112. Safe and Moral AI with Rebecca Raper

Ben Byford:[00:00:09]

This is our second interview with Rebecca, the first one being episode 44, Moral Machines. This was recorded on the 6th of May, 2026. This time me and Rebecca were talking about LLMs and whether they're actually safe. Adding human values to AI, AI governance, guardrails, moral assurance, the lack of interdisciplinary work in robotics, whether AI should be sold as the solution to everything, the apparent sidelining of AI ethics, and the fact that AI might progress the widening inequality of people.

If you'd like to find more episodes, you can go to machine-ethics.net. You can contact us, hello@machine-ethics.net. If you'd like to follow us, you can go to bluesky machine-ethics. Machine-ethics.net, Instagram: machine-ethics-podcast, YouTube: machine-ethics. And if you can, you can support us on Patreon: Patreon.com/machineethics. Thanks again and hope you enjoy.

Hi Rebecca, welcome to the podcast for the second time. And we had you back in episode 44, which seems like a long time ago now that we're on 111. Um, so thanks for coming back on. Um, and a lot's changed since 2020, but again, also some things have completely not changed and, uh, some subjects and interesting things are still, um, whirling around. And, um, before we get to that stuff, I guess if you could just introduce yourself, who you are and what do you do?

Rebecca Raper:[00:01:52]

I'm a robotics lecturer at Cranfield University, and for the past 8 years or so now, I've been doing research at the intersection of philosophy, robotics, cognitive science, basically looking at moral machines.

Awesome. Great. And I think we last time, we, we always asked this question about what AI is. I don't know if you've had any changes in that kind of appreciation for how you think about it now? Because obviously 2020, it feels like a long time ago now.

Sure. So I must admit I can't remember what definition I gave before, but I can, I guess, working as a lecturer now, I can try and give the explanation I give to students. Yeah. So basically when we talk about artificial intelligence, We're talking about— the name kind of gives it away, but how do we artificially create intelligence? What I want to avoid doing is saying that artificial intelligence is just— and I know this is quite a popular narrative to say that it's just machine learning, it's just, um, types of, um, technology or mathematics that's been implemented. So I, I view it as the science of trying to replicate intelligence. I think with that in mind, intelligence isn't everything, and we can have something that's super intelligent, we can have a super powerful computer, a super powerful calculator, but if it's not ethical, if the decisions and its outputs aren't conducive to creating outputs that are beneficial to people, then that's going to be problematic. So I think the attempt to try and design and ensure that machines are safe, that the outputs are ethical and conducive to people, is really important.

Yeah, I mean, it's interesting that you could have a machine that just is— I think this is kind of the traditional before, before times idea of what an intelligent machine would be like, isn't it? You know, something which crunches the numbers and gives you a result and doesn't have maybe a moral intelligence or a social intelligence in the same way that we expect humans to have some sort of social empathetic component to their actions. And it's interesting that I feel like in popular culture...That's what we thought we would get. And I think I've talked about this before, but we, and it's funny that we don't have that almost. We don't have these calculating machines intelligences. I, as the way I see it, but we have something kind of in between where it's kind of like muddling, trying to do this thing and it's getting all this information and just kind of doing some stuff, but it's not necessarily correct. In, in factual sense. Yeah, but also it doesn't necessarily have this moral, reflective, skin-in-the-game intelligence that we also might want it to have, or at least that's the way that I see it. Again, is that kind of like what we're talking about at the moment?

I think so. Um, I think what you're identifying is the distinction potentially between ethical machines, so machines that are designed so that they embody ethical values to some extent, but then the decision-making outputs of a machine. With automation generally, if you think about something being driven by automation, a process being driven by, by an artificial intelligence system, how do we ensure that the outputs of that process are ethical? And I know there's a lot of discussion about whether that, that second criteria that I've just mentioned is something that we should be researching. Since working on part 1 of my PhD and the podcast, I think I've come to the realisation that it's less about embodying, trying to replicate human morality, and more actually about just ensuring safety assurance.

