Mr. Barry Lowry:

It does no harm to take evidence from all the stakeholders in any ecosystem. I refer not only to big tech but also representatives of our SMEs, where some of the most innovative ideas are coming from. AI is a very broad church. For example, robotic process automation is seen to be part of the AI family. There are numerous examples across government where that is being used successfully. For example, the National Shared Services Office, NSSO, is using it. That avoids the need for people to literally read stuff from one screen and type it into another. That is very successful.

We are seeing successful use of AI in areas such as training it to identify tumours, health risks and so on. Unlike the clinician, AI does not get tired. Therefore, it will provide a tremendous source of advice to the clinician that someone could be at risk in a particular area. AI is only part of what a clinician does. The clinician will still do the physical operations, ask questions, take bloods and so on. AI is valuable here and the track record is good.

The reason the overall percentage is so high is due to GenAI or large language models - the ChatGPTs of this world and so on. Unlike IT systems, where we will talk about agile and minimum viable product, the investor is spending all the money upfront on AI and it might not be used. Building a chatbot requires training that chatbot in every single process it is going to answer a question on. However, if the public do not like it or use it, then the investor will have spent all that money and will not get a return on the investment. In that case, if the original plan was to divert a large number of people to the chatbot so that the telephone support could be used by the most vulnerable or the people who cannot follow chatbots so well, they are not going to achieve that objective. That is why we have given the advice on the due care and testing that are required. Are you building something people want as opposed to something you think is going to be a good idea?

There is a lot of marketing around the LLMs at the moment, but they will find their homes. For example, I mentioned three proofs of concept that we carried out in OGCIO as part of informing our advice node. The first was a chatbot. The second was about how to curate structured information, such as existing reports, to make it easier for people to do research. The third, which was related to that, was on structured information, such as emails and Word documents. We have addressed those two requirements with Copilot and reduced the costs dramatically because someone using Copilot will be curating that data there and then. For example, there was an OECD report on AI, which was 145 pages long. I put it into Copilot, asked it some questions and it pointed out the bits I should focus my reading on. It is areas like that where we will not hear of dramatic failures, but there are huge successes even in our department where we are using this. We are not using it for public services. We are using it as our own research assistant, if you like. It is highly beneficial.