What is a foundation model?

Foundation models can also be​‘fine-tuned’ to complete specific tasks.

Foundation models can write text, produce images and create sounds that resemble books, artworks or music produced by humans. More recent foundation models have been trained to predict a wide range of biological sequences around proteins, DNA and RNA, carry out multi-step ​‘reasoning’, and operate computer interfaces., undertake multi-step ​‘reasoning’, and operate computer interfaces.

Notable examples are OpenAI’s GPT models – which power the conversational chatbot ChatGPT, Microsoft’s Copilot, and Google’s Gemini – which produces the AI summaries that appear in Google search results.

It is difficult to make laws that can control and regulate the development and deployment of foundation models. Applications built on foundation models can involve many different companies at different stages of development and use. 

These include the foundation model developer, the application developer and the hosting company. Because of this, if there are faults in the application, or if the application causes harm, it can be hard to identify who is responsible.

The wide range of tasks and applications that foundation models can do, and the fact that their ways of working are hard to predict and understand, make it difficult to test whether they are safe. Similar issues have come up in other industries, such pharmaceuticals and chemicals. Looking at the solutions found in those contexts may help improve AI safety.

In the UK, the AI Security Institute has been conducting voluntary testing of the most advanced foundation models, and the government is experimenting with the use of foundation models across the public sector.