So it's, it's, uh, less about almost making a functioning socially intelligent thing, whatever that might be. Maybe with that, like, um, so I'm referring to your book a bit here, which is Raising Robots to Be Good, um, and in there you kind of talk about the idea that maybe this stuff requires some other stuff that we don't necessarily have or necessarily know we have, like some sort of autonomy, some sort of personality, to get to a point where we can say that this, this system is a moral system because it has these capabilities. And also we can somehow test it or assure it in some way, which is a whole kind of can of worms, right? But what you're saying is that, you know, nominally we want it to be safe, right? And it feels funny saying that that's like the minimum bar, right? But it feels like we're still not there. Like this, you know, safety obviously is a spectrum and we feel like viscerally it feels like we're very far away from being acceptably safe on that spectrum for various reasons. Um, obviously not all these systems are the same. They are different and they have different capabilities, and some of them are image systems, some of them are text systems, etc., etc.

So there's lots that encompass this word AI that people throw around these terms with. But if we're talking specifically about some of the large language models that you might chat with, they are more or less not safe, essentially. And part of that safety is probably, you know, a moral issue, right? They impact us existentially, right? Yeah. They can hurt us. They— I hate these words, right? We haven't got very good semantics for talking about AIs, unfortunately.

I think maybe they can influence us. Maybe that's a term that might be helpful.

Yes, yeah, exactly. So the influence on us is varying and can be catastrophic. And that's just one of the things that we're seeing, as well as environmental stuff, consolidation of power, money, you know, bias, all these things come together. How do you feel about that safety issue in terms of our current situation?

Yeah, I think, to be honest, it's a really pressing issue. Um, I think there's a lot of work— I know there's a lot of work being done in the safety domain, but it really needs to reframe itself as, um, I think part of the problem with the discussions that are taking place at the moment is they're looking at it from maybe a human-centric perspective, and they say— they're seeing how can we get our human values that something— that X is good, and how can we put that into a machine. I think it needs to be looked at from another way, and actually how do we ensure that the systems are safe rather than trying to instil human values. There's so many complexities that are involved in when you're trying to put human values into a machine. One of the key ones is that it's inherently—that approach is biassed

to the system of values itself.

Yeah, or the system of values that you're selecting.

Um, do you not think though that, um, sorry to interrupt, but do you not think that we already do that in the data itself? So there is, for example, again, if we're talking about one of these large language models like Gemini, um, OpenAI's ChatGPT, etc., that they have a set of data and then a certain regime of different things, different reinforcement learning from human feedback and different processes that they train the models on after the fact. And part of that is selecting the data and evaluating in a certain direction, right? So, part of that is going to influence what you get out, right? So, you are kind of implicitly applying values essentially at that point. So, are you talking about a kind of a separate way of work, a separate thing, or a separate way of working to talk about values?

Yeah, I think, so, when I'm talking about values, I'm talking, when I was talking about it before, I was referring to when you're trying to align a system. So, the AI alignment technique, for example, looks at how you can create an AI system so that it's aligned to human values. That's what I take issue with, because I think what you end up doing is almost— you have a risk of prejudicing on a particular set of human values, and you can see that that quite obviously can be problematic. I think what you're referring to would be more bias, right, that's inherent in a, in an LLM, for example, um, where the data used to train a machine learning— well, to train a model is not necessarily representative of the people that it will ultimately implicate.

Yeah, so it's less about the specific values you want to imbue, but more like the stuff is picked up from the data, which is biassed, but a different kind of bias essentially.

Yeah. So, so you're wondering where, I guess, whether it's the same. I think they're different issues.

I guess my, I think the way that I think of it is you get the same outcome. So if you did none of the post-training, you'd still have a model which has a certain value, right? Because of how it is created. You just haven't been explicit in trying to make it have a different value, essentially.

Um, what, what I, I'm— I, I think is you can— it can be value-driven from both directions, or prejudice and value system can be driven on both sides.

Yeah, and I guess if you, um, suggesting that maybe this is extremely problematic because we have to know and select for these things, um, what is, what is the safety kind of direction for that, and how does that differ? Because you were mentioning that it's, it's, it's always going to be problematic, um, selecting for moral values or ethical values, so, yeah, and that we should nominally be selecting, or—

I think so. Something I looked at in my book is actually how we guide the— how we shape the decision-making in the first place. And by shape, I mean how we might govern that decision-making. So rather than if post-training. Yeah, you, you— what you're doing, I guess, by putting a value system on top is almost, um, it feels like kind of patching an AI system. You're assuming this AI system is, is okay, and I'm patching it by trying to put some value system on top. What I think we should be doing is ultimately having a human in the loop at the pre-training stage, um, at the initial training stage, and having a governance mechanism to shape how those decisions are made generally.

