Welcome to today’s session of the Business and Trade Committee, as we wrap up our inquiry on artificial intelligence. Thank you very much to our witnesses who are here in person in a very hot House of Commons. Professor Carl Frey, thank you very much indeed for joining us online. Carl, perhaps I could start with you. You have written a brilliant book about this and you are publishing some excellent work. Tell us, what do you think is going to be the impact of artificial intelligence on the UK labour market? What are your best assumptions about this right now?

Thank you very much for your kind invitation. It is a real pleasure to be with you. Right now, we are not seeing much in terms of impact of AI on labour markets. We are seeing some impact in certain pockets, such as translation work, and other areas, such as coding. Those are mostly tasks that are verifiable. One key reason that we will still need humans in the loop for some time going forward is that most activities still need human verification. AI makes mistakes and we need workers who are capable of verifying the output. That does not mean that there is not going to be an impact on labour markets going forward. One critical question is where those workers are going to be located. For example, if AI makes accounting much easier, it stands to reason that some of those jobs might migrate to places where labour costs are cheaper, for example. We may see more accountants being hired in Manila, rather than in the City of London. AI reducing barriers to entry in knowledge work and professional services is one key mechanism by which we will see an impact of AI on the labour market relatively soon. Unfortunately, we lack the data to track these trends in real time. Another mechanism, which is underrated, but we are beginning to see in the data as well, is that more people are using AI to fix various things themselves. If I use AI, for example, to fix my boiler, that may not replace the heat engineer, but it means that the heat engineer will lose a call and lose one job. This working assumption is that AI is mainly going to have an impact on knowledge work and professional services. I think that as more people do more things themselves, we are going to have an impact on manual trades as well. We will see more firms offload verification to their consumers, because that is easier to do. Further down the line, I think that we will see some outright automation, likely in transportation and logistics. That often takes more time because you need restructuring. Often, when you look at some of the studies that count up the tasks and say that 75% of the job is automatable, but the remaining 25% is not, the question is whether those tasks are needed to begin with. Often the way that automation proceeds is by simplifying procedures. We did not automate away craftsmen by building robots that did the exact same procedures. We did that by restructuring the work in the factory setting and then applying sophisticated machinery. There is the question of what the next generation of companies will look like. We are beginning to see AI-native companies that are much leaner than incumbents. If those firms turn out to be more productive, they may well outcompete some of the incumbents. Then we might see an impact on the labour market through that competition as well. Another key question is to what degree we will see these emerging new firms developing products versus processes. For the most part, when you use a tool for efficiency and improve productivity in the task that you already to do, you are going to replace some workers. If you use it to create new products and new industries, you are going to create new jobs as well. That is one key reason why business dynamism has historically been really important to new job creation. Large firms have scale. They are more likely to use AI for process improvements. Smaller and young firms do not have scale and are more likely to use it to develop new products and industries. Therefore their entry is going to be critical as well. Going forward, we will need better data to assess some of these trends in real time. We also need to better understand how consumers are responding to these technologies. We tend to often focus on machine capabilities. I am quite happy to go to a self-service checkout when I go to the grocery store, but when I have dinner with my wife I do not want to sit in a giant vending machine. What is the difference? We need to understand in which settings humans prefer humans, rather than machines. We lack data on that too, unfortunately.

In terms of policymakers thinking about two potential paths—an automation future or an augmentation future—you are saying that we do not know yet. It could be a bit of both, and understanding what the real-world impacts of that will be is still looking through a glass darkly.

It is most certainly going to be a bit of both. It is also important to remember that an augmentation future is not necessarily beneficial to everyone. A bank teller who can tell only one currency is of no use in a modern bank. Even if digital technology has augmented bankers, who have now become relationship managers, that may not have been beneficial to the workers originally in that job. There are a lot of open questions at this moment, including to what degree, if AI augments work, firms will simply rehire workers with the right skills, rather than retraining their existing workers.

Let us test that and understand what it means for roles and sectors.

Good afternoon, everyone. I will say for the record that the Institute for the Future of Work clerks the APPG on the future of work. That is just to make that clear for the record.

That is on the record.

The Government’s approach to assessing this is to break down jobs into high exposure, low exposure, high complementarity and low complementarity, with my rough guess being that those that have high exposure but low complementarity are the sectors that are most at risk from job disruption. Would you say that approach is the right one? Have they identified the correct sectors? Which sectors do you think are going to be most affected? Is there a regional element to this that we need to understand as well?

