Mr. Seán O'Sullivan:

I think it will happen broadly as we train these systems. For example, when I was evaluating current AI content-detection systems, I found there was a huge accuracy drop-off between OpenAI's GPT-3 and GPT-4 models, simply because the latter was more capable. We can say these models are just predicting the next token but, in doing that, they are also being optimised to mimic human writing to a more and more believable level. This was point-in-time research showing that existing systems, even when provided with past examples of a person's writing, are unable to mimic their style perfectly such that an author-verification system can differentiate. We need to explore what it means to use these tools in writing speeches, for example. If there were a version of GPT-4o that had been finely tuned on every speech the Deputy had ever written, there could be a time in the future when it was able to mimic her style exactly, so that raises questions as to the ethics surrounding that. These all need to be discussed.