17
Examination of witness
Professor Dame Angela McLean.
Q112 The Chair: This is the Communications and Digital Committee. We are continuing our inquiry into large language models. This is our final public hearing. We have two witnesses today. The first is Professor Dame Angela McLean, the Government’s Chief Scientific Adviser. Later, we will be joined by Viscount Camrose, the Minister responsible for AI. Clearly, with Dame Angela, we are looking to ask questions of the leading expert advising the Government on the various aspects of large language models, on the Government’s proposed policy and approach to it, and on the way in which they intend to take this whole area forward. We will be talking to the Minister later about the policy choices and decisions that the Government are making. Let us get going straightaway. I turn to Lord Kamall to kick us off with the first question.
Q113 Lord Kamall: Thank you, Dame Angela, for appearing before us today. I want to start by talking about some conversations that I have had with some of my friends who live in European Union countries, who, when they look at technology, often feel that the precautionary principle is given more consideration than the innovation principle. In looking at the UK’s emerging policy on large language models and AI more generally, we have heard from witnesses that the UK’s policy seems to have pivoted away from innovation and is now heavily focused on risk. Do you agree with that comment? Does there need to be a greater focus on innovation, or have we got the balance right?
Professor Dame Angela McLean: I can see where that comment comes from, because we—certainly the bit of government that I sit in—spent so much of the summer preparing for, running and, in the case of some of us, attending this thing called the AI Safety Summit. As part of that summit, the Government announced the setting up of this thing called the AI Safety Institute. Let me say that my colleagues in DSIT and the Foreign Office did a terrific job; they ran a summit that we can all be proud of.
However, I do not think that shows that we have stopped paying attention to other things. Of course, you are right, and we need to keep three things in balance: short-term safety worries, by which I mean safety issues like bias and misinformation and disinformation, all of which we live with already; longer-term safety issues, including the sorts of things that were discussed more at the safety summit and have been discussed rather widely over the past 12 months; and the massive opportunities. There is plenty of focus, and plenty of us in government are very keen to keep those three things in balance.
Q114 Lord Kamall: Thank you for that. When you think about the UK’s role in the world and in new technologies—AI, large language models, et cetera—what do you think is needed when you look at the current landscape of the UK for us to compete internationally? I am thinking of things such as advanced computing facilities. What do you see from your perspective? What do you think we should be doing more of and who should be doing it? Should it be the Government, the private sector or others?
Professor Dame Angela McLean: First, I see more compute; I think there is more compute on the way. We are already in a very good position in having well-curated data pertinent to the questions in hand. That is an incredibly important asset and one that we need to make sure we make good use of.
What else? Good and wise regulation. That reminds me: we have a whole thing called the science and technology framework, which I have brought with me. It was written by my predecessor, Patrick Vallance, so I can laud it. It is all about exactly this set of questions. If you have a big science and technology advance, what do you do so that you can use it to drive prosperity, security, terrific jobs—all the things you want for your country? I will not go on about it for too long, but if you do not have a copy, take one of my cards. It will take you to where the framework is on the internet. It is 30 sides long and is about what we used to call the 10 big things. It is a great long list of all the things you need to do in order to drive commercial and public sector advances when you have an exciting new technology at your fingertips.
Lord Kamall: When you look at those 10 things—I admit that I have not looked at them yet—is anything missing, or potentially missing, in order for the UK to play a major role in this?
Professor Dame Angela McLean: It is always skills—at least, skills is the first one. The whole point of the framework is that it says, “It is never just one thing”. We have sometimes struggled to generate wealth from the brilliant inventions we make because of systems failure. There are lots of things that we do not get quite right. Nevertheless, I would go back to my immediate answer: with the very rapid advance in platform technology—a technology that you can use for lots of things—we all, not just children, need to up our skills.
Lord Kamall: When you talk about skills, is that domestic plus immigration?
Professor Dame Angela McLean: It is about making sure that our kids are properly skilled, that we reskill, and that we are a tremendously attractive place for the most able people in the world to come and work.
Lord Kamall: I know from speaking to my eldest son—he is 22—that a lot of his friends have done different degrees. Many of them are excited by AI and want to go into data science, but most of the jobs advertised are looking not for entry-level graduates in STEM but for people who have some experience. How do we tackle that gap?