And do you have an example of one of those decisions that it might be an issue?

So with an LLM For example, deciding what information to present, or if you have, um, just, just a general AI system giving you recommendations, that might— the prioritisation of that. So rather than actually saying, oh, it's recommending films that aren't appropriate to certain people, or it's recommending, um, content that is causing harm to some extent. Yeah, rather than fixing that retrospectively, actually we need to look at the— how it's making those recommendations in the first place.

And, and is that like a context— give more context for that, um, or has it changed the model?

Do you think it changed the outputs, the decision outputs?

Yeah. Do you not feel like that's like, I mean, given what we have currently, that would completely change the game. And what we might get is models that are, you know, the most safe but also not very capable because they can't recommend adult films maybe, or they can't, um, the systems aren't able to, uh, do some actions, if you like?

No, because you're not preventing certain behaviours or certain outputs. What you're doing is you're governing them.

Okay, but how, how does that governance mechanism work?

Yeah, but this is— I think this is an area that needs work, right? But ultimately, I mean, so if you think about films when, when I was younger, when films came out of the cinema, they had an age rating. Yeah, that's a simple mechanism to govern films to ensure that people of a certain age can only watch, um, age-appropriate things. Yeah, I think what needs to be discussed, and this is something that needs kind of opening up is how do we put those— a similar— that's quite obviously quite a simple governance framework. How do we put governance frameworks in so that the outputs are appropriate for whatever demographic it's impacting? In the, in the case of Recommender...

Right, right. So I mean, so for example, if you were going to make a, a child's film recommendation website, let's say, and you were backing it with an LLM, um, or some AI thing, right, um, you could, you could somehow assure that it's, it's not going to recommend certain stuff, right? Um, so the, so the actual issue is if you're going to a general system and asking it things, then how do you have the context to govern any single question under the sun, right? So it seems to me that you're actually against the general systems, but, uh, have a design essentially to fix the less general systems.

No, I don't think so. I think actually it's the other way around. I think that, that governance, it's almost like a pre-training governance framework, and that, that what that would do would allow...It essentially shape the decision-making of any AI system. And a problem we have at the moment is that we've got really powerful artificial intelligence that's kind of bringing outputs, and all those outputs need vetting. Essentially, that's how the— it's, um, how it's been managed, because it's happening after the, the training's already taken place. So what you're having to do is almost go through and put guardrails or against every single or every dubious output. And you can't, you can't catch every— everything by doing it that way, especially as AI systems become more and more powerful.

Which is why there's, I guess, lots of discussions around the need to halt AI development at the moment because we're not really in control of that.

Yeah, I mean, I, uh, I have a similar feeling that it would be nice if people— and, you know, they've called for this in the past, some of these, um, top scientists, top people. Some of those companies have asked almost, but it feels like we're in this race, race to the bottom, in terms of, um, making some sort of all-powerful um, general intelligence situation, which may or may not be possible.

Um, so for me, it feels somewhat intractable, what you're suggesting, um, in terms of what we have currently to be able to— because it feels like you have to have so much context. And the flip side to that is that sometimes you don't want the guardrails. So I've got this weird thing that I always think when, when we first had, was it GPT-2, whatever, you know, there was very few guardrails and you could ask it stuff and it would tell you this is, you know, it's not, it was not very useful. It wasn't very nice. It would say weird stuff all the time. But, you know, if you asked it how to build a bomb or whatever, or some sort of taboo subject, it would, it wouldn't take very much for it to tell you something useful.

And what we have now is obviously not that, right? Which is good for a general audience, but actually less good for a non-general audience. You know, if you were a researcher and you wanted to get to more of a base model and be like, no, I actually, I don't want you to be kind to me. I don't want you to give me something, you know, in a nice flowery language. I just want you to smash these two ideas together because I know you have them in there. And they're taboo subjects, and I'm a researcher or journalist or whatever. And it seems to me, unless you're in one of these companies, that actually that doesn't exist anymore in the same way, unless maybe there's some open source models that I haven't tried this on. So it's funny that you have some circumstances where you might want that stuff and some circumstances where you don't.

