Welcome to this Education Select Committee oral evidence session. This is our first evidence session in the Committee’s new inquiry on AI and EdTech. We are excited to get going with this important piece of work. We are joined this morning by a guest, Allison Gardner MP, from the Science and Technology Committee. Allison, you are very welcome. Before I invite our witnesses to introduce themselves, I will put on record that one of our witnesses, Kester Brewin, is a personal friend of mine. He is here in his expert capacity, but I should say that for the record. I am not aware of any other declarations that Members need to make this morning. I invite the witnesses to please introduce yourselves.

Morning, everyone. I am Patrick Milnes, head of policy for people and work at the British Chambers of Commerce.

I am Josh Hillman. I am the director of education at the Nuffield Foundation, where I oversee all our work, research and programmes from early years, through schools and into various post-16 education and training pathways. I should also say that we host the Ada Lovelace Institute, which aims to ensure that the opportunities, benefits and privileges of data and AI are equitably distributed and experienced across society. We are well into a joint programme of work with the institute on various issues around the education and AI agenda.

I am Will Akrigg and I work at the King’s Trust as UK government affairs manager. King’s Trust, formerly the Prince’s Trust, helps young people across all four nations get into education, employment and training.

I am Kester Brewin, associate director at the Institute for the Future of Work. I am speaking today in a personal capacity, though, partly due to my 25 years’ experience as a secondary school teacher. I spent quite a lot of that time writing on education and technology for various publications around the national education press.

Thank you very much, everyone. I will begin our questioning this morning. Could I ask each of you to briefly set out how you would describe the impact that artificial intelligence is having on the world of work? Are we experiencing an evolution, or something much more revolutionary? How are the impacts distributed across different sectors and different parts of the country? We will get into a lot more detail, so I am looking for an overview and headlines at this particular point in time.

The first thing to say is that the capabilities of AI are rising faster than expected, and the evidence on how that is translating into fundamental changes in labour-market demand, in employer behaviour, in productivity, is actually quite limited and contradictory. We are probably overestimating the changes that are happening in the short term and underestimating the long-term change, but the direction of travel is quite clear. AI is changing work across a wide swathe of occupations and sectors unevenly. It is important for today’s purposes to see that entry-level jobs, in particular, are quite exposed, because some junior roles are more susceptible to automation, but also because employers are cautious about hiring, which they are when things are unpredictable. It hits those entry-level roles first. That is a big issue for the education system, which cannot possibly predict the future needs—imagine if you had done that five years ago—but it can prepare young people for adaptability in that more uncertain, fluid and AI infused labour market.

We are seeing that the use of AI is also rapidly accelerating. We did some research earlier this year, and 54% of businesses are now using AI in some capacity. To contextualise that, uptake was only 35% in 2025, and two years earlier it was only one in four. You can see that the uptake is really picking up. What we see is that 95% of firms that are using AI have reported no change in their headcount, and 86% using AI report no change in the nature of the jobs they are doing. It looks more like we are seeing augmentation at the moment rather than a general transformation of the workplace. We are seeing good productivity gains as well as a result of that, but it is quite uneven in terms of the adoption we see at the moment and where the productivity gains are being spread around. There is a risk that the use of AI might widen gaps between firms, within sectors but also across sectors and regionally as well. To follow on from what Josh has said, in terms of education and the skills that people have, the skills system is not aligned with the pace of change that we are seeing at the moment. There is an urgent need to develop a serious plan for long-term AI literacy, and not just in education but lifelong reform of how people can continue to adapt as the technology changes.

We tend to agree with both the previous answers as well as the Milburn interim report. AI is clearly a massive opportunity in terms of productivity and growth. We also agree that it is not the sole cause of the rising NEET rate, but it is definitely changing the picture. We are already seeing it massively impact on the job market generally, with things like online applications—43% of young grads in recent research have encountered AI in an online job application. One line from the Milburn report stuck with us, “Recruitment has become more remote, more automated and less human.” When we talk with young people every day, we see this with them constantly. Many speak quite candidly and dishearteningly about applying for tens if not hundreds of jobs and hearing nothing back.

I think the issue is that this is a highly powerful and very highly divergent series of technologies, so some of the impacts are going to be very good and some of them are going to be very different from that. As people have said, it is not really robots coming for people’s jobs but it is fundamentally changing people’s experience of the work that they do. That obviously is going to impact teachers and the work they are doing, but also how students are experiencing work and how parents are experiencing students doing work. It is that nature of what technologies are always able to do, which is to either improve the discretion that people have in their work and take away some of the toil or the drudgery, or push them into lower-discretion roles. One might see the difference between, say, the black cab driver and the Uber driver. Both are taking someone from A to B. One is a very high-discretion role: you have to have the knowledge, and you have to do all these things to get there and, therefore, there is a wage premium. With Uber, it is driving to the same place, but everything is being driven by the algorithm at a price push down in terms of an algorithmic bid. There is no discretion there for the same role. One of the key things with this very powerful technology is: what is happening to the discretion of people’s work? I have spoken quite a bit to people about thinking about a much simpler technology, like a hammer, which amplifies our ability to act in the world—I cannot push a nail into the wall with my thumb—but it then also reaches back into the heart and says, “What do you want to do with that extra power?” In fact, all technologies have that thing. They amplify what we can do in the world, but they force us to ask questions about what we want to do with that power. Very clearly with social media, we did not get that reflective process right in terms of, “Gosh, here is a very powerful set of technologies. What are the questions we need to ask about how we use them?” Those questions need to happen together. It cannot be a personal reflection. I think that is what is really fundamental in the research that we are seeing around the impact in work and workplaces. It is where workplaces are having those high-engagement conversations, talking to people right through the workplace around the impact of these technologies, that you see the best results. If it is just dropped in, the impacts can be much more low-discretion work, so it really drives down to governance, how people are using that within firms in terms of what they are doing to reflect on the power of these technologies and how they can be used.

Thank you very much. All of you have spoken about advantages, opportunities, risks and problems, and how we mitigate those risks. Do you think that the Government have grasped the scale of the impact that AI is having, and what will happen in the future? Do you think the current approaches to regulation, oversight and the mitigation of risk are fit for purpose for the pace of change we are seeing in the magnitude of the challenges in front of us?

I could quickly carry on in terms of that idea that we have to have a socio-technical approach. After world war two a lot of work was done in coal mines around increasing productivity. They found that specialisation did not really improve productivity because it broke the relationships of the people working in those mines. This was work by the Tavistock Institute in the 1950s. It said that in order to get the best out of the system we have to co-optimise both the technical and the social dimensions of that. One of the concerns I have about some of the approaches that Government have taken so far is that they have been very strong on promoting the technical capabilities, really positive about AI in that, but there is not necessarily so much focus on having to make sure that we match that with investment in the social capabilities to make sure there is that co-optimisation. We are seeing a bit of that now in terms of the people-centred adoption challenge, and much more about people-centred approaches, but it is fundamental, I think, to get that right in terms of having that co-optimised approach. From my experience as a teacher, what worries me is that very often technologies are dropped into schools without any of that proper social approach about understanding what might happen. You would go back in September and suddenly there is a new piece of kit in the room, no instructions, no talk around how to use it, how to use it together, or what the impact on learning might be together. It worries me, particularly in schools that are very resource constrained, that this extremely powerful set of technologies could be dropped in without really thinking about the social implications for the fundamentally important ecosystem of relationships that is so precious in schools within learning. We have to get that right. The Government push on AI is fantastic, but I would really want to see the balance with the social approach as well.

I am happy to come in from the side of young people. It is probably not for the King’s Trust to judge the Government’s overall response to AI as an industry. There have been some promising initiatives, including the TechFirst programme and the AI skills boost. We have been working quite closely with Skills England, including on the AI skills framework. We released a bit of research with Public First last year, surveying 2,000 young people. What we found quite striking was that only a third of young people believe that employers, education providers and charities are doing enough to help them prepare for AI; 53% felt they had been left to figure out AI on their own. We do not think enough is being done with regard to that.

There is another area where I would like to see more Government intervention and that is around the youth labour market. There has been a lot of focus on whether there are going to be fewer job opportunities for young people, and whether the youth labour market is going to be contracted even further by AI. But we also need to look at the nature of those entry-level jobs. Is the learning and career development that has traditionally always happened in those jobs being eroded in any way? When tasks are being automated, those roles require more judgment. They require quite sophisticated experience of using AI in the workplace, and high-level employability skills as well from day one. That is going to be a problem if young people who previously in those early-stage roles develop those judgments through undertaking those tasks and by being observed, getting feedback from more experienced colleagues and guidance. My belief is that we should see that first rung into employment as part of the education and training system. It may not be something that is governed by Government, but it certainly needs to be regulated and given some attention, because this is where young people learn to work, and this is where they learn to apply their employability skills. You might see employees getting a very quick hit from an AI-driven productivity gain, but they also depend on that future talent pipeline that comes through the pathways through which future workers develop the skills, knowledge and judgment to become managers later on. Looking at that early stage of employment as part of something that the Government need to give attention—certainly this Committee needs to give attention—I think is another important area for Government intervention.