Professor Dame Angela McLean: That is a really interesting question. We used to talk about this a lot, actually. I used to work in the MoD and we used to talk about this being a potential way for it to grow its own skills resource and contribute to the country’s skills resource. Basically, we as government need to accept that we need to get a load of people in and train them up in doing this stuff, and that many of them will leave. That is okay, actually, because we need a country that is awash with this set of skills—it is not that expensive—and some of them will stay; we have fantastic questions in government, and we need to set things up so that government is a marvellous place to work if you are skilled at data science. As I mentioned earlier, we have rich data assets, and that is rare if you are a data scientist asking, “Can I get at the data that will let me answer the question I want to answer?”
It is not just us. We need to ask industry to play a role in imparting the skills that are needed. There is a whole range of skills. We do not need everybody to be able to write brilliant new algorithms. We need people to learn about this new skill called prompting and to get really good at that. For me, the most interesting question concerns talking to a person with a problem and thinking about how to turn it into an issue where I can use data science and AI to help.
Lord Kamall: Can I push you a bit on that data skills gap? Lots of young people want to get into it, but do not have experience when all the jobs ask for experience. You are saying that government should step up there, I suspect. How do you convince government and the private sector, because they do not like paying for training and they are losing people?
Professor Dame Angela McLean: That is a great question. The Government take in a lot of graduates. In my own bit of government, we have a new graduate scheme, which is a terrific way for us to access very skilled and able people. We lay on training specifically for them. As I am sure you know, the fast stream is our premier entry route into the Civil Service. I honestly think that we should be expanding all those things and acknowledging that many of those people will leave. One of the four things that I have always said in my post is that we clearly need a more scientific Civil Service.
Q115 Lord Kamall: Generally, from a government perspective, there is lots of public and government data out there. How do we make sure that as much of that data as possible is available to large language models for training?
Professor Dame Angela McLean: That is a great question. I know we have a thing called the Open Data Institute, which was established by Mike’s colleague and mine, Nigel Shadbolt, and by Tim Berners-Lee. I am pretty sure that that was set up to help with that. I know that people inside DSIT are working right now on making sure that, at the same time as making the data available, we value it properly. Some of our data assets are extraordinarily valuable, and we need to make sure that we do not underprice them when other people come along and do fantastic things with them.
Lord Kamall: There are obviously some concerns about sharing data, such as health data. How do you address those concerns?
Professor Dame Angela McLean: I would refer my health colleagues to specialist advice. The Council for Science and Technology, which is the Prime Minister’s group of science advisers, prepared advice on precisely this question about exploiting public data. That has now become a strand of work inside DSIT to make sure that it is done properly.
Lord Kamall: Thank you. If you have any further thoughts, please send them to our committee staff.
Q116 Lord Lipsey: My colleagues are probably not as ignorant as I am, but what is prompting?
Professor Dame Angela McLean: Prompting is asking questions. You are sat there with your large language model and you ask it a question. I might say to it, “Where was Angela McLean born?” and it will give me the wrong answer, which is not tremendously interesting. There is a skill to how you ask questions that can be learned so that you get more useful answers out of it. You would have heard of the version of large language models that you interact with directly, such as ChatGPT.
Lord Lipsey: Would somebody who is good at prompting also be able to tell when a machine is hallucinating?
Professor Dame Angela McLean: I think not. I might refer you to your specialist adviser on that.
The Chair: You mentioned a moment ago the underpriced nature of some of our valuable public sector data assets. Can you give us an example or two of which data assets you are thinking of?
Professor Dame Angela McLean: Did I say “underpriced” or “not priced carefully enough”?
The Chair: I wrote down “underpriced”.
Professor Dame Angela McLean: Underpricing would be an error. I cannot give you an example of where they have been given away.
The Chair: I interpreted what you said as their value not being sufficiently recognised.
Professor Dame Angela McLean: I share the view with DSIT, which has taken advice on this from Saul Klein, a member of the Council for Science and Technology, that it needs to be careful to avoid underpricing our data assets. I do not have an example off the top of my head, but do you want me to find one and send it to you?
The Chair: Yes, it would be helpful to have an example that illustrates that point. My ears just pricked up and I thought, “Where are we undervaluing something that is clearly a precious asset?” So, yes, it would be great if you could do that. Thank you.
You also referred to systems failure alongside skills that has perhaps led to us not exploiting some of the opportunities that emerge from our research. Other than skills, is there another aspect to the systems failure? Are we learning from the past with large language models so that we avoid missing out in this context?
Professor Dame Angela McLean: Absolutely, we are. It would be worth sitting down and thinking about the S&T framework, which was published in March last year. I will give you a few of them; I am not going to read out all 10 of the big things. Thing 3 is to invest in R&D. Thing 4 is skills. Thing 5 is spending money for little S&T companies. Thing 6 is one of my favourites; it is using government procurement to encourage innovation. It goes on. There is a list of 10 things, as I said. It would be super-interesting to sit down and think about what the S&T framework means for how this country is going to generate prosperity and security using LLMs.
