Advancing Machine Unlearning for Sovereign LLMs in the UK

London, UK: Locai Labs Limited, the UK’s foundational AI company, is proud to announce a new research project in machine unlearning supported by the awarding of 20,000 GPU compute hours on the Isambard-AI AI Research Resource (AIRR) service by UKRI. This project, titled "Advancing Machine Unlearning for Sovereign LLMs in the UK", addresses one of the most pressing challenges in modern AI development; how to ensure that powerful open-source AI models can respect data privacy and copyright, comply with regulations like GDPR, and respond to legitimate requests for data removal without the need for costly and unsustainable full retraining of the AI model.

Compute is the engine of AI development, and access to Isambard-AI, the UK’s sovereign high-performance AI research infrastructure, will enable Locai Labs to progress this research.

The Challenge: building affordable and safe UK sovereign AI

Open-source LLMs have made remarkable progress in the past few years, now rivalling proprietary frontier models in performance. However, their training data is often opaque, and potentially includes personal, sensitive, or copyrighted material. While open-source models fuel innovation, they may introduce risks, for example when developers build downstream applications without full knowledge of what their base model has learned during training.

Under data protection laws such as the UK’s GDPR, individuals have a "right to be forgotten". But how can this right be upheld when a model has ingested data it shouldn’t have and there’s no practical way to remove it?

George Drayson, Chief AI Officer, said: “Retraining a model with frontier level performance from scratch is prohibitively expensive, as is building new frontier models. This is why the UK currently uses frontier models developed by the leading AI labs in the USA and China and has no UK-built sovereign foundational model at present. The development of machine unlearning provides the UK with the potential to create safe, UK sovereign AI models with frontier level performance at a fraction of the cost in compute, capital and energy.”

What Is Machine Unlearning?

Machine unlearning is the process of selectively removing the influence of specific data from a trained AI model. Rather than retraining from scratch, the goal is to "forget" certain information through targeted updates, ensuring the model behaves as if it never saw that data, while preserving its overall capabilities.

This is no small task. Knowledge in LLMs is highly entangled; removing one piece can inadvertently affect others, leading to incomplete forgetting or performance degradation. Current methods show promise but lack the precision and scalability needed.

Our vision: safe, affordable UK sovereign AI

At Locai Labs, we develop and post-train open-source large language models. Our first model Locai L1-Large is a 235 billion parameter general model that we use to power our consumer chat application GB1.ai. Locai L1-Large was developed by applying proprietary techniques to improve the performance of the Qwen-3 open-source model developed by Alibaba and to increase its knowledge of UK languages, norms and idioms. We have developed techniques to reduce bias and censorship in open-source models and this new research will develop new techniques to unlearn problematic content; whether due to privacy concerns, copyright claims, or safety risks. This will allow us to deploy safe sovereign AI for the UK without having to build new frontier sovereign AI models from scratch – a prohibitively expensive undertaking. 

What will our research cover?

During this three-month project we will:

·Evaluate existing unlearning methods using established benchmarks such as Tofu and MUSE

·Develop novel techniques that improve forgetting precision, model utility, and safety retention

·Conduct large-scale experiments on post-trained open models

·Share all findings, code, and models openly to support the broader UK AI research community

Editors Notes:

About Locai Labs Limited

Locai Labs is the UK’s foundational AI company. Based in London, the company has developed a proprietary LLM post-training methodology; Forget-Me-Not which it has used to create Locai-L1-Large, the UK’s first foundational large language model, trained using 100% renewable energy. 

For more information, visit www.locailabs.com