How AI gets stress-tested before release

0:00Tom Lue:A CNBC original podcast. Hello and welcome to The Tech Download, a new CNBC original podcast where we unpack the tech stories that matter most.

0:12Dawn Bloxwich:Each season we dive into one big theme and what it means for your money with insights from the industry's most influential voices. This season we're looking at Google DeepMind, the powerhouse driving the tech giant's AI push.

0:25Tom Lue:Ever wondered what would happen if AI got out of control or fell into the wrong hands? And how can governments make sure it's safe for you and I? Our guests for this episode, DeepMind's Senior Director for Responsible Development and Innovation, Dawn Blockswitch, and Tom Liu, VP for Frontier AI Global Affairs, addressed these questions. When we think about responsible AI at Google DeepMind, we're not seeing it as something where it's slowing down progress. Instead, it's about us thoughtfully designing it from the very, very beginning.

1:23Tom Lue:in on this idea of responsibility when it comes to developing the assistance. And just to set this scene, there is a growing chorus of voices that are talking about the risks of AI and particularly fears around AI getting out of control as it gets more powerful, that the makers of these AI systems won't even be able to control them, the societal impact of AI when it comes to people losing jobs and how it's going to impact all of us. And of course, we are placing at this point in time a huge amount of trust in these technology companies, these giant technology companies with vast pockets. And there's very few of them developing this technology to get this right when it comes to the safety and the responsibility of developing these systems.

2:11Tom Lue:And that, I think, is a really profound thought when you think about it, Steve.

2:16Dawn Bloxwich:Here in the United States, the regulatory system is way different than it is over in the EU. And companies here kind of have the benefit of lax regulation. We went through this in the Web 2.0 world and the social media world. Nothing happened. We saw all the problems related to election interference, Cambridge Analytica, you know, all sorts of terrible things happening on YouTube. and no regulation happened. So when you say, you know, we have to put our trust in these companies and the people running them, that's because our, at least here in the United States, our government has shown a lack of desire to really solve these issues.

2:56Dawn Bloxwich:So these AI companies, they have the benefit of just going at it. And, you know, the EU, I know you can speak to this a little bit, they're working on stuff in a more serious way. But still, they got some ways to go. And in the meantime, this is evolving faster than regulators can even keep up.

3:14Tom Lue:Absolutely. And that's the point. The regulators are always playing catch up because this technology is developing so fast. I mean, the tech industry will say, you know, we want the regulation, but we want it to be thoughtful and we want it to be adaptable. And the regulators will say, well, yeah, we need regulation, but they also are trying to themselves get to grip with what this technology is. So it's a difficult position, but these are the kind of topics we're going to unpack in this episode, both from how do you develop these AI systems responsibly from within, but also how does the regulatory approach fit in here and what kind of regulation could be brought forward?

3:50Tom Lue:We've got two super interesting guests and I want to kick off with the first one, Dawn Blockswitch.

3:59Dawn Bloxwich:executive decisions is the new podcast from cnbc where i ask powerful leaders about their decisions that changed everything i'm steve sedgwick here's miss joe malone cbe i started

4:10Tom Lue:that first business of skincare that's when i knew that i was in charge of my own life and that's when the entrepreneur really although i didn't know what the word entrepreneur meant that's when the entrepreneur really took hold that's executive decisions with me steve sedgwick get it wherever you're listening to this don't it'd be great to just kick off with um i guess sort of trying to define what we mean by responsible development what are the kind of things you're thinking about we want to build ai to to benefit the world and humanity that's that's our mission right so for us what that means is being really thoughtful about how we're building and how our AI is being used.

4:50Tom Lue:So our approach always takes into account the AI principles. It's guided by that and it centers around these principles of responsible governance, responsible research and responsible impact. So when we think about responsible AI at Google DeepMind, we're not seeing it as something where it's slowing down progress or it's about hitting the brakes. Instead, it's about us thoughtfully designing it from the very, very beginning, considering it at all aspects and all parts of the development journey. And when I talk about risks too, we think about quite a broad spectrum of risks. So you would have seen our frontier safety framework.

5:32Tom Lue:It thinks about the most severe types of risks, like chemical and biological risks or cyber risks. But we're also thinking about near-term risks as well. So whether that's bias and inequality, because they're really important. And we want to think about these things as a holistic set of risks and opportunities. And we do see them as very connected. So that's why we don't want to just consider one or the other in isolation. They should be considered as a whole and as something that we are considering in tandem and in parallel. Some would say Google DeepMind's goal to build AI that benefits humanity.