The Chair: Procurement has certainly raised with us before as an opportunity that we do not exploit sufficiently.
The Lord Bishop of Leeds: I heard you say that we need a more scientific Civil Service. What would it look like, and would it be at the expense of something else?
Professor Dame Angela McLean: I would say not. A real strength of our Civil Service is the idea of the generalist. We have people who run the Civil Service and could write you a single page on almost anything, but, at the moment, very few members of our senior Civil Service, most of whom work as generalists, come from what I call a deep science and technology background and have worked for, say, 10 years in science and technology. In my head, that could have been in industry, universities, or one of our government labs.
We need to create a pathway so that some of those people can come in, learn how to be a civil servant, which is a skill of its own, and join the senior Civil Service as generalists but generalists with a deep experience and expertise as scientists. Let me give you an example to explain why. I want people who can just look at an experiment and say, “That will not do”. That may be completely obvious to people who have spent parts of their lives working as scientists. It happened to me the other day. I will not tell you where, but somebody showed me an experiment and I said, “There are no replicates in that experiment. Every treatment is different, so you are not able to interpret the results”. That was not because I knew about that bit of science, but because I have spent a lifetime teaching science to 21 year-olds.
Q117 Baroness Fraser of Craigmaddie: Dame Angela, thank you for coming today. The consensus of our witnesses has been that large language models will make things easier, better, quicker and more exciting, but a variety of witnesses have also talked about risks. Can you give us an understanding of what the Government know about risks? How robust is their understanding of risks, and can you articulate the short-term immediate versus the longer-term potentially more existential risks?
Professor Dame Angela McLean: Rarely do I find myself able to say, “I think we’re in a good place at the moment”. That’s because we were so busy on this over the autumn, so we have an articulation of what the risks are, which might be called a sort of call to arms. That paper was written in the run-up to the safety summit, and I commend it to you.
Even further, I commend its appendix, which was written by my office, the Government Office for Science. Instead of being a call to arms, I would call it a piece of honest brokerage. It really tries to say, “On the one hand, some people think this. On the other hand, some people think that”. Other people have described it as one of the most balanced expositions of this problem that they have seen coming out of government. I am very proud of that piece of work and the people who wrote it, some of whom are sitting behind me.
The Government’s risk register about LLMs was published at the same time. Those three documents are publicly available. Of course, this is a fast-moving field, so things go out of date as quickly as you write them. However, one of the things that happened at the safety summit was the commissioning of a document that will be called, I think, the State of the Science, which will be authored by a man called Yoshua Bengio. It will be an even more up to date explanation of the risks. The countries that attended the summit have been invited to nominate somebody to sit on the board of advisers for that document; I will sit on it for the UK. I will not be a writer but one of the advisers helping to guide what comes out of that.
Is that enough?
Q118 Baroness Fraser of Craigmaddie: I would like to ask a little more. Do you think that, here in the UK, we have credible warning indicators for full risks? Yes, the AI Safety Summit looked at this from an international consensus viewpoint, which is exciting and helpful, and well done to you and your colleagues for bringing together the risks: “These are the risks on the one hand, and these are the risks on the other”. But where are we going to land? Where should we land, and who, for the UK, should be responsible for where that landing strip is?
Professor Dame Angela McLean: What are dimensions of this landing strip?
Baroness Fraser of Craigmaddie: In other areas, whether health or something like that, we have indicators of pandemic, catastrophic loss, threat to life—
The Chair:—weather, and that sort of thing.
Baroness Fraser of Craigmaddie: Yes. For this space, should we have those indicators?
Professor Dame Angela McLean: We cannot have those yet, because we do not have what I would describe as an engineer’s description between here and the catastrophic risk. Pandemics happen to be something that I might know about. If I wanted to tell you what would happen between today and, let us say, yesterday’s swine flu case becoming a pandemic, I could spell that out for you in very specific detail.
We do not have that spelled out for the more catastrophic versions of these risks. That is part of the work of the AI Safety Institute: to make better descriptions of things that might go wrong, and find scientific descriptions of how we would measure that. By “scientific”, I mean things like properly designed experiments, properly described work, and an absolute emphasis on reproducibility. If the rest of us in science do an experiment, write it up and publish it, and someone comes along and says, “I couldn’t reproduce what you did”, that is shameful for us. It is like saying that you really messed up, that you are a bad scientist. At the moment, that is not a tradition of this part of science. One of the things that came out of the “What should scientists do” panel at the safety summit was that this needs to become seriously scientific, and reproducibility is a core of the scientific method. So you have not seen those, I would say, because they do not exist for the more catastrophic end of the risks.