To take a slightly broader, more generic approach to the question, we want to see a really strong assurance framework in place from the Government, something that looks at how and why people are using AI. At the moment, the pace of technology is so fast, and the adoption is so excitable in some cases. Kester was touching on people who are just dropping these technologies into their organisations without really considering the implications of what it is doing and how it is doing it. I think there is an ethical standpoint, in terms of ensuring that UK businesses are able to use AI in an effective but fair and morally correct manner. There is more that could be done to ensure that businesses understand the underlying technologies or algorithms that are driving their use of AI and what that actually means in terms of the data that is being shared, who is it is being shared with and all of that.

Thank you very much.

To what extent do the effects of AI on the workplace vary across different sectors of the economy and different regions of the country? Patrick, you already alluded to this. Which sectors are more vulnerable to the effects of AI?

What we are seeing at the moment is obviously a big disparity across sectors and size of business as well. There is still a cohort of small firms—26% of them, according to that research I have already referenced—that say they have no plans to adopt AI, but only 10% of large businesses say they have no plans. You can see the discrepancy there. What we are also seeing in terms of the sectoral difference is that only 42% of manufacturers currently use AI. But if you look at business-to-business services firms, 57% of them are currently using AI. You can see where the gaps are. At a more basic level, you also see a gap in urban and rural in terms of the use of AI, but that is not necessarily driven by the use of AI. That is a more fundamental disconnect in the level of internet service that is available, because you cannot use AI if you do not have high-quality internet—good-quality broadband. This research is a bit older, from 2023, but at that time 82% of businesses in urban areas said they had a reliable broadband service and for rural businesses it dropped to 56%. As I said, that research is from a few years ago, but I would expect those figures to be roughly similar now, based on conversations I have. If you cannot even get reliable internet you are obviously not going to be able to use AI in the same transformative manner. That is a big divide.

Our Gen(eration) AI report, which came out last year, basically looked into the jobs that young people are currently holding and are in. We found that AI was set to change 56% of the jobs that are typically held by young people: 46% will be augmented, with an element of AI introduced into the role, and 10% will be displaced entirely. These are the roles that young people are in—things like retail and wholesale roles—and 60,000 roles are estimated to be displaced and 500,000 significantly augmented. That is sector variation but, geographically, cities are going to be hit harder than rural areas. In Belfast, 13% of the jobs held by young people are at risk of displacement, compared with 7% in rural areas. One thing that Public First, with which we did the research, thought could happen off the back of this is entry-level employment going down in cities and, therefore, fewer young people actually moving to cities, with higher cost of living and rent. We found the difference quite stark.

More generally, this is a series of technologies that could really help to reduce inequalities, including geographic inequalities. It may also further entrench existing inequalities. We know that that can be from a socio-economic perspective and, certainly within schools, different access to different levels of resources of AI could have a very profound impact on the learning outcomes for young people. It has the opportunity to reduce inequalities in the sense of being able to help people in terms of geography and access and those things, but it really does have the potential to further entrench those inequalities because of the costs of these systems, which are still not really particularly well known. The current costing model for using AI is token cost—I think most people are on a monthly subscription—but those token costs could really spike up, a little bit like an oil price. We know that that is having an impact on schools’ thinking in a resource-stretched environment: “If I am going to sign a contract with a major big tech provider, how is that going to work out in terms of suddenly all my tokens have gone? I either have to be able to top that up with a big spike in cost or we are not going to be able to use it.” For those schools that have more resources, that is not going to have such an impact. There are potential socioeconomic impacts of these things, which are a little bit more hidden under the surface.

One thing that builds on all of those points is the importance for young people of careers guidance. It is always important, it always needs to be absolutely core to the educational infrastructure, but the uncertainties geographically, sectorally, among different occupations and in what qualifications you need just underlines the importance of that. Young people need to have accurate labour market information about what jobs are changing and what tasks are changing within those jobs. They need to see where the emerging growth areas are, and where those jobs are going to be. Do they need to move? Do they need to think about recruitment practices? Will was talking about these mass recruitment platforms—do they need help in understanding how AI mediates that? There are a lot of big issues for careers guidance to help give young people opportunities to make the most of the changes in the labour market and potentially to their advantage. That can, maybe, narrow some of the regional gaps that we have been seeing.

You have mentioned a bit about the difference between jobs in certain sectors and B2B. Is there any evidence around how AI is affecting graduate jobs versus non-graduate jobs? Would it be fair to say that there are some non-graduate jobs that actually might even be better protected, perhaps, than graduate jobs?

We have done a bit of research with the Institute for Fiscal Studies around this. Similarly, in agreement with the Milburn interim report, we found that the evidence isn’t specifically that AI is affecting non-graduates more than graduates. We believe AI is affecting them all equally, because there isn’t that decrease in vacancies for graduates that you would expect for AI. If a new graduate cannot get a role in the knowledge economy or a typical graduate role, they will then go for those lower entry-level roles that require lower qualifications, and there will still be a narrowing at that point.

One thing that might be worth adding to that is the fact that graduates on average are probably having more experience of using AI as part of their education—there is plenty of evidence about that. They are acquiring some of that AI literacy, which means that when they are entering those roles, they are probably better able to immediately apply some of that in the workplace in a quite sophisticated way.

It is probably worth saying it is a very highly contested space as well. There was a recent report published by Warwick Business School, I think it was called “The Broken Ladder”, which was trying to argue that actually the decrease in entry-level roles is better explained by the rise in remote work after the pandemic. Again, there are arguments, counter-arguments, one report saying one thing, another saying another. What is very clear is that whether it is AI adoption or remote work, both of those speak to the absolute vital importance of institutional culture. What are we actually here for as an institution in terms of our place of developing pipeline, of building pipeline, and of building a society effectively, one in which we want young people, all people, to flourish? Whether it is one or the other, it is actually fundamentally a culture point about what we are going to do with these new technologies and the powers they give us. We need to carefully reflect on that in terms of the implication for our hiring strategy, and for our talent strategy longer term.

Josh, in an earlier answer you referred to the employability skills that young people entering the workforce in the age of AI will need. Could you set out a little more what you mean by employability skills? What skills are you referring to?

I am talking about the transferable skills that are developed partly through the education system, but then partly through practising those skills in work. We are talking about areas like collaboration, communication skills, creative thinking, organising, planning, prioritising work, problem-solving—these are all increasingly important in the labour market from the get-go rather than being developed through work and through your career. The research we have funded has identified those skills as being shared across wide ranges of the employment sector, but also that the education system needs to evolve to some extent in order to deliver them, because it is a mainly knowledge-based curriculum. Those employability skills need to be combined with that foundational knowledge and judgment, because it is the combination of those that allows young people to be able to evaluate AI outputs, for example. They need to know enough about various topics to be able to spot when something is plausible, when it is partial, when it is right or wrong. So you need the knowledge, you need the employability skills, and you need the AI literacy. By AI literacy, I am talking about not just how to use specific AI tools, but understanding what AI can and cannot do, what the limitations are, the bias, privacy and ethical issues—you need all those three things to be built into the education system at some point before people enter work.

Will, on those employability skills, I think others would call them essential skills or life skills or whatever you want to call them. Do you agree with that assessment about not just AI literacy but a much wider range of what others would call—I hate this term personally—soft skills being important?

Yes. They have been called core skills as well. One thing we really try to do is build confidence and aspiration, which there isn’t too much of around AI at the moment. We are doing a bit of work with Skills England and their AI skills framework around this, as well as working with employers like Google and Capgemini. We find that there is a bit of a comms piece, really. An employer, for instance, will advertise a job but they are not entirely sure themselves, or they will not be using the right language to make it accessible for a young person who might never have had a job before. We know young people are more likely to be using AI for personal use than professional use, and they might have those digital skills anyway. The flip side of that is that we are also working on that same comms piece with young people, trying to understand, when you are reading a job description, what skills they are actually asking for, and looking at themselves, their experience, what they could bring to the table, how to communicate that—soft skills.

Kester, can I ask you a very leading question? Is our very heavily knowledge-rich curriculum preparing young people to go to employers and explain to them their essential skills? Is it preparing them to enter the world of work with the skills that employers are looking for?

I think schools are being asked to do so much, and I think the answer to that is no. The deeper question is: what are the drivers that are pushing schools towards having to focus on knowledge, outputs, exams, grades and so on? I hope it is not too early for a bit of philosophy of mathematics, but Bertrand Russell in a 1920s paper was talking about how we can know things. He describes two sorts of knowledge: descriptive knowledge, which is knowledge about something, and the different facts about whatever that might be; and then there is what he called knowledge of acquaintance. Of course, one would understand that acquaintance knowledge and what we might call the core skills—the soft skills, the relational skills—are absolutely fundamentally important if we are going to truly know in the world. One of the things that schools have struggled with in the past is understanding what exactly a school is and what it is for. There has been that sense of teachers being very knowledge rich, and they are going to pass that on to young people, and we are going to assess how much of that transfer has happened. AI and other technologies have fundamentally changed that. From my sense, a school has to reimagine what it is there for. For me, it is there to really prioritising that knowledge of acquaintance. You are here together to learn together, to do that critical judgment and thinking and those sorts of things. That is why I am afraid, going back and back and back, it fundamentally begins with oracy. It fundamentally begins with early language and reading. Without that, how can you possibly expect to do critical thinking and judgment about what an AI output is giving you? Really fundamentally vital, right at the beginnings of an education, is language acquisition, because we are all large language models and we have to make sure that young people are given that opportunity to have that so that later on, yes, they can acquire knowledge and that can be assessed, but fundamentally they are able to understand and critically judge. That is very important early work that needs to happen.