6:10Tom Lue:And as part of that artificial general intelligence, this idea of this extremely powerful AI systems, that there's this kind of tension between being able to achieve AGI and these powerful AI systems and doing so responsibly, that there is always this kind of tension there. And actually, can the two really coexist? Yeah, I do strongly believe that they coexist, and they're not in competition. And going back to the point I made around designing this thoughtfully in from the beginning. If you do that, it's really well considered and it must be done like now, as opposed to like, we don't want to get to AGI and go, hmm, we should now be thinking about these types of risks, but rather think about it now so that we're getting ahead.

6:53Tom Lue:A good example of this is the work that we did in AlphaFold, which I realize is not necessarily an AGI example, but with AlphaFold, we embedded in the team for many, many years, working through with them and helping them anticipate some of the opportunities and risks. And that meant that we helped sort of shape it from the very, very sort of like beginning. And at the end, we launched it and it was very, very successful and actually saw the benefits that we wanted to see come out of it for society. And I think that is the process that we want to continue replicating, even as we're getting to AGI.

7:28Tom Lue:The unfortunate thing, I guess, about unintended consequences is that we may not know what those unintended consequences might be. How comfortable do you feel about your level of knowing what the risks are at this point? I think we have a good grip on some of the biggest risks. That goes back to the frontier safety framework because these are the big ones that we have considered, but also we've seen considered across the industry. So we know that there is some alignment in terms of the way that people are thinking about that. But we also need to be very mindful and thoughtful about monitoring as well.

8:03Tom Lue:So looking at exactly how the models are being used and in specific use cases so that we can then say, right, based on this, we didn't see this one coming. And so actually we need to do, we need to make some changes and some adjustments. And the other big, I think, debate and concern right now is around misinformation and hallucinations with some of these AI systems. And, you know, granted, using them, they've got a lot better, a lot, lot better, certainly from, you know, when they were first emerging and getting very popular with users. But again, its accuracy of these systems is not 100%. So what's the kind of feedback mechanism for correcting and improving upon the misinformation that comes out of the systems that DeepMind's building?

8:46Tom Lue:Yeah, I think hallucinations are one of those unintended consequences, I think, of the creativity of the models. And so when we're trying to think about how to address that or where it may be appearing, again, we go back to what we can see the users doing. And users will flag things to us if they are seeing issues, but we'll also sort of monitor via the logs. We also have a number of different initiatives that we're putting into place to basically address things like factuality so that we're making sure that we're grounding in the right types of information. Dole, can you help me understand the term, red teaming?

9:24Tom Lue:Red teaming is a really important part of our approach. And so, like, everybody has slightly different definitions, so I'll give you mine. In our world, what it means is an unstructured sort of exploratory way of testing our models. And we complement that actually with structured evaluations. So, data sets of evaluations that we would run as standard on a model. But the red teaming side means we want to have people who either are real experts in their fields or are wonderful at gel-breaking. And we want to see them really go in and explore the models and see what they can find. And then they share that with us.

10:02Tom Lue:We will then make changes, if appropriate. And then we can also then turn them into structured evaluations. So it's a really critical part of how we operate. So we often hear these sort of stories about people, quote unquote, tricking AI systems through prompts to give an answer that perhaps was not intended by the creators of these AI systems. And so is that something similar to what happens in the unstructured side? Yes. So we will have people testing for those jailbreaks. So what novel ways can they get the system to say things that we would rather that the system doesn't? Or it can be in lots of different ways.

10:39Tom Lue:Like it can be trying to see whether specific information can be surfaced from a model. So it's all about, I think, creativity and exploring new ways. And every time a new capability comes out, so a model sort of has a new ability, we will then want to test where are the boundaries of that? And are there any risks that we see associated with that? And that red teaming plays a part there. How does the fact that there is such intense competition commercially with DeepMind and OpenAI and other competitors in this field play into what you do? Because you need to develop quickly, but you need to develop safely.

11:15Tom Lue:Safety and speed, they are really necessary parts of a whole. We do it together. And we know that the competitive pressures are there, and it drives a lot of progress as well. But it can also mean that people may be tempted to cut corners. But I think that we want to make sure that we are very balanced. We want to keep our focus on the safety and responsibility aspects. We do have and we have created forums. So with OpenAI, Microsoft and Anthropic, we created the Frontier Model Forum to specifically discuss topics like safety and help us better align on best practices, which I think has been really helpful.