Do we have good ways of measuring things that are nearer term? There is much more on misinformation and disinformation; we are better at measuring those. I do not think we are very good on accuracy of these models, so those are things that we ought to be developing now.
Baroness Fraser of Craigmaddie: I want to paraphrase what you have said. Am I right that the AI Safety Institute will therefore work on developing a framework to address quantitatively new risks? Meanwhile, the current risks, the immediate risks, of large language models are from threat actors or misinformation—that is, things that we know about. Who is responsible for setting out the framework to address those? Is it the Government’s central risk function? Is it regulators? Who is leading?
Professor Dame Angela McLean: I will check with my colleagues, but I am pretty sure that it is the NSC(R)—the National Security Council (Resilience)—people who came up with the risk register that I told you was the third document that is available.[1] That is a risk register in the sense of a bunch of things that could go wrong, how bad they would be, some vague sense of how likely they are. So we have that, and it sits firmly inside Resilience.
On things like what bad actors might do, work that was started as part of the work done in preparation for the safety summit will continue. That happens behind wires.
Q119 Baroness Fraser of Craigmaddie: Do you think that the regulators that we currently have are sufficiently able to understand the risk register and the fast-moving nature of this area?
Professor Dame Angela McLean: That is a great question. This is something that we have thought about quite a lot in the Government Office for Science. It started before I came. We had this set of what we think of as innovation-friendly regulation, but in the process of doing that work it becomes very clear to us that it is hard for regulators to attract and retain the kind of extremely valuable talent that is required if you are to be a regulator in a very fast-moving and prosperity-generating domain like this.
We made some suggestions about what to do about that. One was that there should be freedoms—I might have to choose my words quite carefully: “Regulators should be granted greater flexibility to develop a cadre of technical experts and determine the right pay and conditions to attract talented individuals”. The other was that surely the regulators needed ways to facilitate secondments so that people could move in both directions, both in and out of the regulators, to help us to build what I would call SQEP—suitably qualified and experienced people—inside the regulators.
Baroness Fraser of Craigmaddie: So, in a nutshell, the regulators might need more money and more skills.
Professor Dame Angela McLean: I would agree with that, yes.
The Chair: To clarify, are you talking about people moving in and out of regulators from the regulated industries or from government?
Professor Dame Angela McLean: No, I meant from the regulated industries. We would like the movement to go both ways. You will be a much better regulator if you really understand the industry that you are regulating.
The Chair: That is true, but it also presents different risks to do with regulatory capture and the sufficient independence of the regulator.
Professor Dame Angela McLean: I do not deny that, but the need for qualifications and experience is great enough that we need to think about how to manage those other risks.
The Chair: Again, just so that I can be clear about what you have been saying in response to these questions, is it your view that the Government are in a pretty good place in their approach to the risk management of the near-term risks, but that there is still work to be done on the longer-term risks, which would be done by the AI Safety Institute?
Professor Dame Angela McLean: I would say that we are in a pretty good place on identifying risks, and some of them we are taking very seriously. We are all extremely worried about the risks for next year’s election—and we are not the only ones; the whole world is very worried about what will happen in next year’s election—and we are in the process right now of thinking about what to do about it. We have hints from other areas where people worry about misinformation and disinformation—the two biggest ones are climate change and medicines in general, vaccines in particular. There is stuff that we can learn from those topics about how you can help people to be prepared to be bombarded with misinformation.
I would not say that I am completely happy that we are doing enough about it, no.
The Chair: What do you think should be done that is not being done?
Professor Dame Angela McLean: This is my personal opinion; you had better not take this as a GO-Science opinion. Right now may not be the right time, but I think we need a public information campaign to let people know what form misinformation and disinformation might take in the run-up to our own upcoming election.
Q120 The Lord Bishop of Leeds: The AI Council had an ethicist on the board and the AI Safety Institute does not. Yet much of what you are talking is what I would call functional. It is fascinating stuff, but what is the role of ethics—the question of why we are doing this, and why we say yes to some stuff and no to other stuff.
Professor Dame Angela McLean: That is a really good point. I completely agree. This question has technical elements, but it is fundamentally a socio-technical question. This is about how we collectively wish to use this technology. Shall I go away and suggest to my friends in the AI Safety Institute that they need an ethicist and I know a good one?