It is never too early for epistemology. Patrick, what is the perspective from businesses on these big questions?

Employability skills—I will use that phrase because it was in the question—come up all the time in terms of people coming out of school or other education without those core skills. We have done some research asking what skills employers think young people should be focusing on the most in order to best enter the world of work? Seventy-two per cent of businesses said oral communication, 67% were talking about confidence and resilience, 62% teamwork and the ability to work with others, and only 47% said digital skills. That shows what others have spoken about here: people are focusing on those transferable skills that work across a variety of different jobs. If I could touch on the careers piece as well. What employers tell us is that people are coming out of education without any sort of work experience because the careers advice and guidance that is in schools at the moment is quite limited. There are a number of reasons as to why that is. It is very limited. For a lot of work experience, you get two weeks when you are 15. I was chatting to business owners the other day who were telling us that their son was trying to find work experience. He could not find anywhere because he was 15 and employers only want to take people who are 16 or above, so there is a fundamental flaw in when the work experience is done in schools that is acting as a huge barrier. What we also see is that careers advice and guidance, not at all schools but at a lot of schools, is just done by a teacher and is an activity that is tacked on top of their other teaching responsibilities. Obviously teaching is a very hard job, they work very long hours, and then it is difficult for them to do careers advice on top of that. They are not careers professionals. Many teachers have only ever been teachers, so it is very difficult to do a thorough careers advice and guidance process that you would want to see. What you would like to see is more initiatives to get businesses involved, to get them into schools and education providers, to offer that support and to show young people all the opportunities that are out there and available.

A quick addendum to that is that we just published a report looking at how AI is being used in careers guidance, and often just as unadulterated AI-driven career guidance. What our research has found is that it works fine as a hybrid model with human support, helping young people interpret the information and understand the choices, building their confidence, making decisions and connecting to those opportunities in the real world. However, there is a real danger of a move towards cheap and cheerful careers guidance that is purely driven through these AI technologies, which may improve efficiency but may not improve the wider career development of those young people. So we would call for hybrid models.

Yes, I have touched on some of this already. It is definitely narrowing the number of opportunities for those young people, and it is also changing the nature of those jobs. It is particularly going to be challenging for those young people who have lower-level qualifications or who have struggled to get work experience and work readiness through their pathways. We are talking here about the million 14 to 24-year-olds who come under the NEET category. That group of young people is really weakly connected to education and work by definition, so they have benefited less from developing those skills. They also have more limited access to technology and to high-quality forms of AI and safe tools and so on, so they have that second challenge. Then their actual use of AI has potentially widened the gap in learning and development—they may have not used it to scaffold their learning; they have used it more to substitute. Then they are also facing what Will was talking about in terms of this mass rejection that comes through these huge AI-driven recruitment systems. They have all those things compounding, which is making it much more challenging for them even just to get that first job. That is a very large group of people who are effectively being disenfranchised through all of these forces. AI is just one of them, but they are all compounding for that group.

Retail is classically an industry for young people with low to no qualifications to take their first step into the world of work. That is clearly just one sector that AI is narrowing, essentially. We found evidence as well that does seem to show that, in terms of vacancies, AI is having an effect. In terms of young people’s aspirations, we have found that one in five young people have already changed their career plans because of AI, 71% believe AI will make competition for jobs much more intense, and more than a quarter decided not to apply for a certain job because they thought AI might replace them. Clearly, it is already having a narrower effect in terms of even young people going for the job. That is another barrier.

It is quite difficult in this question to separate the impact of AI on the number of entry-level pathways available for young people versus the wider factors that are also impacting entry-level pathways for young people. At the moment a lot of employers are operating on very tight margins. There is a lot of cost pressure. Things like the Employment Rights Act are going to make it harder for employers to operate. The national living wage and national insurance increases from last year, the impacts of the war in Iran—all these things are adding up pressure on employers. What we are seeing is that because of the uncertainty and ongoing issues, a lot of employers are withdrawing from making any investment or hiring decisions at all. They are making do with what they have. They are going to look at their internal staff, what resource and capabilities they have already, and therefore they are not going to bring any people in. Then when they are bringing people in, because of how tight their margins are and the way they feel they have to operate at the moment, they are trying to take risk-averse approaches to hiring. The interpretation that a lot of employers will take when they are trying to be risk averse is to bring in someone with a proven track record—someone who has a lot of experience and someone who they think is going to be able to hit the ground running in a way that a young person may not necessarily be able to do—because, again, they do not feel they have the breathing room and the capacity to help to get a young person up to speed, because of how tight their operations are at the moment. Yes, it is difficult then to separate out what is the AI impact compared with the rest of it.

It will be adding on and, again, what you see is that where you do have an employer who finds some limited money to make an investment in their business, they may invest in an AI tool instead of a young person, because they view the return on an investment in the short term as better. Part of the issue is that businesses have to make a lot of short-term investment decisions, because there is not the financial freedom and economic stability to be planning for five or 10 years down the line. They are looking at what they need to do now to ensure their business is still operating in six or seven months’ time.

I totally agree with that. One of the problems is that when businesses come to that hiring decision, they may be making pre-emptive decisions like, “Oh, I should be investing in AI because everyone else is investing in AI,” rather than looking at the facts of what is happening on the ground. We are not necessarily seeing businesses make decisions based on an informed choice of, “It is going to give me this productivity boost,” but in a general ecosystem liberation of AI being the problem and AI being the answer—all of those things—whereas, in fact, it is a polycrisis and there are so many different things going on. What we do know is that—this is work that the institute published—the basket of skills that young people are being asked to bring to a first job has grown so that the demand is bigger for them, which puts a huge onus on young people. It would be fair to say that if young people are applying for 100 jobs, there has to be a question: are you really understanding or coming out of school with a sense of purpose? What are you really driven to want to do? That speaks to the careers piece, where it is a really mixed bag. There is some fantastic work on careers being done, helping young people understand what their intrinsic and internal motivations are to drive them towards—“Yes, I really want to be passionate about that”—and therefore making those connections, and really thinking about the relational aspects. It is a big picture across cross-transitions that needs to be considered.

We need to need to move on. We have quite a few questions still to ask and I need to wrap this panel up by about 11 o’clock, so I encourage you all to be as concise as possible in answering. That would really help us to get through everything.

We have already touched on my question, because I was going to ask you about the Milburn report and as many as one in six people in the UK being NEET within the next five years. We have talked about entry-level jobs. Josh, and maybe Kester, how can we use AI to reverse the trend and improve entry-level opportunities for young people, rather than making it harder?

It is a really important question. There may well be new training and progression opportunities for new jobs in emerging sectors in AI-enabled occupations that could be amenable to some of the emerging qualifications that are coming through now we have T-levels. V-levels are in progress and are going to become more important. We are going to potentially see quite a significant increase in apprenticeships. All those routes could potentially be important routes to move some of those young people who are currently NEET into really quite worthwhile employment using AI, and those routes are potentially quite exciting because they combine learning, there is some mentoring, there is workplace exposure, there is employers being involved in the design of some of those qualifications. The positive side of this—I gave quite a doom-laden description of AI needs earlier—is that some of those are potentially quite exciting.

Do you know what sort of new jobs?

We do not exactly know. A really interesting piece of research showed that I think it is something like 40% of current job titles did not exist 10 years ago, and that process will continue. We are already seeing quite a lot of new roles, not just in the development of tools that are using AI but roles that are applying AI in normal business practices.

Kester, any thoughts?

Yes, I will add to that by saying I think one of the functions that AI has is to reduce the frictions that can exist. Much of this is a matching problem: here is someone who wants to work, here is someone who might want to offer work—how do you get those two people together? It might be that someone is in the wrong place; remote work can deal with that friction. It might be that they don’t necessarily have the right skills; we know that AI can definitely deal with that by taking a lot of the extraordinary content online already and algorithmically putting together a course that fits directly into what a young person is passionate about. It is also about decreasing the information friction and the frictions about how to actually do something. For many firms, including those run by friends of mine, it seems like a big burden to get over the administrative hurdles to be able to offer an apprenticeship. If AI can be used to reduce those frictions, that allows more firms a lower level of energy of entry to be able to offer apprenticeships. Similarly, we have also seen a big uptick in the amount of entrepreneurship and innovation. Using AI to start a little company is absolutely fantastic. What if every young person—we have brilliant young enterprise work going on—when they came out of school had already started a company and there were different skills involved in that? That is not about using AI; it is using AI to lower the frictions to allow greater innovation and entrepreneurship.

Great. Thank you.

Did anyone else want to comment?

Slightly similarly—not thinking about broad digital and AI skills—we have been running a pilot in Glasgow using generative AI to get a job. That is focusing on the massive online job boards where you just throw out a CV. We are basically running a session for young people on using AI to set up maybe an automated job search for things they are interested in, or writing a cover letter but understanding a bit more what they are looking at and putting a bit more of themselves into it. It is still early days, but we have had some promising results from that so far.

