Model Compression Techniques

Compression methods like quantization, distillation, and pruning can reduce energy usage by up to 44%, without a drop in performance. These techniques shrink model size and complexity, allowing them to run on less powerful hardware and at lower energy costs.

In addition, the report highlights the benefits of shorter, more concise prompts and responses, which can reduce energy usage by more than 50%—a simple yet overlooked strategy that users can adopt immediately.

Advancing AI Access and Equity

The report also emphasizes that smaller, resource-efficient AI models are more accessible in low-income and low-infrastructure environments. According to the International Telecommunication Union (ITU), only 5% of Africa’s AI talent has access to the computing power required to train or use modern generative AI tools.

By promoting compact models and efficient techniques, UNESCO and UCL argue, the world can democratize AI access, especially in areas where energy, water, and connectivity are limited.

“These findings open the door to a more just and equitable AI landscape,” said one of the report’s lead authors. “This is not only about energy—it’s about who gets to participate in the digital future.”

Toward a Sustainable Digital Transformation

The report situates these technical innovations within the broader context of climate change, environmental justice, and global digital development. It urges policymakers to:

• Mandate energy transparency in AI model reporting

• Support open-source small model development

• Integrate environmental sustainability into AI procurement and deployment policies

• Fund international collaborations to ensure equitable access to efficient AI systems

As UNESCO continues to advocate for ethical AI governance, this report sends a clear message: sustainability and performance are not mutually exclusive. By rethinking design choices and empowering users, the world can build a smarter, more inclusive digital future—without burning out the planet.