11:57Tom Lue:So when things are moving fast, that we have a way that we understand what everybody is trying to do. And we have a joint view on risks, which I think is good. Dawn, we've spoken about, I guess, some of the more immediate risks. You were talking about some of the things you have to think about longer term. What are those for you? Yeah. So the two big buckets of things that we're really thinking about, particularly when we think about the agentic era coming into play, is thinking about the frontier safety framework risks, but also some of the socio-effective risks. So the frontier safety framework includes CBRN, so that's the chem, bio, radiological, cyber, harmful manipulation, which we think is very important to consider, and then things like loss of control and deceptive alignment.

12:44Tom Lue:These are all things, very broad categories, of which there will be lots of aspects to explore within that. But the other area of socio-effective, which is about that social and emotional ways that the model connects with users, that to us is going to be a very important area to understand over the coming year and years, I would say. So that includes things like delusions. It includes things like relationships and companionship. So it's going to contain quite a lot of different topics that we're going to need to explore, which we'll also need to do in collaboration with a number of different groups, including third parties, I think, civil society, academia, because we're going to need to agree on some standards there.

13:33Tom Lue:And there's some really tricky questions that we're going to have to deal with. I suspect you and your team are thinking about how you approach some of the questions around transparency, around working with third parties. And I do just want to talk about some of those parts as well, because there was a letter earlier this year, I'm sure you saw from Paul's AI. And one of the criticisms and allegations they made was that Google DeepMind wasn't living up to the commitments it made in the AI summit in Seoul. And the crux of it was that when Google released Gemini 2.5 Pro, they said there was no safety evaluation that kind of accompanied it.

14:07Tom Lue:I know Google and DeepMind have sort of responded to that, but there is this kind of broader view, I think, and it speaks to that there are some groups who believe AI labs should be more transparent with the data going into their models, how they're training their models, and all of these other parts. How are you thinking about this idea of transparency? AI is moving so fast. We are seeing it as we've been talking about it in so many different parts of our lives, so in knowledge, in productivity, in creativity. It's just everywhere. And I think in some of these discussions, we will find that there are going to be some conversations and topics and debates about certain topics.

14:46Tom Lue:And I think that is a good thing to have. From a transparency perspective, we're very thoughtful about releasing information in our tech reports, in our model cards to make sure that users, whether they're customers, enterprise customers or developers, have visibility of the testing that we have done. and where we are confident and where we may have questions, but to make sure that they have visibility of the safety testing and the testing that we have done holistically. And we were one of the first companies to release standardized reports in this way. And we think that that is something that is going to continue to be important.

15:25Tom Lue:We want people to understand more about the models. And we want also for people to rely on the information that we're providing. Are you sort of constantly debating how much you should release or can release even, just given some of those considerations around, well, we need to also protect our competitive edge? Yes, there will always be that balance with the competitive aspects and that will go on, I suspect. but I think that we are always aiming to be as transparent as we can particularly from the safety and responsibility perspective because we want people to understand this is the testing that we have done in frontier safety areas this is the testing that we're doing you know around bias and hallucinations and we we we want that sort of transparency and trust.

16:13Tom Lue:Dawn how long have you been at DeepMind now? Nearly eight years seven and a half nearly eight years and you previously Obviously, we're at Salesforce, IBM, and in PwC. What sort of made you jump into this world of AI? I've always thought that AI sounded like an amazing space to be in. The possibilities are huge to be able to support humanity, to think about so many different complex problems which we've not been able to tackle by ourselves. So, you know, this aspect of sustainability, things around medicine and health I think there's so much more we need to understand and so when this role came up to work at DeepMind and in the area of ethics and society I was like this is just something I couldn't say no to and very excited that I am still here.

17:05Tom Lue:I think the bigger question is why did you leave Australia to come to rainy Britain I mean that's my my bigger question. It's a question I ask myself a lot I'm gonna be honest. Dawn the other the other part of of the equation here is you know you have internal processes you have internal teams you have ai principles you have in a lot of internal systems in place as you try to develop these ai systems and the other question is how do you work with third parties because then there'll be a cohort of people say well this is just deep mind grading its own homework effectively so where is that checks and balances we really value independence as a part of our overall evaluations approach so So we have our model development teams doing their own testing.