The Lord Bishop of Leeds: That would be marvellous.
Professor Dame Angela McLean: If you know a good one, let me know. I think that is a great idea.
Lord Lipsey: I am fascinated by this thing about attracting sufficient capacity to the public regulators and so on in the face of private firms that can pay a lot of money. This is partly about the psychology of the people you might get. Are there lots of technically adept people who think, “I would love to serve the public interest by working for a public regulator”? Some of those people are very good; some are less good, ethically. Are there such people around, or do you have to match—well, you cannot match what Bill Gates gets—what a successful entrepreneur might get in order to attract people to public service at all?
Professor Dame Angela McLean: I know extremely clever, very able people who have worked in the regulation of medicines for most of their lives. They were not paid anything like as much as they would have been paid if they had worked in a drug company. There is no question that that is possible. We do not have an AI regulator at the moment, so I cannot say in relation to that specific issue, but there are other regulators.
This country is renowned for having fantastic regulators. Let us not fall into the trap of thinking that we can have either regulation or innovation. You probably know better than I do that good regulation drives innovation. Clear, rapid, well-informed, proportionate regulation, particularly if it is stable, creates a terrific environment in which to innovate.
Lord Hall of Birkenhead: I would like to ask a bit more about the public information campaign. I know this is your personal view, but we are all worried about the general election and misinformation. If you were to give a brief on what that public information campaign should be, from where you are sitting, what should we be really worried about?
Professor Dame Angela McLean: I can tell what I would do: I would go away and talk to the experts I already know who work on disinformation and misinformation in other arenas. My understanding is that they would tell me: “Show people what the tricks are. Show people how they would be tricked”. I do not know what those are off the top of my head.
Baroness Harding of Winscombe: You mentioned that there is no AI regulator today. The AI White Paper is very clear that the Government do not want one and, instead, want existing independent regulators to cover their areas with—these are my words, not the White Paper’s—a light-touch central risk function. Should we be concerned, because, as far as we have been able to discover so far, that light-touch central function does not exist yet? Are we wrong? Does it exist? If it does not, should it?
Professor Dame Angela McLean: I will pass on that question. I am really sorry; I do not know. Could I ask you to ask Viscount Camrose?
Baroness Harding of Winscombe: I will be doing exactly that in an hour’s time.
Professor Dame Angela McLean: He is a much better person to ask than I am. Forgive me.
Baroness Harding of Winscombe: No problem.
Lord Foster of Bath: When you were answering questions from Lord Kamall and a subsequent question from the Chair, you talked about making health and science data available for, among other things, LLM training. At the same time, you stressed the importance of valuing it and went on to talk about how we must avoid underpricing it. Do I take it from that that you believe that LLM developers should pay for the use of this data?
Professor Dame Angela McLean: It depends on who they are doing it for. If they are developing something that is just for the public’s use, then no, but if they are developing something that they will sell, then yes.
Lord Foster of Bath: In your view, should that apply to all other forms of data?
Professor Dame Angela McLean: I see where you are going. I have already written about this and said that it requires an ongoing dialogue about where we are with copyright and how we want to manage copyright to properly balance the rights of creators and the ability to move forward with new technologies.
Q121 Lord Foster of Bath: We talked earlier about the balance between innovation and safety, and I suppose we are now talking about the balance between innovation and copyright holders. You said that you want to go away and think about it, as more thought should be given to this. We are very keen to put something in our report about it. Your predecessor, Patrick Vallance, was very clear about his belief that text-mining and data-mining exceptions should be extended to make it easier for people to get the data to innovate with LLMs and so on. Do you share that view, or are you saying just that it needs more thinking about?
Professor Dame Angela McLean: I will find you something that I have already said, so that I am at least consistent, if you will forgive me. I think this is in the hands of the Intellectual Property Office at the moment, and I would be inclined to await further comment from me while the IPO works on it. I said: “The government response … confirmed that a code of practice would be developed to enable AI innovators and the creative industries to grow together in partnership”. My understanding is that those recommendations currently sit with the IPO.
Lord Foster of Bath: The problem with the code of practice is that it has been suggested that it will be voluntary. A lot of copyright holders, who are vital to the creative industries, are often small individuals who do not really know how they are able to exercise any of rights they might have. If the code is voluntary, there is no way of knowing whether people have abided by it or not. It is not really a solution. Do you think that a voluntary code is really the way forward?
Professor Dame Angela McLean: Can I read another bit?
Lord Foster of Bath: Yes, please do.