Can I back out a bit from the question? I will come back to answering it. For me, I feel as though, as policymakers, it is not a question of if; it is a question of when and how. We are often spending too much time worrying about—not worrying, but overly considering—the micro of this, when the macro is undeniable. I want to reinforce that. I have literally no idea which jobs are going to be most augmented and which will be least augmented. I believe that anybody who tells you that they have a clear view of the next three to five years is not likely to be that correct. It will be different. In my opinion, every role will be affected. It is going to be an intentional way that we step into that future. I believe that we can either let every role be affected, or we can make sure that every role is redesigned together and effect the change together. That may sound like a slight nuance. I will give you an example, because you mentioned tradespeople. I completely agree, Carl, with a lot of what you said. To pick up on your example of an electrician and take that further, I might, heaven forbid, try to mend things—I would fuse my whole house and blow up everybody I know. That could create more work for electricians, because I have actually done a worse job. It could create less work or different work. I agree with you that there will still be work. The thing I believe very deeply is that there will be an electrician who is working with AI and an electrician who is working without AI. The one working with AI is going to be the one who will be more competitive, be more successful and have more wealth in the future. I believe, as policymakers, we should be tilting into a future like that for this country. We should not spend so much time being so anxious about what the anxieties around particular sectors necessarily are. We should take the view that this has to be about wide participation, broad access and trying to minimise inequalities in how people are able to use it, so that we do not do what we have done in the last 30 years of technology, which is to exaggerate inequalities to a large degree. Sorry if I slightly did not answer your question.

That sounds like the Indian AI-for-all model that we heard about when we took evidence on this in Mumbai.

Yes, possibly. I think that there will be a UK version of this, because we start from a very different place and we have different attitudes culturally to technology, and specifically to AI right now, where, understandably, there is a lot of fear. There is some anger and some conflagration. You see the statistics often that people are very angry with OpenAI in the abstract, but cannot live without ChatGPT in their daily life. You get that combination of complexity. There will be a UK version of this. It is very important not to design so much for a specific sector at this point in the journey and miss the bigger picture here, which is that there is still a huge opportunity for the UK to get ahead of this and stop doing what we tend to always do in policy terms, which is dance around a bit and not go big and bold. We should be thinking about how everyone in this country has deep access, proper training, reskilling and opportunity. We can see that some jobs and tasks are changing right now, and Anna, Adam and Carl are much more expert in this than I am, but that does not change what is still the macro picture, which is what we should be striving for in a three to five-year horizon.

Adam, what is your perspective on Mr Madders’s question?

I agree. We are in a position of radical uncertainty and we need to assess the evidence as it emerges. The challenge for policymakers and unions otherwise is how we adopt a no-regrets framework, where we plan for every eventuality and anticipate every eventuality, but also think about what we can do now that we know would be good anyway. We know that we are already building an AI-enabled economy in a situation where we have the lowest in-work investment in skills in the private sector in the OECD. We know that we have the lowest business investment in the OECD. We have an economy that is plagued by short-termism, shareholder primacy, weak investment and low participation of the workforce in shaping things. I will give you a contrast to that. I was in Sweden the other day, meeting Swedish trade unionists. One of them said something very striking: “We don’t fear new technology. We fear old technology,” and that is because old technology makes them less productive, less competitive and less able to upgrade their abilities to participate in the economy as a whole. The reason why they do not fear this is that they have a stake in the economy as a whole. They have 88% collective bargaining coverage in Sweden. In the UK, it is 13% in the private sector. That means that you have a situation where the wider system, whether it is the social security net, the skills system or the taxation regime, is set up to be high productivity and high innovation, but also high equity and high resilience for the workforce. If we are thinking about AI as a transformative technology, which I agree it may be, or is, but perhaps over a longer term than we are led to believe by some of the CEOs, we need to think about what institutional architecture we need to build to reflect that. To briefly sum up, in contrast to that, right now there are 500 workers in British Gas who are about to be made redundant in call centres because people are using customer self-serve with AI-enabled chatbots. The company has introduced this, customers want it and those workers are going to be made redundant. That happens; it is being partially driven by AI. There is an interesting thing, which I call the drudgery elimination fallacy. The workers who remain are picking up the most difficult calls, so the ones where the customer is angry or the issue is intractable. They have gone to the employer and said, “How is this reflected in our pay packet and our working conditions?” because they were doing more demanding work. The employer said, “We are not going to talk about it.” That comes back to how you capture the gains of this. What is the incentive structure for the gains to be reinvested into the workforce and into innovation? Over the past two years, British Gas has seen a 33% increase in shareholder dividends in 2024 and a 22% increase in 2025. A lot of that will be super-profits from energy prices, but how are the gains of this transformation going to be shared and reinvested in the workplace, rather than just gravitating to the top? These are some of the big questions that we need to grapple with.

