King’s Speech

Lord Tarassenko Excerpts

Tuesday 19th May 2026 Lords Chamber

[Lord Tarassenko (CB)]

My Lords, it is always a pleasure to follow maiden speeches, and I congratulate the noble Lords, Lord Hobby and Lord Blackwater, on their distinctive and excellent speeches.

There was no AI Bill in the gracious Speech, but AI will play a major role in NHS modernisation and education for all. Here, I declare an interest as a senior adviser to the Alan Turing Institute, the national AI institute. Over the past six months, the institute has argued that the time has come for the UK to train its own sovereign large-language model. At present, the choice is between closed-source models, such as ChatGPT or Gemini, from US big tech companies or open-weight models, mostly from China. Neither type of model can be fully trusted as they may be poisoned with sleeper agents or may harbour back doors. For uses of national importance—and I argue that these include education and health care, as well as defence and national security—the only way to have 100% security when using an LLM is to have complete knowledge of both the training data and the model’s weights. Other countries, such as France, Switzerland, the UAE and South Korea, keen to reduce their dependence on the US and China, have announced their own sovereign LLMs in the past 12 months.

There are further advantages for the UK in going down this route. First, a sovereign LLM will be trained using UK values with full transparency, respect for the law on copyright and the possibility of creating a process to remunerate training data providers. This would help to end the standoff between the UK’s growing AI sector and its world-leading creative industries. Secondly, the UK has unique datasets in both its health service and its school education system, including BBC resources, that could and should be treated as sovereign data assets prioritised for use in training the country’s sovereign LLM.

Is it affordable when it costs US big tech companies tens if not hundreds of billions of dollars to train their frontier LLMs? In contrast, the Alan Turing Institute, supported by evidence from the Swiss model, estimates that it would cost only between £5 million and £10 million pounds to train the UK’s sovereign LLM. How can that be? US big tech companies do raise hundreds of billions of dollars to train the next general-purpose frontier model, but the key lesson from DeepSeek 18 months ago was that frontier-level performance on application-specific tasks can be achieved with much simpler models through distillation and fine-tuning.

The UK sovereign LLM, distilled and fine-tuned on NHS data, will be able to give healthcare advice as accurate as, if not more accurate than, GPT-5. It will not be able to translate Hamlet’s soliloquy into Arabic, as a general-purpose LLM such as GPT-5 can, but that is of no consequence to an NHS doctor or a patient seeking a second opinion.

With the Health Data Research Service due to launch its first services by the end of this year and with AI tutoring tools now being developed for secondary schools, we urgently need a UK sovereign LLM. This would enable fully trusted, application-specific models fine-tuned on our unique NHS and education datasets. The Government launched a sovereign AI unit last month; it should now support the development of sovereign AI models.