Professor Dame Angela McLean: “While the development of a voluntary code of practice is an important, pragmatic step in resolving these tensions, it should not run counter to IP value creation, as IP is central to the sector accruing value and its capacity for growth”.
Lord Foster of Bath: I strongly share that view and I am very pleased that you have written it, but it means that we need to do more than is happening currently. Surely we need clarity as to where the law actually stands, because there are lots of LLM developers scraping data, which definitely have not paid for it, do not have licences for it and so on, and our creators are losing out. That may be the price we pay for innovation, but that is a decision that people have to make. We are very keen to know your view on it, but I think you are saying that you do not have one at this stage.
Professor Dame Angela McLean: This “should not run counter to IP value creation, as IP is central to the sector”. That looks like a view to me. I have told you what I have written. I look forward to seeing what you write.
Lord Foster of Bath: Okay. Thank you very much. That is it.
Q122 Lord Hall of Birkenhead: The answers that you have been giving touched on advice, as of course you would, but could you pull together your views on expert, external, scientific advice to the Government? This year, we have been through the Foundation Model Taskforce, which morphed into the Frontier AI Taskforce, and now we have the AI Safety Institute, which is focusing much more on risk. Do you think that the right balance is being achieved in the expert scientific advice to the Government? Are there gaps or places that you think will not be covered by the institute? Should we and the Government be worried?
Professor Dame Angela McLean: It is very much part of my job to make sure that advice to government comes from a broad spectrum of opinion. That means commercial opinion, academic opinion, opinion from our own experts in government labs, and opinions from across the spectrum. That means opinion from people who think that truly dreadful things might happen and from people who say, “I just don’t see a pathway to that”.
I opened by describing the work that my office did, which was very careful to say, “Here’s the range of opinion. Please don’t think that one side of this opinion sits in the commercial world and the other sits in the academic. That is not where the split lies”. One reason why there is room for so much opinion is that there is not that much evidence. Some of this will be resolved with the acquisition of evidence. For example, as next year plays out, we will see what kinds of capabilities are generated in the next round of these very large models.
Q123 Lord Hall of Birkenhead: The AI Safety Institute has a brief. I am interested that you describe your role as bringing in people to add to the pool of thinking, which must be very interesting and exciting. What parts of the bigger picture is the AI Safety Institute not looking at, which you have to make sure are promoted and pushed within government?
Professor Dame Angela McLean: We are extremely interested, for example, in helping the Government themselves to get ready for these huge changes that are coming. We have a document that describes five different scenarios for 2030. It says, “Imagine these five different ways in which things could pan out”. They are basically stories that we will take across government to different departments and say, “Look at this. If this happened, how would your department react? What would it mean for you? If the world were like this in 2030, what might your bit of government do to make it better before we got there, or, if it ended up like that, what would you have to do to manage in a scenario like that?”
That is a Foresight report, which is a standard product from GO-Science. I think it is great. I did not write it; my team wrote it. We did its first run-through with some Permanent Secretaries last Friday and had a full and interested discussion. So that is one thing. Of course, this is not particularly about safety; safety is one of five big dimensions that we think about here.
Lord Hall of Birkenhead: This is more about innovation.
Professor Dame Angela McLean: It thinks about capability and adoption. When we imagine these things, we ask, “Are they more closed models or is it an open model?” We think about the geopolitical context in which all this plays out. Those five things are there as guides for thoughtful but busy people to ruminate, in what we hope is a structured and helpful way, on how a technology will change the working lives of those of us who lead in government.
We also suggest people to the AI Safety Institute. We help it to draw up its rosters of experts who can advise it. I chivvy my mates, the other GCSAs in other countries, to join the expert panel to help with the State of the Science report. Those are main things that we are involved in.
I go back to the science and technology framework, because all the things that are happening across government with it ought to help with the commercialisation of these technologies.
Lord Hall of Birkenhead: To go back to the scenario and what you did with the Perm Secs, that is a good way of getting organisations to think hard beyond the everyday stuff that they tend to think about constantly. What was the worst scenario? What was the scenario that pushed them the most and made them think the most?
Professor Dame Angela McLean: We put them into three separate groups, so it was not a proper experiment; there were no replicates. My least favourite scenario is AI disappointing us—that is, it being just another damp squib, and it turns out that we never managed to sort out the hallucination problem so we can never use it and we have to go back to Excel spreadsheets and Google. That is my least favourite scenario. I have not brought them with me. Did you ask for the worst?
Lord Hall of Birkenhead: I asked for the most challenging.