There was an American trade union leader called Harry Bridges, who once said that his philosophy was about how to make sure that workers, in his words, got a piece of the machine. It sounds like quite a similar argument you are advancing.

Yes, and that is what we hear from members as well. It is not about being anti-technology. It is about being anti-exploitation, or a concern about the resilience. Part of this is the material things about how you gain a share of the wealth that we all create. The other part is the dignity component. CWU surveyed its workers, of whom 80% have never been informed or told about AI being deployed in their workplaces. They are not aware, so they have not been able to participate. If you are not able to participate, you are going to feel, and be, vulnerable. The bigger question is the things about the economics, but there is also whether we feel valued when we go to work. Are we actually valued? Do we have a stake in the workplace? Do we have a stake in the wider economy? We have institutional mechanisms to support that but, back to Martha’s point, we need to be bold about doing that. We can do that without getting trapped in the analysis paralysis of, “What is happening to AI? What is happening?” We know that there are weaknesses in the economy in different places. There are models and techniques that we can use to improve that. It is not easy, but there are things that we can do.

To build on Adam’s point, there are some brilliant projects that start with often young people, sometimes out of the workforce, who are designing what roles might look like in companies in the future and actively looking at problems. Employers will come in and say, “Have a look at all the things we do. Work out what roles we might need.” There are practical things happening. To reinforce Adam’s point, the analysis paralysis is real right now and we should start doing things. There are examples of it.

Anna Thomas, what is your view on where the impacts might lie and what we can do about it?

The scale of the challenge is real. There are some very loud voices that get that part right, but perhaps not the cause or the driver of it. You highlighted, Chair, the importance of augmentation. By that, I would mean people’s choices, capabilities and skills. Redesigning work and potentially business models in a way that is directed and puts certain goals and people first is hugely important. It is not a coincidence that everyone on this panel has highlighted the importance of participation and involvement, which perhaps invites us to go upstream and downstream in thinking about sharing benefits. That is hugely important, but we do not have to start there. Professor Frey made some very important points about other factors driving the labour market and the fact that, overall, in terms of numbers of jobs, there is not a lot of change. Although I agree that there should not be an analysis paralysis, the latest research pulls out the point that we need to refocus on work transitions and supporting work transitions. That is happening inside jobs, firms and sectors, as tasks, skills and roles move around. There is enough evidence now, including evidence from the Institute for the Future of Work, which is an independent research and development institute, as you may know, that suggests that we are at a point where there are signs of a new equilibrium in that, and that we must not assume the best results. As Baroness Lane-Fox says, the potential is there, but often those points of opportunity are also points of risk. They must be understood and targeted perhaps more sharply than they are now and at a bigger scale, as we have all said.

Could I just put a specific scenario to you to understand whether we are ready for the change? We know that driverless cars are happening across the world. We know that the Government are very keen to get trials under way. We can assume that that is going to be a feature of our economy in the not-too-distant future. We do not know exactly when and how it is going to take shape, but we know that it is coming. Are the Government in any way prepared for the job displacement that is going to take place as a result of that?

I have not seen much mention or plan for that. I have met drivers, and reps in GMB and Unite for drivers in London, where it is being rolled out. The evidence that we have so far from where driverless cars have been rolled out extensively is from the US and China, but the US is perhaps more equivalent to our labour market. There are job losses, but the biggest thing that we see is downward pressure on wages. If you look in Seattle or in Phoenix, you see an around 10% decrease in the take rates of the drivers who remain driving. The companies are going in—it is the same model that Uber used when it came to London—and undercutting the existing market because they have capital to be able to do that and a “land and expand” mechanism. Obviously, more supply drives down the rates of those drivers. In London, there are 100,000 private hire vehicle drivers. Those are individuals, but supporting families in different ways. I am not aware that there is a substantial plan for how their income is going to be affected, how they will transition if they need to, or other things. For example, the new companies or operators are going to create new jobs. There are going to be depots for cleaning people’s sick out after a night out and for maintenance, and the call centres and control rooms. Are those jobs going to be outsourced? Are they going to be secure? Are they going to be hostile to unions? Are they going to be on good contracts, or is there going to be further precaritisation of a sector that has been at the sharp end of algorithmic management and dynamic pay, which is the abuse of technology to exploit the pay rates that workers take? Those are the conditions that we are arriving into. Again, that is the institutional architecture; that is not the technology. That is where public policy has power, especially in markets such as London. London is huge and has leverage. If any of those companies want to come play here, we should be setting terms and signalling what those terms are.