That leads quite nicely into my question. I wanted to know about the curriculum and assessment review and how we can design a curriculum to best fit the skills that we want. We have talked about changing computer science to computing, micro-credentials, and also the importance of teaching the skills in context of the individual subject areas. The ideas you have both just listed seem to fit with that. Josh and Will, how well do the Government’s reforms to the curriculum and the qualification systems reflect the growing importance of AI in the workplace?

The curriculum and assessment review had a very wide brief, and it is a useful foundation to build on. AI is moving so quickly that the DFE in the implementation phase, which is where we are now, needs to be much more explicit and practical about how AI and digital literacy gets built in, not just for computing but right across the curriculum. I am not suggesting that you need a separate review on AI, or even a separate subject called AI, but you need to map AI literacy right across all subjects. You can see great examples of it potentially coming through—for example, in citizenship, understanding misinformation, digital rights and democratic resilience. Then in English, how do you evaluate generated text and evidence in maths, understanding probability in statistics and uncertainty? There are all sorts of ways in which it can be built across the curriculum. The really important point is about assessment, which is central because that shapes what is prioritised and evaluated within the curriculum, within teaching, within qualifications, within the accountability system. So much of the debate about AI and assessment has been about cheating and the authenticity of children’s work, and how awarding bodies can build AI into the way they practice various stages of the assessment process, but the much deeper issue is about the validity. What is the assessment trying to measure and how is that AI skill and that AI literacy built into what young people leave the education system having understood? That is where I feel a lot more work needs to be done.

At secondary school age, the King’s Trust delivers enrichment and extracurricular activities, which is the capacity we fed into the curriculum consultation. That said, given it is so new, and our research finds that institutions are not doing enough, we are finding ourselves having to educate young people on it more. We have a pilot with an EdTech coming up in the next couple of months—FE Tech—looking into working with employers to find out what skills are actually needed and training young people up in those. So watch this space, potentially.

That leads to a quick supplementary that I will direct to Patrick and Kester. How has the Government engaged with employers in industry to ensure that the curriculum changes reflect the skills required in the workplace?

I forgot to say something very quickly about the previous point. There are some really early signals in the US about a lot of backlash against AI and about the overuse of tech within the school space and within the education space. It is very important to be alert to those. I was on a roundtable at King’s College talking about AI resistance. With a background in engineering my understanding of resistance is that it is not antagonising the system. A resistor in an electrical circuit doesn’t antagonise the system; it stops it getting overwhelmed and blown. My worry is that unfettered AI will literally blow the education system, so we need to have an appropriate level of resistance to make sure that it is useful in terms of its capacity and not blowing things up. That requires a careful reflection of whether we should be using AI within this particular subject, and whether we should be having lessons whereby we are using high discretion around that. It can be, as we have seen with phone-free schools and other things, that the parents are actually quite looking forward to that. There is another thing that is really important about culture. I was speaking to colleagues the other day. The workload in terms of marking and assessment has been reduced, but in some situations as a result of that they are setting more assessments. That is completely bonkers. What that speaks to is a culture within schools that is still driven by, “I have to get more data about my young people so that the senior leadership are really impressed by the amount of data.” Even if you are using AI to reduce workload, you still have to do the culture piece around what do we do with that extra time in terms of enriching people’s experience of education rather than doing more assessment. To speak very briefly to the other point, the cultural piece is super important. We need people who are able to use AI reflectively. Sometimes it is your teacher, sometimes it is your co-worker, sometimes it is your servant, and sometimes you don’t use it at all. Being able to have young people coming into businesses who really get that is important for their skills and for the skills of businesses. It is also profoundly important environmentally because you are using a high-level AI model just to search for the best local pub. This is absolutely bonkers in terms of the amount of water that is being used. Young people reflecting on appropriate use and schools doing that in a values-driven situation is very important.

That has a big impact on cognitive development, which we will talk about later. Patrick, have the Government engage with employees and industry on the assessment review?

Yes. We would like to see more engagement with employers around all of this—not just the use of AI, but the skills system and infrastructure in general. Employers are very keen to feed into their local FE and HE colleges and schools, and make sure they are on the same page and speaking the same language. One of the things we see at the moment is that the education providers and employers think about skills—not just AI skills but all skills—in quite different capacities. That can make joining up the two groups quite difficult in terms of preparing young people for the world of work—and not just young people but older people as well. There is a lot of communication that needs to be done there to make sure that what skills colleges and education providers are training people in are the skills that businesses are saying we actually need. At the moment there is a bit of a disconnect there. There are networks available, like the local skills improvement plans, and the chamber network runs to either 32 or 33 of those across the country now. We would like to see more of that type of collaborative thinking done where employers and providers are brought into a room to ensure that there is a real alignment there.

As I say, we deliver extracurricular rather than engaging with the actual curriculum content. I would probably just stress that what we are hearing from young people is that they are not getting enough from schools.

Yes, and I think this idea of avoiding cognitive offload is incredibly important. That is the purpose of an education system: to help to train people. I go back to that principle of making sure that language is acquired in order that that can happen. I hear a lot of talk about AI literacy, and I want to split that up and say there has to be literacy and then you can talk about AI literacy. It is fundamentally important that that happens first in order that when you then have very powerful language models, you are able to bring your own language skills into a fair game with them.

You would not want to see schools do a lesson that is, “Here is how to use ChatGPT,” or, “Here is how to use Copilot.” They should be teaching the skills that we have spoken about a lot—critical thinking, problem-solving, how to adapt to a problem—and then you plug the tools into that afterwards. Because they have been taught those other skills, they can use AI to enhance or augment the work they are doing in that place. Then again, that needs to sit with a wider reframing and improvement in the careers’ information, advice and guidance that is available to young people as well. You could then develop an understanding of the different sectors and career paths that are in your local area, but also the pathways to get into those, and then how AI can be used within those pathways as well. It is a much more holistic approach.

I think most of the discussion about AI literacy has been about helping young people understand what the biases are, what is a hallucination, what are the ethical implications, what are the environmental implications—all those sorts of things. I really want to double down on this point that understanding the cognitive processes that are either supported or bypassed by using AI is something that children themselves need to understand from an early age. Most of the discussion about cognitive offloading has been almost from the academic side of things—“What is this doing?”—but I think it is something that children need to understand themselves, with appropriate language. They will understand that they can produce in the short term a better output maybe than they could have done by using an AI, but they need to understand that part of the process of education is that deeper cognitive, metacognitive processing that is required for that learning to stick and to be durable. That is a really important part of AI literacy that actually gets forgotten. The key question is: is the thing being offloaded in cognitive offloading part of the learning? If AI is helping them scaffold their thinking, then it may support them, but if it starts substituting for that, then it is going to be problematic for their longer-term development. It is a really hard balance, because on the one hand we are saying that young people need to understand in detail. You do not understand something unless you have actually used it; otherwise, it is just a purely theoretical lesson on what is AI. That is going to be, I think, the No. 1 challenge for the curriculum.

The only thing I would add to that is the absolutely vital importance of doing this together. We have seen classes of children where they are sat in front of the Chromebook, and they are not a class anymore; they are a group of individuals working on an algorithmic learning system. It is fundamental that all of those things—totally supporting what Patrick and Josh are saying—are done together. That is what a class is. The magic of being in a classroom is being there together and learning from one another and talking together about that. AI literacy as a group effort, as a cultural piece together within a community, within a local place where we are bringing our young people together to say what the important things are about the society that we are trying to build, we have to do that together rather than make it a very personalised, individualised and atomised experience.

I have a quick question for Patrick. My constituency is 610th out of 650 in terms of mobile connectivity, so I am absolutely at the forefront of the digital divide. What would you like to see the Government do to reduce inequalities, particularly between rural and urban areas, so that we do not widen this divide as AI becomes more widespread.

I am conscious of how long is left in the session, but the BCC recently published a report around improving connectivity in rural areas. Rather than regurgitate that now, I can get that sent across to you later for some bedtime reading, if you would like.

Do we need to address this with urgency?

Absolutely, yes.

Can I add to that? One bit of nuance in all our research is that young people from the most socioeconomically disadvantaged backgrounds are less confident using AI, less confident about the impact AI will have on their careers, and much more pessimistic about it all. That is to flag that that disadvantage will only grow.

Some more bedtime reading we will be sending you, Caroline, is work that we funded on a minimum digital living standard, which shows that around 45% of families with children do not have the combination of devices, connectivity and digital skills that are necessary for everything that we have been talking about this morning, and that is particularly focused on low-income families. That is before you even start thinking about how those technologies are being used to support learning transitions and support. That divide increasing is a massive danger.

That is great because I cannot sleep in this heat, so I will call that bedtime reading.

On that note, I will wrap up this panel. Thank you all very much for coming to give your evidence to us this morning. It has been a fascinating discussion. If there are additional points—bedtime reading or otherwise—that you were not able to get across to the Committee, or further thoughts that you have, please do write to us and we would very much welcome that. For now, thank you very much. Witnesses: Professor Rose Luckin CBE, David Monis-Weston, John Roberts and Professor Neil Selwyn.

Welcome back to our second panel of the Education Select Committee’s oral evidence session on AI and EdTech. We have four further witnesses to hear from this morning, one of whom is joining us online from Australia—we are grateful to Professor Neil Selwyn for giving his time at a late hour to be with us today. Will all our witnesses please introduce themselves?