17:48Tom Lue:We do testing. But also we then have these third parties. We've been doing this for years is that we wanted to be able to have like an independent view so that we could identify any of the unknown unknowns. So we've engaged with a number of different providers, very sort of like specialist providers, to try and help us understand more about what they're seeing in the models. And that will be something that we continue to do. And we are, from a transparency perspective, sharing more on in our reports. AI labs around the world are sort of coming up with their own set of principles and standards as well.

18:24Tom Lue:But do you think there needs to be something that's more collective, more standardized globally? I do think that it's important to be having these conversations, which is why we did have the Frontier Model Forum. because it's exactly that, that we want to be able to talk about those standards and how we are calibrating across the labs on the types of risks that we're looking at, how we're actually addressing those, because we don't want this sort of like completely jagged view of like, you know, we value something or we're very concerned about something, but another lab isn't. So those conversations are very, very important.

18:59Tom Lue:We really value them and we will continue to do them. Do you think, just based on the conversations you've had with your counterparts and some of the other companies, there is a genuine collective sense of responsibility? Because I think that's the other concern from the general public, is that there's these very powerful companies, very rich companies developing these systems that we are all going to be using. Can we trust them? I think people from across the safety community, regardless of lab or what company they're working for, deeply care about this area. They're very passionate about it and they've been thinking about it for years.

19:38Tom Lue:I do think that there is a genuine good intention from these groups and a strong desire to work together. I think the safety groups are an invaluable part of how we're developing models and I think they will continue to be. How confident are you now that when AGI is here, that it can be controlled? We are going to continue to build safety and responsibility into everything that we do. And we've done it and we will continue to do it. So I think AGI is, in some ways, we don't know exactly what it's going to look like. All we can do really is continue to apply a very scientific approach and have a rigorous approach to how we're thinking about safety and responsibility and looking at each new capability that comes down the line and addressing them then and there.

20:32Tom Lue:And I think that will continue to be this sort of like practical way that we can actually address those questions well in advance of getting to AGI. So you're confident sort of that kind of approach will allow you to be in control of those systems. Yeah. Dawn, that was such a great insight into what's going on here at Google DeepMind. I appreciate your time. Thank you. Thank you.

20:55Tom Lue:So it was really cool to get an insight there, I thought, into the way DeepMind is approaching the development of the AI systems. It sounds like from the very start, as they're developing these models, they're thinking of all of these potential risks that could come out of them. And there's stress testing. I asked Dawn there about this idea of red teaming, you know, effectively trying to find loopholes in these systems. But the other tension I thought that was quite interesting here was this tension between tech companies wanting to ship product quickly because it is, of course, an incredibly intense competitive environment, but also needing to do so safely.

21:35Dawn Bloxwich:Right. And this reminded me so much of the conversations we're having with about another Google company several years ago. And that's YouTube. So around 2015 to 2017-ish, let's call it, YouTube was going through a lot of problems. And for so long, the messaging out of the company was, you know, we're an open platform, let anything go. You know, we're not the arbiters of truth. And it took them so long to realize when you have billions of users watching billions of hours of video a month, you're going to run into some of these issues. And you kind of do have a responsibility to control what's going on out there.

22:12Dawn Bloxwich:If not for like just being a good global citizen, advertisers don't like it. And so it was really interesting to see this idea that, look, they're thinking about it. They're doing it. They're red teaming it. It sounds like they're trying to stress test Gemini in a way they weren't doing in the early days of YouTube. So that was refreshing to hear. Are there going to be problems? Yes. Would they admit there are going to be problems? Yes. We can go on right now and find those issues. But the fact that they're doing it and thinking about it and talking about it right now is quite refreshing.

22:45Tom Lue:I said at the start of the episode, we're kind of having to put some sort of trust into these companies to get this right, and particularly on the responsibility side. And so we want to know, you know, what they're doing. And we want to know that they're at least having a good go at trying to make these systems safe from the start. You know, that is one part of the equation. The other part is, how do companies like DeepMind and Google more broadly work with the regulators when it comes to thinking about policy around AI? That's a conversation I had with Tom Liu.

23:20Tom Lue:Tom, let's just kick off, I guess, with the global regulatory landscape a little bit, because it's fair to say regulators around the world are still figuring out AI and what to do with it, how to regulate it, how to legislate it potentially. In the EU, there's the AI Act. The US and UK are taking different approaches here as well. So it's quite fragmented. What's the operational challenge for you and for DeepMind as you deal with this kind of fragmentation?