Yeah, I agree with what you're saying. So I think that's what the problem is with, I guess you might call, the guard-railed approach. But I'm suggesting instead of— so if the guard-railed approach is almost like policing the outputs of an a... What you have to do is if you do that is you need essentially more and more, as the systems become more powerful, you need more police.

Yeah, yeah.

And you don't always want to police things like you say, because that to some extent hinders the benefit of these systems in the first place. So actually what you need is a governance framework, pre-training, some kind to assure the safety of the outputs. And that's kind of the reframing that I try to address in my book.

Yeah, and you go into what that might look like. So, I mean, did you want to speak a little bit about the kind of the moral assurance? Because I think that's— it's a thing that maybe doesn't get talked about or thought about much, that we have this place that we want to get to, right? That we, you know, we have all these tests for mathematical capability, general reasoning, all sorts of stuff that we might apply and, and find an answer. Um, but we also have this idea that we might want to have some sort of moral, ethical, you know, assurance, reasoning, stats.

Yeah, it's, um, it's linked to the safety, um, ultimately. So this is where my thinking's come on a bit since my book. But by moral assurance, what I'm saying is that ensuring that the decisions, the decision outputs of an AI system such as an LLM, but this can extend as well to, and it becomes even more important actually when you start to think of AI agents, an AI agency, machine agency more generally. So what we want to do is assure that the decisions made by those agents are ultimately in line with ethics.

And then having some way of testing that to know that that's the case.

So what I propose in my book is basically some kind of framework for developing a morally assured system. And then in the same way you might in society, you might almost check that something is morally assured before letting it out. It's through various mechanisms to check it there as well.

Yeah, and I think you were mentioning that, because obviously you work in Robotics Department now, and you are interested in how the things change if you apply some of this to the robotics situation or the embodied situation. So do you have a kind of introduction to how that differs, or what that— what's going on there when you were talking about more intelligent robots?

Yes, so obviously we have the AI alignment problem, which is linked to more software-based systems. What I'm increasingly seeing is artificial intelligence, and particularly LLMs, embedded into physical machines. So I know, um, I think there's a Gemini version of a robot where you basically say, fetch me the— just a robot arm on a table, and you say, fetch me the blue cube, and it'll, using an LLM-based decision architecture, fetch the blue cube. Now, that's interesting. That gives a capability to machines and robots that I think is...allows them to perform tasks a bit easier. Or if we go back to the AI assurance, you can see that maybe some of the risks transfer over into that domain. Yeah, when they're putting that kind of decision-making into a physical machine.

If you imagine, for example, a machine that has to fetch red bricks, but that machine's got bias, um, what it'll end up doing is that bias will be amplified in embodied behaviour. So it might think that— trying to think something that might be problematic, um, because something's square it's a— and it's red, it's a red box. So maybe you train your robot to, um, put all red boxes in the bin.

Yeah, just so happens that, um, your favourite toy, your um, something really important to you, really treasured, looks like a red box. And because of bias the machine identifies it as a red box. Yeah, you, it starts to venture into physical safety there as well.

Yeah, I think for me it feels like this, there's so many, uh, with like an agent system, like let's say agent's been kind of taking on this other thing, but like a robot with agency, it has a system which it's making decisions of what to do, whatever, and it can do those things. But you might want it to do certain things, and, you know, otherwise it's not gonna, it's not gonna do anything, right? It's gonna sit there and just await, or it has some, you know, default programming that says, if you're not doing anything, do these things, maybe clean up, whatever. But cleanup maybe doesn't incorporate— I think there's some examples in the literature about, you know, cleaning up the pets, or like, you know, not really being aware of maybe the capability to be aware of enough things might not be there, which is a problem. But also maybe that it doesn't actually know about certain things or know that certain things are categorised in certain ways. So cleaning up might incorporate finding a bit of, I don't know, thrown away bit of paper or tin foil or something and putting that in the bin, and then putting all your books in the bin, and then putting all your pens, and you know what I mean?

The cleaning up is, in an embodied sense, still a very hard problem. You know, there's lots of things to say there and going on which are interesting, and I feel like that's one of the reasons we haven't got robots in our in our houses and workplaces so much at the moment because there are so many issues and so many things that can go wrong after the basic, the robot can move and actually action things and has the capability to do useful work. So, yeah, I mean, there's the moral situation. It feels like still quite far away from us, even though we're talking about it. But it will catch up soon. I'm sure it will.