Good morning. My name is David Monis-Weston. I am head of AI and EdTech at Purposeful Ventures, which is a not-for-profit that partners with social entrepreneurs and philanthropists to improve the education and wellbeing of young people from their earliest years. I am also a former maths and physics teacher.

Hi. I am John Roberts. I am the interim chief executive at Oak National Academy and also a former physics teacher. Oak is a publicly funded and independent arm’s length body of the Department for Education. We provide digital products and give teachers and schools across England access to free, optional, open and adaptable curriculum and resources.

Hi. I am Professor Rose Luckin, professor emerita at University College London. I have taught in schools, colleges and universities, and I am also CEO and founder of Educate Ventures Research, which is a bespoke consultancy company that helps the education sector in lots of different parts of the world get to grips with AI and its implications for them.

My name is Neil Selwyn. I am a professor at Monash University in Melbourne, Australia. I have been researching education and technology since 1995. I previously worked at UCL in the institute of education, at the University of Cardiff and the at University of Bristol.

Lovely. Thank you all very much for being with us this morning. I will start with the same question that I had for the other panel. How would you describe the impact that AI and EdTech are having on education? Are we experiencing an evolution or is it more revolutionary?

We are only just beginning to see the true potential of AI in education and the impact it can have. Nine out of 10 teachers use AI for their work, from recent data, and 93% of pupils in the UK have used it at the GCSE or A-level age for their revision and for their schoolwork or homework. AI can already beat us all at chess and it has cracked Nobel prize-winning challenges, but education is categorically different and harder. That difference matters for how we deploy and build for AI in education. Chess has fixed rules with clear win conditions, and learning is not like that. I remember from my days as a teacher: it is complex, it is social, and it is emotional. That is at the forefront of everything that we are doing at Oak, particularly around developing safe and high-quality AI in education, particularly for teachers. The digital resources that we have developed over the last few years are being used now by around 250,000 teachers in England. Two years ago, we developed a generative AI tool for teachers, Aila, and we are willing to give the Committee an opportunity to see a demo of that being used in a school if they were willing to do so. This was the first public generative AI tool released by a central Government body for general public use. It is safe, it is accurate, and it is high quality. We have put a lot of work into doing that and have learned a lot on the way. It has been used by 60,000 teachers across England now, saving them around three hours a week. However, we absolutely need to make sure that the AI being used in schools is safe, high-quality and tailored to our education system here in England and the national curriculum that we have. We have a critical window at the moment to make sure that we do not fall into some of the similar traps that we have seen with social media and safeguarding of young people around that. Every country in the world is facing this challenge. The Department probably deserves a little bit of credit here for some of the work that it has already done. We have, as Oak, been supporting that work, and the ingredients of a good AI strategy for education are assembling. We have had the AI product safety standards released, a machine-readable national curriculum is being developed, and we have been supporting that, and the content store to provide materials to innovate on top of around the market, an AI tutoring research programme for disadvantaged pupils, and sovereign benchmarks.

We might get into some of that detail, if that is all right. Can I go to Professor Selwyn next, please?

It is an interesting question. I will backtrack and talk about EdTech. I said I started in 1995. If I look back to 1995, you could argue on one hand that everything is completely different. We have digitalisation of classes. Everything is learned through the screen. We have learning management systems and one-to-one devices. When the tech goes down, it is often now impossible for learning to continue. We saw that with the Canvas shutdown in late April with higher education. On the other hand, as the previous speaker was saying, we still have a continuation of what we call the basic grammars of schooling. We still have a national curriculum. Assessment is still similar. Timetabling is still similar. We still have face-to-face delivery. We have fundamentally the same structures and processes and practices as we had 30, 40 and 50 years ago. Technology does not transform or disrupt or revolutionise education, but it changes things and amplifies certain things and dampens down others. The main impacts, arguably, have been on administration and management of education, and the role of corporations coming in and having more of a say of what goes on in schools and universities. However, as you were talking about previously, we still have the digital divide. We still have lots of entrenched problems. Everything has changed and lots of things have not changed.

I absolutely agree with that. In fact, when I was looking at the recent Government report on EdTech, I was amused that the description of the barriers to uptake of education, technology and AI would have been totally recognisable to anyone looking at EdTech challenges 20 years ago. We see relatively limited impact of certainly large language models in classrooms. We do not see them used very often. We are seeing teachers beginning to use them more. Outside of school, we are seeing them used much more than that. The actual nuts and bolts of what goes on in classrooms has not been changed a great deal. To some extent that is positive because, as previous speakers have said, the brain is the brain. The brain has not changed even though the tools have. We are still trying to optimise a system to help the same design of brains as 20 or 30 years ago. On the other hand, this is one area that is fraught with simple-sounding solutions that turn out on a deeper look to be harmful. We are approaching another one of these waves. We have seen waves with the introduction of the electronic calculator, with Google search, and with interactive whiteboards. The most important thing we can do at this stage is to learn the lessons of what went wrong with those and make sure we do not repeat those. In general, the two categoric responses—“Everything is great. We must use it everywhere,” and, “Everything is terrible. We must completely ban it”—have categorically been proven to be wrong on every single previous occasion, and I confidently say will be wrong this time as well.

Many great points have already been made, so I will say something a little bit different, I hope. It is uneven, it is unfolding, it is evolving, and that is uncomfortable for a lot of people out there in education trying to deal with this. However, it is important to recognise that AI is not a single entity. We have all talked about AI, but most of what has been said—Neil went back to EdTech and so took us back a while—is about generative AI. It is important to recognise the difference because they are not the same technologies. I have been doing this for over 30 years. For most of those years, I have been knocking on the door of education saying, “Hi, I am Rose. Can I talk to you about AI?” People would say, “No, not today.” Then November 2022 happened and everybody wanted to talk about AI. It is important to recognise the difference between the AI that has deep research and evidence about its use in education that we were talking about before November 2022, and the generative AI that is often the subject of conversation today, because the generative AI is riskier than the more old-fashioned forms of AI and it brings different challenges and different opportunities. Recognising that difference is important. The second difference that is important to recognise is the purpose of the AI we are talking about. We can think about AI of all types being used as a tool to support lesson planning—as with Oak—or an AI tutor to support back office with document management, summarising or whatever. This is a tool that is being used. We can think of it as a subject that needs to be understood and we know that, in terms of using the tool, you are much more effective if you understand the fundamentals of the behaviours the AI is capable of. These are interconnected. Then the third way we need to think about AI, which is the most important and I suspect is at the heart of what you want to get at, is that it catalyses a need for a change in how we look at our education systems, in how we prepare people for a different workplace, in how we help them be able to thrive in the future. There is that catalyst element of what is happening with AI. Each of those three parts of the conversation are interconnected. It is important that we recognise that they exist.

Building on what you have said, given the capabilities of some aspects of AI, is there a question about the fundamental purpose of education that we will have to get into in the coming years to help us to get to grips with these problems?

Absolutely. We are facing a revolution, whether we like it or not. We know that the vast majority of students are using these tools, although the data is unclear. One of the biggest challenges is that we are still building the database. I do not doubt the figure that you gave, John. I am sure it is accurate. Interestingly, a prearrival questionnaire that was done by multiple universities in the UK for arrivals last October showed that only just over 60% of students said they had used generative AI or any AI at all, and 39% said they had no confidence whatsoever in using it. We are seeing quite contradictory data, which makes your job harder. Recognising the need for that evidence is an important fact. We have to see what is happening here. You just had a panel on the future of work. It is perfectly clear that the nature of work is changing and we have to help young people be prepared for that. However, as previous speakers have already said, that is not about helping them learn to use the technology effectively. It is about focusing on the fact they need to be good at learning because we cannot predict exactly what they will need to know in three years, five years, eight years or 10 years. We cannot. We cannot predict exactly what their jobs will look like, but we can predict that they will need to be good at learning. We need to hunker down on all that metacognition, all that critical thinking, all that epistemic cognition and all those advanced thinking and learning skills. Of course, you have to learn something to learn those, but we need to think very carefully about what we want our education systems to be able to enable our students to do. That has a huge impact on something like assessment.

Thank you very much. I will move us on in the interests of time.

I will drill down a bit more into the impact that AI is having on teachers, lecturers and early years practitioners. You have already spoken a little bit about it in your opening answers, and I will start with David because he is a fellow maths teacher, and I always have sympathy for those. What impact is it having? Is it having a lot of impact on teacher workload, on teacher practice, or any other relevant area you can think of right now?

It is not having an enormous impact on those areas. There are certainly examples of teachers who are engaging with using lesson planning tools and are more informally using these tools at home to think about their lessons, but we do not have any evidence that it is making significant inroads into teacher workload. Teacher workload is one area that is always a challenge inasmuch as if we did save 50% of teacher planning time, no teacher has ever gone home, as you will remember, without an enormous list of jobs that they have not yet done. In a sense, we can slightly shrink some tasks so that they have more time for others, but we will not necessarily be dealing with teacher workload. My concern is—this is not a new thing with AI—that if we are looking to give teachers shortcuts on lesson planning or marking, these are vitally important processes for the teacher brain to prepare for your next lesson. They need to think about what they will teach in the next lesson so that they are ready to teach it. They need to evaluate what their students have done so that they know how to go in and respond to that the next time. We can make that more efficient, certainly, but I worry both about AI technologies and I worry about people giving fully prepackaged lessons teachers need to read out. Both are the wrong solution for professionals who need to be very reactive in the classroom.