Professor Dame Angela McLean: Probably the AI Wild West, where things that are fairly safe are all driven off open models and the idea is that somehow the open LLMs win the competition and become the best models. So instead of there being four of five companies that we can squash when they are bad, there is an anthill of LLM providers, and then we have to think about how we will have the invisible hand of the market regulate this mess.
As you can imagine, we had a super-interesting discussion where one person sat back and said, “This is fine. The market will sort it all”, and another person said, “No, we’ll have to be really clever regulators and design some AIs to regulate the AIs”. We had a very rich discussion, but that was just the start. I hope that we will have more interaction, using those scenarios, right across Whitehall in the coming year.
Lord Hall of Birkenhead: This sort of points to the view that civil servants are very open to the sort of advice they need from you and the people you bring to them. Is that right, or are there still patches of greater difficulty?
Professor Dame Angela McLean: So long as the problem is not too difficult. This was a very open discussion. It was in a session that had been designed specifically so that there would be time for people to put away the difficult problem of today and have a day to think about AI. I gave one of the talks, actually.
Lord Hall of Birkenhead: Is it about having a more scientific Civil Service—that is, a more scientific top of the Civil Service? Presumably, that includes politicians and Ministers too.
Professor Dame Angela McLean: That is for the electorate to decide. I do not think I can change that, but I can try to help. There are two things here. One is that everybody should just be more scientific, because it is fun and interesting. Another is that being a top civil servant is a rare skill. Really good civil servants do amazing things. We have these things called CSA networks where you pluck out people like me, plonk them in and say, “Get on with it”. But we have to have loyal wingmen to whom we say, “I’d like to do this thing”, and then they operate the machine. I want there to be a career stream for people like that who will go on to be part of the senior Civil Service.
Lord Hall of Birkenhead: And that is what they are aiming for.
Professor Dame Angela McLean: Yes. I would like there to be a whole bunch of them. I have loads of wonderful colleagues who would be fantastic senior civil servants—I used to be the CSA for the MoD, so I know loads of scientists who work at Dstl, the defence lab—but I do not think that they could do it tomorrow; they would have to come and learn. I do not know what the magic is, because I am not a career civil servant, but they would have to spend a couple of years learning how to be a top-class civil servant.
Lord Hall of Birkenhead: Thank you. That was fascinating.
The Chair: I could comment on that, but I will resist.
Lord Lipsey: That was a fascinating answer. Obviously, the scientific approach is a strong one, but what has struck me in listening to months of evidence is the sheer degree of disparity of view, particularly about risks. At one stage, we had one lot of people signing a letter saying that we must have a six-month pause in all this—some of them did it for commercial reasons, but some of them did it because they believed it—and, on the other hand, we have had people being completely liberal about it, relatively speaking, and worrying about regulation being a way of blocking innovation. Maybe this is true of the early stages of any scientific development, but this degree of disparity seems to make the challenge of dealing with it all much harder than it otherwise would be, even before we know what happened with Sam Altman.
Professor Dame Angela McLean: Yes, it is difficult. I agree with your statement, but I am not sure what your question is.
Lord Lipsey: The question is: do you agree?
Professor Dame Angela McLean: I do.
Q124 Lord Kamall: You mentioned your international counterparts. From your conversations with them, in your opinion, which countries have really thought about this? Where have you thought, “Gosh, we really need to do a lot more thinking in this area”? Whose thinking do you admire, and have you thought, “We need to think more like them and think about the issues they have raised”? Have I thrown too much at you there?
Professor Dame Angela McLean: No. Clearly, the Americans have given this a lot of thought. Actually, the Europeans have given it a lot of thought and have come up with a different answer. I have not read what they think, but I am told that Singapore is very front footed on this. To be honest, the equivalents that I know best are Australia, Canada, the US and New Zealand, because we are in a sort of club together. One of the points of the safety summit was to draw together international discussion on this issue. We cannot do this on our own. We can regulate it until we are blue in the face, but what happens then? It is quite interesting to read about the Chinese attitude towards regulation too.
Q125 Lord Kamall: You have given us some bedtime reading in that case. I have a second question. You probably do not want to go into too much detail on this—I apologise—but one debate that we are looking at is whether there is enough publicly available data to train the models. If not, are you concerned about a scenario in which a lot of the trainers decide not to access copyrighted material because it is too difficult, meaning that the models are less effective because they have not had the fullest dataset possible?
Professor Dame Angela McLean: You are getting into a very technical area that is beyond my expertise. Rather than give you a silly answer, I should say that I am not best placed to answer that question. However, my understanding is that, for more limited tasks—for some in government, our tasks are quite limited—what matters is having the right data, particularly in the later stages of training.