See, I'm not, I'm not so sure actually. Um, I mean, I'm not so sure that there is a lot of work being done on, in, in robotics, for example, where they are putting AI into machines. Um, and maybe they're not publicly acceptable so that we don't see humanoids in our houses, but they are being used in other domains. And I think the people— part of the problem is the people working in that area aren't aware of the AI safety risks, right?

So the people making and researching how to put different capabilities into the into embodied systems are maybe not being as cross-disciplinary as they could be.

Yeah, there's a lot of— in my experience, I see a lot of people almost just kind of taking a packaged AI system and putting that into a machine.

Yeah, yeah, yeah, because it's— it must be the simplest option almost, I guess.

I think it's hard to be knowledgeable about everything.

Yeah, yeah, yeah.

And that's why, that's why interdisciplinary, um, interdisciplinary work is so important and why I think people in different areas obviously need to be talking to each other.

Yeah, yeah, yeah. Have you seen some of the issues first-hand or, or know about some things that, uh, driven you to this realisation?

Yeah, I, so I'm a robotics lecturer and I do, I'm actually, I teach, I teach an apprenticeship degree in robotics, so which means I do a lot of work with industry. And when, when it comes to industrial robotics, it's worth noting that this is a field that has been going since like the '50s They've had robot arms building cars, automated solutions. So they're a bit naive, but AI's kind of come along and they're taking it into factories, I think. So there's definitely a piece on public awareness of AI safety, I think, or general awareness across everywhere.

Yeah, and especially if you're working with those systems already, essentially.

Yeah, you mean the people working— I don't— I guess the problem is if you don't know about it, if you don't know that you need to know something.

Yeah, that's quite hard.

Yeah, because you're being presented to it, presented to this, to the thing as a something that is happening that must be safe or whatever, otherwise would— it wouldn't be happening, right?

Yeah, I think safety is definitely a consideration within engineering if we're thinking about, um, robotics engineering. The problem is AI is— this is why I take issue— AI has been sold, and particularly given recent, um, it's been sold as kind of a solution to everything.

And you mean specifically the large language models or multimodal models?

Not just large language models. It's been sold as, um, a solution to the economy, to jobs, and I think the risks associated with it aren't— I don't think everyone's aware of those.

Yeah, yeah, yeah. And obviously if you're listening to this podcast, then you should already have a very good idea of like the breadth of those risks and, you know, the many, many episodes we've talked to people talking about lots of different things in different areas that, you know, either being thought about or we need to think more about different places. And do you think, is there something that can like really irks you at the moment? Um, for me, it's— it feels like the environmental stuff. But is there something that you think about at nighttime and go, oh, I wish we could just sort that one out?

Um, just what I kind of said, really. The way in which AI's been pushed. What, what is kind of is worrying me is AI ethics has almost been pushed aside for economic purposes. And that's kind of what seems to be driving everybody at the moment. Obviously it's an issue, obviously cost of living is an issue, but AI on its own without those boundaries is not— to solve.

Isn't AI supposed to solve everything, all the problems?

And there need to be more public conversations and awareness pieces about it. And I think it's actually the responsibility of everyone working in across AI ethics and machine ethics and, um, to almost shout about that.

Yeah, I feel like, um, similarly, um, I haven't written it yet, but I think I was thinking of writing an article or something, uh, click— like, like a clickbaity title, right? Um, like AI ethics is dead and capitalism killed it or something like that. Because I think I'm just, um, agreeing with you that we, we're trying to do good work and we're trying to point out the issues, but everyone's sidelining the issues because of essentially the money that's been spent, but also the the outcomes which we haven't necessarily got are still so big and alluring still to people.

Yeah, you mean the potential benefits of AI are alluring?

Yeah, I mean, I feel like maybe the potential benefits, but the financial benefits maybe are the thing that's more alluring, right?

it's still unclear whether the. I have like this weighing scales situation, right? And like, um, it's a bit like any technology, right? If it's, if it's, um, the benefits are outweighed by the negatives, then you should seriously question why is it being used, right? And it feels like that scale is, you know, everyone has their own answer to this situation of whether it's being outweighed, but it feels like to me that we've definitely working with a technology, let's say large-scale version of AI, the large language models, all this sort of stuff, which maybe the negatives are outweighing the positives at the moment, especially for people who maybe aren't benefiting financially from the situation. And it seems unclear to me how they will financially benefit from the situation. At all, you know what I mean? So, um, so it feels like, um, you know, maybe we need to sort that out, uh, you know, as well. Um, I think we, you know, we've discussed a little bit about the financial incentives, but, um,

I think, um, there's a risk that actually with AI, what it ends up doing is causing a monopoly of the people who have AI, which means that the difference between people who are rich and poor actually widens. So if AI is only— if you think of it as— if AI is only in the way that things, I guess, are disseminated, if it's only introduced to, say, big corporations without sufficient ethical guidance in place, it could just end up benefiting even more the big corporations rather than things like communities.