Gathering the evidence is important around this and, as a teacher as well, trying to get some of your Sunday nights back. We talk at Oak sometimes about trying to give teachers their Sunday nights back through some of these tools and products. We are currently undergoing an EEF and NFER trial for Aila to test that workload saving to see whether it is having a genuine impact. That will report later in the year. Our own evidence is suggesting that that is about three hours a week for those teachers who use it at the moment. It is important—we might come on to the cognitive offloading or the shortcuts, as David talked about—that we trust professional teachers to take the resources that they have available to them, adapt them and make sure that they are fit and right for the context of the classroom that they are working in.

I do not disagree with that at all, but what training do we need to put in place so that they can do that?

Taking a step back, we designed the tools that we were working with by grounding them carefully in the curriculum datasets that we have already and had built up over a number of years. The process of generating or creating a lesson with Aila, with our tool, is not about it generating a one-shot resource that you then go and teach. I know from being a teacher that if I had just taken a resource off the shelf, my lesson would have gone badly. The behaviour would have been poor, but certainly the impact of the lesson would have been poor, too. It is all about being a part of that co-creation exercise and perhaps offloading the things that are not important, but making sure that you are having the space to think about the things that genuinely are important for your lesson and being a part of that planning process. Sure, it does generate and help that process and perhaps save some time, but the key goal there is to not remove that planning process. Planning is a verb and the important thing is the process of planning, not the output of that planning process in terms of the resource.

I do not disagree at all. How do we make sure that teachers do not forget that?

We have to trust them. We have to trust teachers. Like I said, lessons will not go well if that is what is happening. Taking a single lesson remotely and trying to deliver that without going through the planning process would be challenging in itself, regardless of whether AI was used in the process as well.

I understand your question around workload, and I can understand why you are asking it, but it is the wrong question to focus on. When back to teaching, probably the most rewarding teaching that I ever did was teaching in further education and adults with special educational needs. I had no training to do that. That was how FE was many years ago. I look back and I know how much of a better job I could have done knowing what I now know about AI. That is the thing: this is about helping people be much better at what they want to be good at. No teacher wants to be a bad teacher. All teachers want to be good teachers. Yes, we might give them their Sunday evening back, but maybe even more exciting is helping them achieve stuff they did not know they could achieve before because we leverage the technologies effectively.

We should do both.

We can help them understand their students better and have more time with their students. It is not about screens. It is about time with students. It is about thinking all the time about how my use of AI, whether I am a student or a teacher or a leader or me, helps me achieve things I simply could not have achieved previously. That will mean that we do the most important thing of all, which is to treat this as a moment in time when we as humans become more intelligent. Our brains are not static. We are still evolving. We can become much more sophisticated in how we think and we learn and behave. Looking around the world, we could do with that at the moment. Let us focus on that human intelligence in everything we do and have it as a bit of a litmus test. Are we becoming smarter? Are we using these tools wisely to make us more sophisticated? That, as an educator, as you can tell, I still find exciting.

Professor Selwyn, what is the impact of AI right now on teacher workloads, teacher practice and so on?

I will come back to what the first two speakers were speaking about. There is a huge industry push for these teacher tools, as you said—these one-shot tools to do lesson plans and everything else. We are doing quite a large project in Australia where we are looking at how teachers are using these tools. We are giving teachers these tools to see how they are picking them up. A minority of teachers are often enthusiastic about the technology and are working out great ways of using this tech to save them time and work more smartly, and they are very noisy about this, but the vast majority are not doing this at all. As the previous speakers have said, teachers are seeing the outputs that these things can spit out and working out that you have to put quite a lot of work around them. John talked about adapting them and fitting them. The pedagogical expertise needed to make the outputs of any generative AI fit with your class, fit with your curriculum and fit with your lesson plans is quite large, but it is also about the social context, getting these things to fit with the students that you are working with, getting them to fit with the classes that you are working with. What we are finding is that when teachers are using these tools, they are putting in a huge amount of work behind the scenes to make these outputs work, not necessarily saving time. They are not forgetting that their pedagogical expertise is important, and they are not forgetting about being empathetic social teachers, which is the good news. A lot of the teachers we speak to, though, are worried about the next generation of newly qualified teachers who will not have had the experience of working without AI to realise that that is what you have to do. If there is a training piece, it possibly is in initial teacher education to make sure that students teachers are learning to teach without AI. My other two big concerns from this are slightly wider, and are about the ways in which this automated EdTech and AI can be used to sideline and maybe displace teachers. This idea of de-professionalisation and de-skilling but also the ways in which this technology can be used as a workplace surveillance and performativity tool are important. It is important to ask labour-related questions about the tech.

Brilliant. Briefly from me because I am about to run out of time, are there any differences between AI applications in different subjects? Can any of you point to differences between different subjects?

Briefly, yes, there are huge differences, not least because people who are grounded in different subjects have different expectations of what the AI will do. Yes, there are subject differences in the way the tools are used and the way that people from different subject backgrounds approach the use of the tool. It starts quite early. I am sure others have more to add.

Yes, absolutely. There have clearly been differences and challenges, and we will all have seen examples of incorrect maths answers from ChatGPT or from AI products. One thing that we did as part of the process of building Aila was trying to benchmark whether it was genuinely creating sound outputs. One area we focused on—and we wrote a paper with MIT Open Learning on this—was around the quality of multiple-choice questions generated. Certainly in the earlier times, it looked superficially great, but there were quite a lot of nuances there. If you are an expert teacher, writing multiple choice questions, particularly the incorrect distractors, so that it is not guessable is hard. We found effects like the correct answer being longer quite frequently, and then it becomes guessable and undermines the quality of the actual question in itself. We have written benchmarks to try to test this and understand where it is good and where it is bad. It is not always a case that when the models improve that they improve linearly as well. You sometimes get regressions in some cases.

Super briefly, yes, there are differences, but these differences are changing wildly. A year and a half ago, we could have confidently said that large language models are bad at maths. Now the latest models are getting exceptionally good at maths. Any generalisations we make will be wrong and will change incredibly fast.

That is a good question, and it is a big question. First, we need to differentiate different types of AI tutor. There are AI tutors that support teachers, and so it is not just somebody interacting with the student. There are AI tutors that are evidence informed, probably not using generative AI, that have been shown to be effective at supporting learners to achieve learning. Then there is a more modern generation of AI tutors with a much looser evidence base. These are more the chatbot-type tutors. Not all tutors are the same thing. You would expect me to say that. It is a very academic answer. However, we know from decades of research that a well-designed tutor can certainly support effective learning. Because of the way that AI can help us personalise, it means that we can adapt to students wherever they are, which is a huge boon for students who are struggling and students who can excel. However, we have to ask the question: “And what else?” Yes, we could design a set of tutors that would help students from different demographic backgrounds with different ability levels to attain greater academic success but, if we do not also at the same time help them to understand themselves as learners, we are not equipping them for the workforce. We have to be careful about how we design and how we implement any type of AI tutor that we decide to roll out. Yes, the potential is there without question to do some good, but it is not straightforward. I have one last point. Even the best-designed tool will not have the impact you want unless it is carefully implemented. That goes back to your question, Mark, about training. We are seeing all over the world that you get a good impact only when you take the time to build the capability in the people who will use these technologies to use them effectively. It is not just about the design of the tool. It is about that training implication. I can come across as a big AI enthusiast, and I am because I know the potential benefits, but I also recognise the huge risks and the challenges. That capacity building in the people is the most important part of anything we do when it comes to getting the best from AI and mitigating the risks.

Is there any evidence that AI tutors do the things that we know from face-to-face tutoring, such as building confidence, forming a rapport, giving particularly the most disadvantaged children or those who might be in care, for example, who have had a disrupted education, a trusted adult? Is there any evidence that you can replicate that with an AI tool?

You can, but it does not necessarily mean the tutor is interacting directly with the child. That is a good point. Some lovely work is being done in Manchester at the moment. It is a small case study, but it is a great example of the tutor being used by the teaching assistant to help them better support the individual. That is the other thing about these tutors. We must not think of them as only a student interacting with the tutor. The tutor can support the human to have richer interactions in some wonderful ways. If I interpret your great question, you would not want to see a child interacting with an AI all the time and I would not, either. No evidence suggests that that would be a good outcome, but there is evidence to show that well-designed tools, carefully implemented, can have a positive impact, particularly when we look at the ecosystem, we look beyond the learner-AI interface, and when we look at the other resources that are part of that ecosystem.