I agree with that. I do not think that we are ready. Taxi drivers are a good example where there is not a buffer. Pay is one example of that. There are not transferable skills. There are not organisational or employee factors and resources in place to support those transitions, so there is not adaptive capacity. In contrast to some jobs that are also very exposed, such as administrative roles in finance, there are not clear pathways. That is one of the red flags that invite a robust policy response and focus on transitional support.

I will start by saying that it is really contextual. You can have AI deployment, for example, with automated vehicles, as we have just discussed. They are not listening to the employees at all. They do not need to engage them. It is not about augmentation; it is about like-for-like displacement in that context. In professional services, you are talking about gen AI. It is a different context altogether. We are seeing that AI is really broadly adopted, but it is very shallow. It is not turning up in the statistics, as the famous saying goes. There is a growing and really interesting recognition that workers have expertise, experience and insight that is necessary to define the problem that AI is meant to solve. AI is often said to be like a solution in search of a problem. Boards and executives also need training. People talk a lot about workers needing to be upskilled, but there are a lot of executives who are really drinking the Kool-Aid on this and then being like, “Where are my productivity gains?” Now there is a walk-back from that, because there has been a lot of spending on AI without any big productivity gains. There is a recognition that you need people’s buy-in. Otherwise, they are going to hide AI usage. There is this phenomenon called shadow AI or shadow IT. You cannot just have top-down mandates because you otherwise have this thing called adoption theatre, where people are being clocked up on how much they are using AI, rather than it actually solving a problem or issue. This goes back to the participation bit. If you bring in the workforce and recognise their expertise and experience in sectors where that is valued, they can help to define the problem, help to define the training that is needed, design the features, be brought into the system and flag issues as they come up. You create an intelligent organisation.

I am going to approach from a slightly different angle. I will declare two interests. I have just stopped being president of the British Chambers of Commerce and I am on the board of Multiverse. I am going to talk about SMEs because they are 85% of the economy. It is a slightly different perspective from Adam’s. I completely agree about context. At the BCC—British Chambers of Commerce—we saw two or three years ago that 84%, I think, of those small and medium-sized businesses said, “I don’t know what to do about AI.” They knew it was super-important in their priority list, but did not know what to actually do. That has now gone down to 64%, or it may now be more around 50%. That is partly because of a programme of work the BCC has been doing, which is very specific. It is helping people look at the problems in their business, as opposed to starting with the tools. It sounds like a very small shift, but in a small business people are being sold Copilot or A. N. Other licences—I am not going to pick on Microsoft particularly—and are asking, as Adam quite rightly said, “Where is my productivity change?” If you just give people a tool, that is not going to work. You have to start with the problem and redesign the process in the business. That can take some work, cultural shift and training. It is like any digital transformation that we have been through in the last couple of decades. It has to be very specific and detailed, and based around business problems, not starting with the tools and the technology, but starting with what you are trying to achieve. That requires quite a lot of intention, especially at a time when SMEs are feeling completely up against it, as we know, pretty cross with the Government, the economy and everything, feeling like regulations are everywhere, and feeling as though they have just had costs layered on top of them. The projected view in the next year is not that great: trade down, Brexit, all this stuff—you guys know way better than I do. We have to be cognisant of the fact that this is hard work, and help, including political help, that can be given to companies to do it is valuable.

You are going to have to remind me of the question.

We think that this kind of employee buy-in is pretty critical to moving this forward, but what is your view?

It is very important, yes. We can pitch it a bit higher than that. It is on the basis of analysis, not paralysis, that many different disciplines, including organisational management, social sciences and economists, show that involvement mediates better results, in the same way that the Pissarides review into work and wellbeing showed very clearly that focusing on work quality and good work mediated better results at different levels—so for the work of the firm and at a labour market level. Involvement and participation is a really core part of that. In terms of involvement, it is not simple because we are talking about different levels. We are talking about tools, jobs, skills and pathways, and then support that is in place. There is a good case for involvement in all those buckets. We have important existing infrastructures, including union involvement, which could perhaps be utilised more in this. Also, as Martha says, we need to look at it from different perspectives, and perhaps look for areas of alignment and see what is happening as best practice too. There are some examples of that, but they are few and far between. We have done eight case studies for Innovate UK of responsible AI adoption in different types of firms, including SMEs, and the public sector as well. There were pockets of good practice. There are plenty of examples of those pockets that are published in those case studies, which were done with the CIPD. There was no one case where all these different touchpoints and levels were systematically caught. We need to go big and offer guidance. We need to not be afraid of offering guidance and support, and incentives too, to reward good behaviour when people are increasing levels of involvement, work quality and so on.

