Cambridge Language Sciences—written evidence (LLM0053)

House of Lords Communications and Digital Select Committee inquiry: Large language models

I am Dr Guy Emerson, writing on behalf of Cambridge Language Sciences, an Interdisciplinary Research Centre at the University of Cambridge, where I am Executive Director. An interdisciplinary perspective is important for a full understanding of LLMs, including not only their technical functionality, but also their social context.

1. How will large language models develop over the next three years?

1.1. Usage of ChatGPT and other LLM-powered chatbots is already falling, as the novelty wears off and users become aware of the limitations of the technology.[1] The opportunities of LLMs should not be taken for granted, but rather critically evaluated.

1.2. The consensus among experts is that the progress in LLMs in the past year (exemplified by ChatGPT) has come from increased scale, rather than new methods. The details of proprietary systems such as ChatGPT are not known, but large open-source competitors such as BLOOMZ,[2] where the architecture and training data are openly documented,[3] achieve similar or better performance across a wide range of tasks.[4] Given the amount of data and computational resources already used, there is limited scope for further increases in scale in the next three years. Most training data was scraped from the web, and the easy sources of data have already been exhausted. Furthermore, collecting more web data now runs the risk of contamination from LLM-generated text. Automatically detecting generated text is not reliable,[5] and training an LLM on generated text has been shown to degrade performance.[6]

1.3. However, even in the absence of major technical developments, we can still expect many attempts to deploy the technology in the next three years, and this is where regulatory attention will be needed. It is important to view LLMs in a wider context, since next-word prediction is rarely a useful task for an end user. Usually, an LLM is generally deployed within a larger system, such as a chatbot (as in the case of GPT deployed within ChatGPT). LLMs are already beginning to be integrated into digital infrastructure, and it is essential to critically assess who benefits and who is harmed by such developments.

2. What are the greatest opportunities and risks over the next three years?

2.1. Many harms of LLMs are already present. In particular, misinformation is easier to spread than ever before,[7] and there is no reliable way to detect LLM-generated text.[8] The amount of generated text has been compared by Prof. Emily M. Bender to an oil spill in our information ecosystem:[9] "The reason I make the analogy to oil spills is that this isn’t just about the harms to the person who initially receives the information. There are systemic risks as well: the more polluted our information ecosystem becomes with synthetic text, the harder it will be to find trustworthy sources of information and the harder it will be to trust them when we’ve found them." Mitigating such risks is an essential task for the next three years, including to make sure that LLM-generated misinformation does not undermine democratic processes.

3. How adequately does the AI White Paper (alongside other Government policy) deal with large language models? Is a tailored regulatory approach needed?

a) What are the implications of open-source models proliferating?

3.1. The five principles in the AI White Paper are important, but some of them may be difficult to evaluate for LLMs, because of their wide range of potential uses. For example, to measure the safety of an LLM, what uses should be considered? However, transparency is an important principle which can be applied more straightforwardly to LLMs. Furthermore, proprietary LLMs fare badly in terms of transparency,[10] and this will need to be improved in order to comply with the EU AI Act.[11]

3.2. Specific uses of LLMs can be clearly regulated, and this does not necessarily require legislation specific to LLMs. The AI White Paper points out several examples where existing legislation covers uses of AI, including laws on discrimination, data protection, and product safety. As another example, the UK already has robust laws on defamation, which can and should be applied in cases of defamation involving LLM-generated text. Providers and users of the technology need to be held accountable.

3.3. Open-source LLMs should be welcomed, because the proliferation of models makes the need for transparent documentation even more important. If poor documentation becomes the norm, it will be easy for bad actors to release models with hidden "backdoors", such as generating misinformation on specific topics.[12]

5 September 2023