I should declare that I am an SRO on that programme that you are referring to from the Government. To bring together some of what Rose and David have said, the importance of evidence here and how to implement this is absolutely critical, particularly given the pace and change of the technology and the different modalities or ways of interacting with AI. It will not always be a chatbot, or in fact that might not be the most appropriate method of interacting. The tutoring programme that is being run between DSIT and DFE and sponsored by No. 10 at the moment is about co-design with eight providers in the market at the moment. Rose talked about implementation. We have a duty to be involved in thinking about what that implementation looks like to support disadvantaged pupils because the evidence at the moment shows that we see some RCTs being released and random control trials and studies that show some evidence of impact. However, when you drill down into it, sometimes it is for higher attainers, sometimes it is for lower attainers. It is not about necessarily the underlying AI model. It is about how it is implemented, how it is designed, and who it is targeted at. That programme and thinking about how we deliver that from within the Government and supporting the market as part of that co-design is about collecting some of that, being part of that process, and also being part of understanding the nuance of product safety, for example, which is an incredibly important aspect of AI. To give an example about the interactions perhaps around something like “Romeo and Juliet”, taught almost ubiquitously across the curriculum, it seems a benign subject. The ending is a double suicide of two young people. That is a challenging topic to have an interaction with a chatbot about. We need to be careful. There is a lot of nuance there about where that might go. We need to be careful about how we hand off. Is that a safeguarding issue? Is that genuine interaction about how the ending of “Romeo and Juliet” should go? Imagine if that was set as a homework exercise. We need to be carefully thinking about these sorts of things and being involved in the process of understanding that so that, because the pace of change of the technology is so fast, we know and can implement some of the safety standards in the right way and the certification consultation and also be a part of building that evidence base.

My reflection on the AI tutors is that a lot of the early promise in a lot of the early studies turned out to be studies where the population was narrowed down. If you pick the children who used the tool properly, it turns out they got great results. Essentially, everyone who had the resilience, support and so on was getting good results. It goes to your point. A number of children who will be a lot more vulnerable need that relationship. The idea that you have someone there who can be empathetic towards you, whom you build a relationship and rapport with and whom you do not want to let down so you strive a bit harder, who can be that little bit more adaptive and give that emotional read, turns out to be important for tutoring. The idea of these systems as something hybrid that works alongside the human is helpful. The worry here would be that if we are struggling to get the most specialist teachers into some of the schools that are most disadvantaged, then, in those cases, why do we not just give them the AI tutor? That could be extremely impoverishing and could particularly disadvantage those who do not already have lots of support at home and lots of people to talk to and peers who are already supportive. The Government are doing a good job here. A lot of people around the world are looking at how we have invested in it and some of the safeguards around it. It is in general a pretty good approach that we are taking. However, we also need to distinguish between, “Here’s a little box, the classic ChatGPT. Sit there and teach yourself,” which turns out not to work very well, versus the machine learning and adaptive learning, where a lot of structure is built in and it will help figure out exactly the right question for you next. Thinking about how these technologies combine gets us to a good place with AI tutors as long as we are not removing humans from the loop completely.

As Rose says, AI tutors can be used in lots of different ways. The most common way that seems to be taking place, particularly if you look at the US, is the model of putting the kids in front of an AI tutor for one or two hours and freeing up the teacher to then concentrate on the students who need the help. That is the promise. When you look at how these things work in the classroom, though, we are finding that teachers are first of all having to do a huge amount of behind-the-scenes work to keep the technology running. It is quite clunky. The maintenance and repair work to keep the technology going cannot be underestimated. A bit like the automated tills in a supermarket, people are running around making sure it looks like it is working automatically. You often find teachers get bogged down in that work. Also, teachers spend a lot of time filling in the gaps about the student that the technology does not know and that only a teacher can. Personalised learning in classrooms is often mass customisation. The technology can know only so much about students. We are finding that teachers are also having to do a lot of work around the technology to humanise what the AI tutor is trying to do. At the moment, they are doing a fairly good job at making the technology work, but it is certainly putting a strain on teachers. It is not making for less work, but possibly making for more work for teachers. Having a human in the room is important but, again, I am worried that we are burdening teachers with quite low-level work and certainly not freeing them up to be expert pedagogists or the promises that we see from industry.

Can I jump in quickly? One issue that we need to address is that overdependence from a teacher’s point of view can risk deskilling. We need to be wary of that.

I would like to talk about the evidence on how AI and EdTech is affecting the learning process. Maybe you could talk a little bit about what we know so far about the impact of these tools on learning outcomes and cognitive development, maybe among different age groups, and whether they might be narrowing the learning by encouraging pupils and students to concentrate more on the output and the presentation rather than a deep understanding of what they are studying.

That is a big question. Let us be specific about some of this. A clear example you can see in higher education. A piece of research done by Wonkhe this year showed that 47% of students did not believe their grades represented what they knew because they had used the thing to produce the output and they did not understand it, and 38% of them said that they did not understand what they had submitted. Yes, let’s recognise there is a huge problem here. Lots of different evidence sources coming in across the ages demonstrate the potential for students not to learn, to perform, but not learn. A very frank vice-chancellor in Australia, Simon Biggs—you might know him, Neil—says that generative AI has shone a light on poor pedagogy, poor assessment and poor curriculum. If somebody can produce something with an AI without understanding it, we are not looking at what they have understood, and everything falls off that. We need to be careful that whatever we are using these AI tools for, it is for learning, not for that performance. There is a clear ask that we are clear with students of all ages about our expectation of what they will achieve. If they are allowed to use an AI as part of their work, the expectation should be higher. They should be able to achieve more because they have a smart tool to help them. Do you see what I mean? We have to be careful about ages. I have huge concerns about very young learners using any AI. Let us get that out there. We need to be very aware of that. However, for older learners, we do need to gradually introduce them and help them to use these tools effectively, but help them to use them for learning, not for producing an output. It is important.

This will be an annoying academic answer, but I would shy away from looking for evidence that something causes something. We have been looking for evidence that technologies enhance learning or hinder learning for 30 or 40 years. It is almost impossible to do it because there is so many different moving parts in any learning instance that to pinpoint the difference on the technology is super difficult. Every learning instance is different for me in terms of context. My great use of a technology in my classroom one day may not play out in someone else’s classroom because of context and everything else. You will never find evidence that adding technology X to your school system scales up and increases learning outcomes and pushes up PISA. We are seeing this with generative AI now. There is a big rush to try to get out studies and meta-analyses that prove gen AI boosts learning. We can get great small-scale case studies and proofs of concept that things work in particular contexts, but none of that evidence scales up to a system-wide level. We are seeing this now. It is a bit of a mess and pseudoscientific at best. Clearly, using technology does affect how students engage in learning, but how the relationships play out on a system-wide level or a country-wide level will take years and years to play out. We have to be happy with a level of uncertainty here. Pinpointing learning is complicated. If it were that straightforward, we would have proven that technology enhances learning 20 years ago. It is tricky. The evidence wars at the moment are messy.

Thank you. Does anybody have anything else to add? We are a little bit short of time.

Yes. Firstly, there is plenty of evidence at the minute that one of the most important factors of whether you will use AI well is how much you know about the topic. Anthropic produced a study recently that the people who were most effective at coding were not the ones who would been taught AI skills but the ones who knew most about the thing they were coding around, about the medical area or something. You need to build the knowledge to use the AI. The irony is that if you use the AI too much when building the knowledge, you go around the key effort you need to get the knowledge. There is clearly a sensitive moment when to learn about the thing to use the AI, you need to not use the AI so that you can use it effectively later. This is not new. We saw the same with calculators as well. Again, people said, “All calculators are terrible. Ban them,” or “We should use them for everything.” It turns out that with calculators in particular there is a certain sensitive period around years four or five. Beforehand, it does not matter if you use them because you are not thinking so much about that particular area. If you use them during that period, you are not doing the thinking about the arithmetic skills to develop them yourself. If you use them after, it turns out that all they do is bypass wasted effort that lets you focus on the right bit of lesson. It will be the same here. We need to ask not whether they will harm learning or hinder learning, but where specifically. Sometimes it is great to cognitively offload. We probably all do it with colleagues and tools all the time, but not when it is the main thing you have to think about. Again, we can learn lessons from what happened before. The same thing is also true with Google. I had many homework pieces handed to me that were copied and pasted from the first link on Google. The same will be happening now. Children have always found ways to bypass thinking. It is around structuring it and thinking of the tasks that help them think hard about the right thing at the right time.

I will quickly say that I agree with everything the panellists have said. Learning requires productive and effortful struggle. We should not offload that effortful struggle. Even in preparation for this today, of course, I suspect David and maybe Rose used AI in preparation, but I wrote my notes out a number of different times because memory, as Dylan Wiliam has said, is the residue of thought. We have to think these things through and not completely outsource the thinking to AI. You would not want me to have done that at all.

Following on from Caroline’s question, I am concerned about the issue of cognitive offloading, particularly in adolescence. Your statement, Professor Luckin, was that we need people to be good at learning because we cannot predict. In adolescence, the brain is still developing. It is still developing its working memory. It is still developing its executive function and its critical reasoning. I am concerned that we could cause long-term changes to the structure of the developing brain if we do not use these tools correctly. How important are perseverance, learning from mistakes and delayed gratification in education? Is that important? Do AI and EdTech tools risk undermining those aspects of the learning processes and, importantly, the cognitive development of children?

There is a huge risk, and you are right to be concerned. We often forget when we think about AI that in the time that AI technology has advanced quickly, we have also advanced in our understanding of the brain as well. We need to join the two things together. I am sure that, used wisely, a well-designed AI intervention can build metacognition, can build critical thinking, can build the right sophisticated thinking and learning capabilities that we want our learners to have, but it does not happen by accident. It does not happen just because the tool is there. You have to make a clear effort. Yes, you are right. Of course we have to have productive struggle, which is why it is about using the tools to help people achieve more and to build more. However, at the same time, your question speaks at the heart of something that is a challenge for all of us, and that is how we generate robust evidence about what is happening when these technologies are changing so quickly. We need to think carefully as a community of researchers and policymakers about the new ways in which we can think about evidence so that we can get you an answer quickly, to give you some ongoing input to answer that question that you are asking. We do have the tools to find out, but we do not have all the answers, and you are right to be asking those questions. We want to ask those questions and get answers because, otherwise, we risk not getting some of the benefits that we could get from these technologies.