That would be really helpful. Carl Frey, is there anything else you want to add into that story about the importance of employee engagement in advancing AI in a healthy and sensible way?

The most compelling evidence I can think of comes from the computer revolution, where you quite clearly see that the firms that do best in terms of productivity during this period tend to be relatively decentralised firms, where people at the lower levels of that organisation have decision-making rights. To me, that is intuitive, because when the technology is new you need experimentation. The people actually using the technology and trying to figure out the use cases are best placed to figure out what those use cases are. If they do not have decision-making rights, it will slow down adoption and implementation across the board. We are seeing, similarly, today that more decentralised firms, with fewer layers at least, are adopting AI more quickly as well. I think that the same intuition holds with AI too.

I have two things. This has now been realised more. Microsoft, in its 2025 future of work report, says that employee participation is essential for a fair and effective AI adoption. There are not loads of case studies. I guess that there is a selection bias for me at the TUC, because the problems come up, but I am sure that there are some. To add to that—and you will forgive me for banging on about Nordics—Denmark has the highest AI adoption in peer countries. We need to think about why that is. It also has a really robust economy. Obviously it is different and we cannot bring everything from there over to here. Another example is that Ford Motor Company recently rehired 300 experienced engineers, because the AI quality assurance that it introduced into its production line was not up to scratch and failed. I am paraphrasing, but the executive said, “We didn’t really appreciate the skill and expertise that those engineers had before we replaced them with AI.” They do not really understand the work that they do, especially in big companies where you are very remote from the shop. They rehired them, but they have gone on to say, “We just needed to rehire them so that they can train the machines,” and then they let them go again. People will do that, because you need the work. The big picture is that, if that happens at a macro scale, that is an extraction up of a wealth of experience and expertise that has been gained and earned, and then you get replaced by a robot. How are you participating in that? We have seen that particularly in the creative industries, where that work is being scooped up and they are replacing a lot of workers in that context.

I wanted to ask the members of the panel about the extent to which they think successful AI adoption in enterprises depends on really good leadership and management skills. Some of the evidence we have heard on this Committee to date has been less about AI and digital skills and more about the lack of skills that managers and leaders have for knowing the direction that they could take. I would like to say for the record that, before I came to this place, I worked for the TUC for the best part of a decade. That is listed in my declaration of interests.

I think that it was the Chartered Management Institute that came out with a report the other day, or a survey or some findings, saying that managers do not have the skills for change management—so, in organisational development and design, and job reconfiguration. That is the story of innovation and technology diffusion. It was not the invention of the steam engine; it was the reorganisation of the factory around new technologies. In short, yes, but, to Dan’s point, you will find the best use cases of AI being deployed effectively and responsibly where management and governance of a company or organisation is fair and effective generally. It is not exempt, separate or segregated from other good or bad things. I will leave it there, other than to reinforce the point that workers hold a wealth of information and expertise. That should be empowered and leveraged.

This relates a little bit to my earlier point about decentralisation. Knowing what you do and do not know is really important. AI is still a technology that is in an experimental phase. We still see new use cases emerging with new model releases. Trying to figure out what the technology can be used for requires experimentation and giving people the autonomy to experiment. If you experiment with something new, it is going to slow you down in other activities, so making sure that you are not punished for that is going to be really important as well. That said, there is a somewhat premature emphasis on adoption. I want to emphasise that what really matters is not the adoption rate, but what the technology is used for. If we all use AI for emails, the result of that is just going to be more emails. We are not going to see productivity gains as a result of that. If you use AI to develop new products and for research and development, that is a very different story. We need to focus much more on figuring out the use cases and product development around AI rather than emphasising driving adoption from the top down.

Martha, particularly with your BCC hat on, my manufacturing SMEs would tell me that they do not know what they do not know. We have heard some amazing evidence about the impact of AI in manufacturing settings. I would appreciate your thoughts.