Q126 Lord Kamall: My last question is about transparency. People say that they want to see more transparency, which means transparency of algorithms and of datasets. Even if we have that transparency, will people understand what that means?
Professor Dame Angela McLean: I think experts would tell you that you always have to trade efficacy against transparency, and is that really what you want? I might add that we have plenty of medicines that work but we do not know why, and we have other ways to make sure that they are safe enough and to keep track when surprising things go wrong. There are other situations where we use technology without complete transparency about why it works.
Q127 The Chair: I have a couple of questions that are both linked and come at the sort of issue that Lord Lipsey raised but from a slightly different angle. You said earlier that one of the influences on the broad range of opinion was the lack of evidence. I thought that was interesting, because is there not sufficient evidence from the way in which other technologies have developed that ought to inform what people think about the balance of risk and opportunities or, indeed, how this technology should be regulated? That is my first question.
Professor Dame Angela McLean: My question to add to that would be whether we can think of an example of a technology with such broad application, where we get a step change in apparent capability, such that even the people who have built it cannot articulate exactly how it works.
The Chair: Does that not apply to the internet?
Professor Dame Angela McLean: I do not think so, no.
The Chair: Why?
Professor Dame Angela McLean: Because the internet is a very large network of information across which various capabilities, particularly in trade, emerged in interesting ways that can be studied pretty openly by anybody. I am happy to debate this further, but I do not see that as a close analogy. I thought you were going to say the published book.
The Chair: Is that an example you want to draw a comparison with?
Professor Dame Angela McLean: I am not a historian, but an interesting comparison would be to look at what the fears were, because my understanding is that there were a lot of fears. I bet there are some historians here who could tell me about the fears about what would happen if we democratised all that information and made it available, because it all used to be locked away.
That is a really interesting question. If you find out more, let me know. The question is basically how a fundamental technology like this emerges. It is a super-interesting question. The internet is one and the book is another. I am thinking about the Industrial Revolution and the agricultural revolution, which we all studied so much at school.
Lord Griffiths of Burry Port: What about nuclear physics and energy?
Professor Dame Angela McLean: Nuclear physics is totally engineered, is it not? Every step inside our application of nuclear physics is understood to the point of being able to describe it, very accurately, with an equation. That is not true of this technology.
Q128 Baroness Harding of Winscombe: This question is connected. I want to follow up on your comment that there are medicines that work but we do not understand why, and yet we have a regulatory regime for safety. If you take that as a good analogy for the regulation of large language models, what does it really mean us doing?
Professor Dame Angela McLean: I have had it vigorously argued to me that the way we regulate medicines is not a good model for LLMs. But suppose we did it that way: we would have trials. I have had interesting discussions with people in the AI Safety Institute about what stage this is at. If we take that model, the stage we are at at the moment is to think about what we go through for a new medicine. First, we do tests in a test tube, then on some sort of model animal, usually, and then on very tiny numbers of people. We do not even have the equivalent of those tests yet, and that is the stage that the AI Safety Institute is at at the moment: it has to devise the equivalent of in vitro tests, in vivo tests and phase 1 or first-in-human tests. What would the equivalents be for those three? That is quite an interesting analogy, even if we did not end up having phase 1, 2 and 3 trials in the wild.
Lord Young of Norwood Green: I was thinking about social media being just as powerful, in its way, and the challenges that there have been in trying to regulate it.
Professor Dame Angela McLean: I regret the things that we did not do already, don’t you think?
Lord Young of Norwood Green: I see this as another challenge. I was fascinated by the Sunday Times article on the sacking of Sam Altman and that organisation, and the reinstating of him and the arguments within the organisation about impact. There is no absolute answer when we are trying to assess risk against innovation.
Professor Dame Angela McLean: I am sure you are right, but we await your wisdom.
Q129 The Chair: My final question goes back to the broad spectrum of opinion and your view that there is a lack of evidence. I am intrigued to know how you, as a scientist, view the debate—among technologists or scientists—on this technology as it exists at the moment and effective altruism. Have you considered and taken a view on this?
Professor Dame Angela McLean: No. I would leave effective altruism out of it.
The Chair: In what way would you leave it out of it?
Professor Dame Angela McLean: The question of how we are going to benefit from the promise of generative AI is sociotechnical, and we must figure out how to benefit from it, while first knowing and then protecting ourselves from the risks. It is a huge question. I view effective altruism as a side issue about lifestyle.
The Chair: That was very helpful. Professor Dame Angela McLean, thank you so much for your time and your answers to our many questions.