I completely agree with all of that. Productive struggle is important. It is nuanced by age as well. How we apply AI through earlier years and if at all, indeed, is an important question versus how it is used in later school age and beyond. One thing that we have done more positively at Oak around this recently is to release digital literacy lessons. Two weeks ago, we released digital literacy lessons that were built in partnership with the Raspberry Pi Foundation. They go from key stages 1 to 4, years 1 to 9, around information and data, safety and security, digital problem solving, and using technology and AI effectively. With the changes around the curriculum—

Will that help with brain development and the cognitive development of children?

The importance here is teaching and supporting pupils to use some of that technology effectively and helping teachers to then teach pupils to do that as well. Those lessons now exist and we know that they are being used and taught. Teachers can take them and use them to teach and hopefully support pupils to use AI or EdTech or technology generally in ways that will meaningfully help their education rather than doing exactly as you describe.

On cognitive offloading, depending on who and where, if you go back 50 years, plenty of people have cognitively offloaded because they asked their parents or they asked their friends or they copied the homework. That is bad. They did not end up learning, and they got caught out later in the test. Also, some people asking their friends and their parents at exactly the right time would get a useful result because they were saying, “I am stuck. I cannot get any further. Help me cognitively offload my problem for a minute and then help me re-see it and then come back to it”. Simply saying, “This tool can help me cognitively offload,” I would not necessarily agree is bad. We need to think about how we stop people using it so that it is unproductive and saying, “I will bypass all my thinking,” which children have always found ways to do, but also find ways so that when children are genuinely stuck, it should not only be the ones with parents who can help them at home, or peers who happen to be able to help them, who have somewhere to go. It is a matter of how we use these tools and structure them so that children will not use them at the wrong times, but also try to make tools available for some children who will not have any other places to go. I worry about oversimplifying. We can see the same issues for children who are being asked to memorise times tables. Some of them wrote out answers because they did them on their calculator, but calculators can be useful. It is not demonising the tool. You are absolutely right that cognitive offloading is the right thing we should be looking at. Let us drill down to specifically where and how and why. How do we make it used for good in some places and not used for bad in others?

You might want to look at a nice paper by Leslie Loble and Jason Lodge from Australia—Neil will know them—on beneficial cognitive offloading and detrimental cognitive offloading.

One conversation I had with my constituents around the social media ban was a frustration from some parents that they felt that they were being told to make sure that their young people were spending less time on screens, while at the same time schools were often asking them to support their young people to, for example, do their homework on exactly those same devices. How can we balance this challenge between making sure that our young people are digitally literate and AI literate, while also recognising that screen time can have wider negative effects on children’s wellbeing?

That is a good point and important. This area of education unfortunately does not do balance or nuance particularly well. You are right that there is a huge discrepancy between what the Government are saying about getting screens out of classrooms and then pushing for AI, for example. One thing I would say quickly is that the screen time debate is often talked about, but most media researchers would say that screen time is the wrong thing to be talking about. It is not the quantity of screen time, but the quality of what you are doing on the screen. You can read for three hours on a screen to deep dive on a screen in a Shakespeare text, which might be fantastic—far more beneficial than 30 minutes mindlessly scrolling through TikTok. We want to talk about quality engagements with tech. However, everything that we are discussing today in this session needs to be set against the context of the popular mood music outside of this forum on getting screens out of classrooms. Looking from the outside, everybody else apart from the UK—and I guess you will be doing this soon—is pushing to get devices out of classrooms. Sweden is already getting devices out of primary classrooms. The Los Angeles Unified School District, which a few years ago gave an iPad to every student, is now prohibiting in-school screen time. Even high school students are only allowed one and a half hours a day. The same thing has happened where I am in Victoria, Australia. There is a real push to get screens out of classrooms. Norway has announced a near ban on AI in all elementary up to 13. We are talking a lot about the benefits and the potential. There is a growing parent and possible even teacher push in the US for AI pauses and AI moratoriums. Students booed Eric Schmidt at a graduation ceremony. The sentiment is rising and AI is becoming quite unpopular with the public over the past few months. We are seeing a big swing back to getting all tech out. We do want to push for balance. We want to change public understandings of tech and public understandings of tech in education. It is about balance, it is about nuance, it is about quality rather than quantity and trying to get over this idea of all or nothing. I am worried that we will swing all the way back to nothing again and never get some of the benefits from technology.

Professor Selwyn, you are right that we are looking at other countries. In fact, we went as a Committee to Estonia, where they are taking a very proactive approach to AI and education. One thing that struck me in Estonia was that they are having to, I suppose willingly, factor in how to respond to certain students who are AI refuseniks and choosing not to engage with AI. They are having to incorporate them within a wider framework. I saw lots of nodding heads there while Professor Selwyn was talking. If I can be potentially a little unfair, does anyone have any points of differentiation for the sake of time in terms of how we manage this tension?

We need to separate screens from AI because AI will largely be delivered via glasses or via voice. It is not about screens alone. I sympathise with you. I am a parent and a grandparent. I sympathise with you and with your constituents over screen time. We have to look at it carefully. However, when it comes to AI, we need to recognise that a lot of the ways that we will be interacting with it will not necessarily involve a screen. We need to think about that more nuanced perspective. I would add that to what has already been said.

It is worth saying quickly that the vast majority of lessons taught up and down the country in schools at any single time are not using technology in any way. It is absolutely about quality over quantity, and I completely agree with that. We do not have a ubiquitous screen time issue necessarily in schools and in classrooms right now. This is not intentionally trying to change that.

I would also say on this debate that it is frustrating to me that we cannot collectively put more pressure on tech companies on this. The fact is that all devices go to all young people with default everything turned on and a parent has to have a degree in engineering—I have a degree in engineering, and I still do not know how to figure it all out—to turn off all the right settings. The settings will change every five minutes, and they will be moved around. Why is this the case? Why is it not the default that everything is locked down and then we give simple steps where parents can open things up rather than say, “Now the children cannot get the device because we cannot be sure that people are able to lock them down.” That is fundamentally a challenge. The other thing I would say is that the US is fascinating on this. It has always spent so much more money on EdTech than we have. When you talk to Americans about classrooms, they assume there is a screen in every classroom. It is exactly that challenge we talked about before. If in a 60-minute lesson, the teacher is wasting three minutes of her time helping fiddle with tech to try to get it fixed, that is 5% of lesson time. Over a 10-year career, that is half a year of teaching time potentially lost. We massively under-weigh the impact that small frictions that devices introduce in lessons can have on the quality of teaching that children are getting. We need to ask what those three minutes of friction could be used for elsewhere. What could all that money be used on? We could be hiring specialist teachers. If every time you introduce a device you have to train loads of teachers on how to use it, you could be training them on the basics of better feedback and teaching some fundamentals better. We do not take the trade-offs as clearly as we should. You can introduce tech in lessons without giving children screens. There is an interesting product where, for example, children hold up mini whiteboards, the teacher says that three quarters, half or one quarter have the answer right and puts it into the system, and it adaptively says, “You might want to ask this question next.. These are ways to use great tech, but we do not have to have every child in front of something.

In Estonia, they have also done a great job at building capability. There is a lot to learn there.

Special needs inclusion has always been an important area. It is not very profitable so it is not very commercial. Funds do not get pushed to it. The social inclusion aspect is much more tricky. I am worried that there is a tendency to see AI as a quick fix for students who are struggling at schools or school refusers that will give them more AI. We need to push against that. There is no quick fix to supporting learners who are from socially deprived, multiply disadvantaged backgrounds. We need to invest in people and resources and time. Technology can be part of that in the background, but technology will never be the solution to any social disadvantage issues.

Yes, I agree. For most of the time I was an academic, I believed AI would be a great tool for breaking down barriers and for greater inclusion, and it is not rolling out that way at the moment. We need to be careful about that. Potentially it could, but we have to be super careful.

it is worth looking at the EEF evidence around some of this. A high-quality, well-sequenced curriculum is core to effective teaching, which is the most important lever we have to support pupils with disadvantage. That is fundamental to the purpose of what we are trying to do. At Oaks, as we are all moving into a more AI-enabled world, it is about thinking about the infrastructure, some of the data and some of the content that we can put in place so that it can be used in ways that might support wider use with inclusion and be grounded in that rich curriculum, but also for innovations perhaps around other assistive technologies as well. As Neil summed up, it is not always a commercially viable, unfortunately, or commercially focused technology, but assistive technologies using AI have, potentially, real benefits.

I have nothing to add.

Thank you very much. That pretty much brings our evidence session to a close. We did not quite cover, and we are interested in your thoughts on, the integrity of assessment with AI, but we are pretty much out of time. If you want to drop a note, if any of you have any expertise in that area and there are matters that you think the Committee should consider, we would be hugely grateful for that. For now, thank you very much indeed for joining us today. That brings our session to a close.