To go back one question—sorry, I am coming to yours—there are so many examples that we can share offline about companies doing this. Jaguar Land Rover pops into my mind, as does John Lewis. I know them through being on the board of Multiverse, and I have declared the interest. They were making a bunch of people redundant. For the people they kept in the business, they designed how their jobs were going to look using AI and gave them AI training to do that. That was a hugely successful project. When Capita was collapsing, a new boss comes in and the first thing he says is, “Anyone in the company can redesign any bit of the company using AI. Tell me directly as the CEO.” He had a huge cascade programme. There are lots of examples and I would be happy to share them offline. While coming from a management point, of course, I completely agree with Adam. This is just about good management, full stop, and good leadership. I am not sure that there is good AI leadership. It is about recognising that you should always engage people in any change. It is a cultural shift as much as a technical ability shift. When I think back to the digital transformations that I have been personally involved in, that is always the bit that is harder, because it takes a longer time. You are going to inevitably have to shift and move roles and people. There will be change and that can be uncomfortable. I am not sure I see a specific AI piece of this. It is just about good management and good leadership. I also totally recognise that we have a challenge in helping leaders understand the art of the possible right now. That is as true in a FTSE 100 board as it is in a small or medium‑sized business and BCC member. There is work to do in helping people feel confident and like they know the art of the possible. The two are related.

One implication of this could be that we expand programmes, such as Be the Business, that are there to surge in extra management training, which is a problem that has bedevilled the British economy for a long time.

I am sure that you have talked to the CIPD, but it has a very clear thesis about how management skills have collapsed in this country and how we need to invest in managers and make management much more of an active skill, if you like. Yes, I agree.

This is only a PS that, but within that bucket of management skills, our research shows that high-involvement management is a particular bucket that needs attention. The 2,000—I think it was—firms surveyed for the Pissarides review showed that high-involvement management was a neglected, important mediating piece, so that perhaps needs attention. In our case studies, which we actually did with the CIPD, we ended up thinking about high-performance work systems and the role of proactive work and role design, which actively thought about how you could reconfigure the tasks and skills through the AI implementation process. That is perhaps also another area for attention and potentially, as Professor Frey said, experimentation, or to build into the growth labs and other initiatives, including support for SMEs.

We have heard various reports about how AI is used in job applications to filter applicants and how algorithmic management is used to hand out tasks, sometimes to set pay rates in particular sectors, and possibly also in redundancy selection exercises. Do you feel that the current regulatory environment is sufficient to deal with that range of issues?

Algorithmic management, in particular, is distinct from this broader question of task automation. You can think about algorithmic management as digitalisation or automation of the human resources employment relationship. As you know, because you participated in the TUC AI Bill working group, the short answer is that, no, we do not think that. The TUC drafted a model piece of legislation that set out how to undergird—this is coming back to a point I was going to make earlier—good culture with the right incentive structure, or, in short, the right carrots and sticks. If you rely on good culture, the good guys who want to take the high road get undercut by the ones who do not. The logistics company that does not want to rinse its employees for everything gets undercut by the ones that do, and they can use technology. We need to set a floor, which is why we have called for legislation on this point. In the Government’s Make Work Pay commitment, there was the workplace monitoring technology component. That has been released for consultation today, so we will feed into that. We are saying that that process of algorithmic management breaks down the human relations component that, as we have just discussed, is really important for effective, dignified, transparent, empowering workplaces, if people are managed predominantly by algorithm. The other bit that I would mention is the interface, increasingly, between algorithmic management and task automation, which we are predominantly talking about today. We know that, for example, workers are being asked to wear forward-facing cameras when they are working logistics or in factories, to collect training data for robotics automation. Drivers’ data has been used for automation of vehicles. It is interfacing increasingly in a way in which algorithmic management is used for algorithmic data collection in order to produce the training data for task automation. This is a new frontier that we are interested in exploring because, with large language models, the internet has been the training data basis for robotics and other things. You need to collect other bits of data. The question we always come back to is how people participate in and gain from that transition in the workplace or are supported through disruption in society. To come to your question, no, we think that there is a lot more work to be done there to raise the floor of good management practices.

While agreeing with my friend Adam here, it is still somewhat surprising to me that, in the last 30 years of my dabbling around on the edges of technology, there has not been substantive regulation, in a technological sense, of any of the new waves of technology—and I mean that very broadly—apart, arguably, from the Online Safety Act. We need to reckon with what that is going to look like in the next five to 10 years. I do not know what that answer is, but I would completely support two things. First, we should make sure that we are deploying the regulations we already have well across the things that are happening. There are a lot of things that we have already in existence and could probably use more effectively, rather than creating more. We just have to always have that backdrop. I have that small business from the BCC on my shoulder saying, “No, don’t make us do more stuff,” There should not be too much of a challenge there. Everything should be additive, be helpful and make sure that there is more productivity and growth in the economy. First is to make sure that what we have already works better. There is no question that there is something about the algorithmic opaqueness that is unhelpful generally. As a very entry-level 101 to some of the laddering that Adam has been talking about, that feels like very low-hanging fruit—that everything is transparent as opposed to opaque. How are you using it? Where is the data flowing? What is coming from you? When I am sending in my CV, what is that transaction as a human being and how do I understand that?