And yeah, so in the next 5, 10 years, um, is there an idea that you have a view on maybe what we should be doing and what the best outcome might look like, or what would be really cool to see?

Yeah, I think obviously AI and robots in the, in of themselves and technology is really cool. We just need to be careful or considerate with how it's applied. So in terms of a view for the future, that's what I think.

Yeah, and I'm guessing that implies that it's not being very well done at the moment.

I think there's probably a bit of over-excitement about AI.

Yeah, I feel like I've got— I've definitely got AI overload at the moment. I'm just like, you know, some of us who've been in AI a long time, I mean, I've been talking about it on the podcast for a a long time, for like 10 years, but I haven't been in academia that long, like not that long, sorry. I haven't been in academia for 20 years talking about AI, you know, some people have, for example, but now that everyone's talking about it, I'm like, could everyone stop talking about it? Ah!

Yeah, maybe, and I mean, I've not been in AI as long as you, but maybe it's time for a brief winter or something.

The AI winter is coming. Well, I mean, if anyone believes in the kind of the bubble bursting, then maybe this is the year for that. We still haven't seen it yet, as we were talking about at the beginning of the year in one of the podcasts, but it still might come because the financial stuff isn't necessarily working out still. And maybe, maybe that's the rest that we need to actually take stock, apply things where they can be applied safer, better, for more useful things. And I can stop having to talk to everyone about AI in the pub and have everyone have an opinion about it, and we can all get on with our lives.

Yes, that sounds good. Maybe, um, Think of other things that are more fun.

Yes. Yeah, exactly. I'd like more things. That would be great. More fun things. Well, with that kind of strange and hopeful, I want to say hopeful ending. Thank you very much, Rebecca, for coming on for the second time to the podcast. If people want to find you, follow you, talk to you, how do they do that?

I'm on both Twitter and LinkedIn. You can find me just by searching my name, um, or also look up my book, which you can get online as well, Raising Robots to Be Good.

Awesome. Yep, check it out. And, uh, yeah, thanks very much for your time.

Thank you.

Hello and welcome to the end of the podcast. Thanks again for Rebecca for coming on for her second time on the podcast. One of the great things about talking to Rebecca and a few of the people I've had on the podcast is that there are very few people in what we termed machine ethics. I know the podcast is called the Machine Ethics Podcast, but the academic field of machine ethics. So it's always great to talk to Rebecca and others about moral machines, essentially, or adding morality to machines. I still find it strange and sort of abstract how people are talking about guardrails, governance within AI at the moment, and I'm not sure how that brings us all the solutions. It feels like that's what people are trying to do, bringing the solutionizing with governance and guardrails, and it feels like they're not actually dealing with the problem head-on itself. Maybe that's just a personal opinion or feeling that I have at the moment. There's a lot of new terms going out, harness, AI harnessing, harnesses, sorry, the use of governance instead of the use of what was responsible AI in the first place in business and things like that.

I think there needs to be an acknowledgement of what is the capabilities of the situation of the AI services we have at the moment and, you know, using them in safe ways given those capabilities. And I think part of that is, you know, all the AI ethics or all the AI research that we've done over the last, you know, 10, 15 years, um, culminating and going, it's really good at these things, it's not so good at these other things, it's practical to use it here, it needs much more safety controls over here. But again, that comes back to what Rebecca were talking about, about the, the hype the massive financial pit that is AI at the moment and solutionizing everything with AI. Maybe that isn't in itself the solution.

Thanks again, and if you can, you can support us on Patreon, Patreon.com/machineethics. You can always get hold of us, hello@machine-ethics.net. Thanks very much, and I'll speak to you next time.

112. Safe and Moral AI with Rebecca Raper

Transcription:

Episode host: Ben Byford