Would the panel generally agree that transparency is an absolute baseline for this?

It is a minimum, yes.

It is entry level.

Exactly, it is entry level. I support very much what has been said about gaps in these areas that are sat perhaps between the Employment Rights Act and the data Act. It is a difficult area. Martha is right. There has not been a substantive piece of legislation in this area, and the rest of the world is thinking about it, including most recently South Korea. Even Singapore, I think, has started thinking about it. Perhaps it may be a different overarching piece. Our work suggests that it may be forward-looking, anticipatory and built around higher levels of information sharing, reciprocity and consultation right across the AI value chain and job cycle. It may be something that can respect rights, of course, but that is also principles‑based and built up and deployed over time, and that sits with existing things and with sector‑specific and regulator‑specific guidance. That will need time.

Let me add a footnote to what has already been said. I think that this is going to be an area of increasing tension going forward. AI algorithms basically already have access to everything that has been written on the internet. What has not been incorporated yet is much of the tacit knowledge that exists in the workplace. There are going to be some very significant benefits to the firms that are able to codify that knowledge and bring it into their AI tools, not least in terms of fusing AI with advanced robotics. Algorithmic management is going to be very much front and centre of those developments. That will create some very significant tensions between AI progress and privacy going forward.

I am wondering whether we should move on to the point about bargaining power and impact on wages.

You have anticipated my question, Chair. An issue that has come up over recent years is the use of algorithms to set pay rates, particularly in the gig economy. The transparency issue is certainly part of that. What could we do in terms of regulation to give workers the information, never mind the bargaining power, to understand how their pay rates have been set?

It would be worth weaving into this your view on what may happen to pay rates. Professor Frey, do you want to kick off?

I do not have a strong view on that particular question, to be honest with you.

On algorithmic pay within the platform economy and other places, if you do not have pay transparency, you cannot bargain on pay. It is as simple as that. It is the foundational component of bargaining as an individual when you go to accept a job or as a collective through some formal representative mechanism. There is not transparency in the gig economy, where dynamic pay is used and algorithms set pay, including giving different rates of pay for the same job to different drivers or workers. Our position is that it should be banned. It is not necessary. It does not do anything other than scoop up more extractive profit for gig economy platform operators. Most fundamentally, a lot of our employment rights flow through status as an employee. We know that gig economy workers are not employees. If they are not employees, it is really difficult for them to have rights through the law or exercise their rights through collective mechanisms such as unions or other forms of workplace representation.

Do you think the minimum wage regulations are fit for purpose for this type of pay?

I am not sure that I could authoritatively answer on that question. I will say something briefly about the other component parts, outside of the gig economy, which are obviously in different places. When workers, as individuals or a collective, bargain for pay, they are bargaining time and skill, or some variation of that. We know that AI can uplift people’s skills in a certain context, because people can access new information. They can access, as Carl mentioned, new content, including tearing down barriers to information that has usually been the preserve of cloistered professions. We also know that it can have a deskilling component. Anthropic, the makers of Claude, did a study. It had two sets of computer programmers and gave them a task. One used AI and the other did not. It found that the ones who used AI had a 17%—this is Anthropic’s metric—decline in skill mastery. That is the process of cognitive offloading. You have given more of the value of what has been produced into the machine. Therefore you have been deskilled, which means that, as an individual or as a group of workers, you have less collective bargaining power when you go into the workplace. That is a process that has been happening for quite a while. The question for us, which brings us to the skills question and other things, is how we make sure that AI is deployed in a way that uplifts people’s skills, rather than atomising, degrading or cognitively offloading their skills and therefore diminishing their individual and collective bargaining power. This is the final thing, because I have been talking for a while. One thing we are seeing is wage stratification or wage polarisation—for example, for computer programmers, which is the only place where AI adoption has been really significant and systematised. Experienced programmers are seeing, broadly, significant increases in pay because they are able to boost their productivity by using coding assistants and AI agents. They can do more than they were able to do. We are seeing a contraction in hiring of juniors. There is less demand because a single worker is doing more. There is a productivity boost, but a labour market impact parallel to that.

Adam, you were first to mention analysis paralysis, but I think you would exclude from that the Government’s ability to monitor what is going on with AI, to look at the workplace and to see the changes that are coming. I do not know whether you are familiar with the AI and future of work unit, which is tasked with this. I wonder how you think the Government are placed to monitor accurately these changes in real time and looking ahead.

The TUC has some participation in that through Mike Clancy, the general secretary of Prospect union, who is the TUC representative on it. Overall, we think that that is really important and a good thing. The Government should be gathering the best insight, expertise and evidence, and that should be informing policy, but I will go back to the point that we cannot anticipate so much. As humans, we are quite bad at predictions, so we need to be continually updating our information. We also need to be making judgment calls. We need to have a no-regrets framework, as I called it. Overall, it is a positive thing, absolutely.

Baroness Lane-Fox, it seems fairly clear that AI is a great servant and a bad master. You are asking us to be quite bold about how we embrace this. This is a huge question, but there is a balance to be struck here. We have to help people through the transition; we also have to harness the benefits of AI. What is your message to us, as legislators, about this? What can we do in terms of policy to get that balance right and to harness the potential benefits?

Easy-peasy. I am going to start slightly differently, in that there is a danger in where we are right now, with the journey of AI and us culturally in Britain, that we have ignored some of the things that might be important and overplayed some of the things that might not. I am going to pick a quite specific example that I hope wraps up some of the things we have been talking about. If you look at young people’s fear of AI, it is disproportionate to other bits of the population, for understandable reasons. They are sending 45 CVs that they may have made themselves using AI, but they are all being algorithmically chucked out at them. There is a disconnect between their view of the world, with what they are seeing and how AI is affecting them personally, and what we all probably believe is going to happen and be productive and helpful in the economy of the future. That is a very profound challenge. If we have a whole group of young people—so, both the Milburn cohort and another group of young people who may well be skilled right now and may be trying to find work but are unable to—that is a real social cohesion point. I would approach from a slightly different perspective. I would say, a bit to Adam’s point, “What are the no-regrets things with the cohorts of people right now? What are the messages that we as Government need to land about this stuff?” We need to make sure that the UK is a place that is not only building the AI sector, which we have done incredibly effectively. We have done an amazing job at becoming the third biggest place for investment into the AI sector, after silicon valley and China. It is remarkable. London and the UK should be really proud of what we have been building. That is a lot to do with the DeepMind diaspora and all the things that I am sure you have thought about a lot. In parallel to that, we need to think about how we are going to message what we are going to do for the broader economy and specific chunks of it. That is one piece of it that may sound slightly flaky, but it is not meant to. It is actually the quite deep, difficult work of how Government join up all the different bits of business support and all the different growth messages to always continually land that this is the direction of travel. This is the way we are going. These are the things. When we make an intervention, we are going to do it because of an AI-first world. I am not going to go back and re-legislate stuff, but the employment law that just went through was not designed for an AI-first world. We need to be really cognisant of what we are doing, with intention, in this building all the time. Are we really building for 2030, for 2025 or, even worse, for 2020? Maybe we are not really even building for it at all. That is the first thing. We need to upskill as many of the people in this building as we can to what we all believe is a shared view of what this future could be. Pick off some cohorts that are in, I would argue, more danger right now. Entry-level jobs is one of those places. That is not just to do with AI. It is clearly to do with the macroeconomy as well. There is horrible mix going on right now with a whole bunch of things together. The fundamental underpinning of all this stuff is that all the programmes that Government do should be questioning whether they are designed for an AI-first world. Are we really helping people? When people go into a jobcentre, are they sitting opposite a desk with someone tapping away on the computer telling them something, or are they sitting behind the desk with the person, co-designing their CV, able to use the tools themselves and see what the different roles, and what the transitions and pathways, which Anna talks about so well, might be? I am not avoiding your question, but it is not one thing. It is about every single signal that Government can send and about having a really joined-up and collective view.

You are clear that Government have to take the lead in this.

Yes, 100%. We are doing aspects of it, but I do not believe that it is joined up, urgent or at scale enough.

I am afraid that that takes us slightly over time, so we are going to have to wrap up there. This has been an incredibly useful panel. Thank you so much indeed. That has changed the way I think about a number of the questions that are at stake in this inquiry. Thank you very much for the evidence. Please follow up if you have case studies or things that you think we should reflect on while we write this report over the summer. I think I can say with some confidence that this will be the first of a number of reports we run over the course of this Parliament.