Original Broadcaster Captioning (Enhanced). Treat it as a searchable index of what was broadcast, not a quotation record.
00:00:00Now on BBC News. Join our I decoded decoded team as we unpack, explore
00:00:03explore and take a deep dive into
00:00:05into the world of artificial intelligence.
00:00:22Hello again. Welcome to AI
00:00:24Decoded. $2.5 trillion. That is what
00:00:28what the world will spend on artificial artificial intelligence this year.
00:00:31year. That's up nearly 50% on last
00:00:33last year. But what is the return
00:00:36return for that investment? Every Every boardroom has been told the
00:00:38the same story. This will transform transform productivity, cut costs,
00:00:42costs, change the way that we all
00:00:44all work. A year ago, the research
00:00:47research suggested that wasn't happening. happening. Barely any of it was showing showing up on the balance sheet.
00:00:51sheet. So where are we now? Well, Well, the most recent studies show
00:00:55show there have been improvements improvements in some areas. Give
00:00:58Give the AI a clearly defined task. task. Write code, answer customer
00:01:02customer queries. The systems are
00:01:04are measurably faster, but most businesses
00:01:07businesses are a chain and all too
00:01:08too often speeding up one link. Just
00:01:11Just parse up the work at the next. next. So how do we solve that? With
00:01:14With us this week, Peter Grant, CEO
00:01:18CEO of go.ai, and James Coote, who who is the founder of pair, both
00:01:21both spend their working lives inside inside companies trying to close close this gap. So hopefully they'll
00:01:25they'll be able to answer the question. question. And with us this week,
00:01:28week, our co-host and author of Technology Technology Is Not Neutral. Doctor
00:01:32Doctor Stephanie Hare is here to to welcome to you all. Peter, let
00:01:35let me start with you. Going back
00:01:38back to the MIT study a year ago,
00:01:42ago, 95% of these projects were delivering
00:01:46delivering nothing measurable. Is
00:01:47Is that changing? I think it is slowly
00:01:51slowly starting to change, but I I do think that's a canary in the
00:01:54the coal mine. I think that report report has come under some criticism, criticism, Christian, because it
00:01:57it was a very small, sample. It was
00:02:01was taken of 153 companies, coming coming back from them. But a more
00:02:05more recent report by PwC that was was released at Davos this year,
00:02:08year, the World Economic Forum was
00:02:10was across 4500 CEOs in 95 countries. countries. But that found that 56%
00:02:1456% of them are yet to see a return return on investment from the investments
00:02:18investments they've made in generative generative AI, and only 1 in 8 can
00:02:21can actually claim a cost saving
00:02:24saving or revenue generation. I think think the short answer is one is
00:02:28is trust, two is complexity and three
00:02:30three is speed. Enterprises are very very different to how consumers buy
00:02:34buy generative AI. So is it failing
00:02:37failing because companies are buying buying things technologies faster
00:02:40faster than they know how to use
00:02:42use them? Or do they know exactly exactly what they want and can't
00:02:46can't make it work? I think it's
00:02:48it's probably the latter, but go go back to trust. So these closed
00:02:52closed models, they want data. They're They're like a plant and imagine
00:02:55imagine the plants, the algorithm algorithm and water is what feeds
00:02:58feeds the plant. And the issue is is that they need this data. They
00:03:01They need this information. But enterprises, enterprises, their IP is basically
00:03:05basically locked inside their four four walls. And they need to protect protect that to make sure those organisations
00:03:09organisations don't have that information. information. So they're very cautious cautious about how they implement
00:03:13implement generative AI. But also also trust comes down to the answers answers that you're getting from
00:03:17from the agents as well. So as we we know these systems hallucinate
00:03:21hallucinate in the consumer space, space, that's fine. But if you're you're a bank, if you're lending
00:03:24lending money, if you're in pharmaceuticals, pharmaceuticals, you're doing research research and everything else, you
00:03:28you have to be correct in what you're you're doing. So that's why organisations organisations are cautious. That's
00:03:32That's why it's taking a bit more more time as well. But I think there's
00:03:35there's a mis set expectation. If If you think Google took five years
00:03:38years to get to 100 million users,
00:03:40users, ChatGPT took 60 days. That's That's 30 times faster than anything
00:03:44anything we've ever seen before. before. So I think organisations
00:03:47organisations are struggling with with that to really digest it and
00:03:51and make it meaningful in their companies. companies. Yes, shareholders are
00:03:55are an impatient mob. James, I know
00:03:57know you're going to tell me, that
00:04:00that technology is not the hard part. part. We're already at PhD level
00:04:04level systems. It's about adoption. adoption. Yeah. I mean, we've never
00:04:07never had a challenge like this before before when suddenly we've got PhD
00:04:11PhD level intelligence in almost almost every subject, but occasionally
00:04:14occasionally lies on tap. And it's it's very difficult for your average
00:04:18average employee in an organisation organisation to use those tools to
00:04:21to their full potential. And right right now almost everyone has access.
00:04:24access. They've all got co-pilot, co-pilot, they've got ChatGPT or or they're using Claude, or maybe
00:04:28maybe their boss doesn't know, but but very few of them are adopting
00:04:32adopting it meaningfully, meaningfully meaningfully and actually using these,
00:04:35these, these tools to their capability. capability. And that's what we call
00:04:38call the adoption gap, right? And And actually it's big and growing
00:04:41growing bigger as the frontier of of models progresses. What do you you think is the problem? Stefanie?
00:04:44Stefanie? Do you think it's a technology technology problem or a human behaviour behaviour problem? So first of all,
00:04:48all, it's not enough to just give give people a tool and say work it
00:04:51it out. Most people are not receiving receiving the kind of training that that you would need to make the most
00:04:55most of these tools. And that's quite quite scary if you think about it, it, given that it is everywhere,
00:04:58everywhere, it's in hospitals. So So you've got doctors who are maybe
00:05:01maybe chatting with you and using
00:05:04using ChatGPT along with the patients
00:05:06patients to discuss symptoms, check
00:05:09check diagnoses, check treatments, treatments, not going through the the ethics of that, the governance
00:05:12governance of that. Is it safe to to upload your medical results and
00:05:16and tests into these models or not? not? They hallucinate so they just
00:05:19just get things wrong. They also also just lie. They just make things
00:05:23things up and have to be challenged. challenged. It all depends on how
00:05:25how you prompt this art of prompting. prompting. People don't do training
00:05:29training because training is a cost. cost. So that's one reason that I I think we're not seeing the value
00:05:32value yet. But second is I think think we see value in things like
00:05:36like software development or customer customer service, but we're not seeing
00:05:39seeing it necessarily throughout throughout the entire organisation organisation because you might speed
00:05:42speed up productivity in one part part of your business, but that doesn't doesn't mean that. Can you give me
00:05:46me an example? Where would you see see within that chain that I described?
00:05:50described? What would you see a speeding speeding up of the business processes
00:05:53processes and where would you see see the bottleneck? So let's look look at the business of healthcare.
00:05:57healthcare. So you go into a hospital, hospital, everybody can understand understand radiology, right? So the
00:06:00the people are looking at your skin.
00:06:02skin. So it's your X-rays or MRIs,
00:06:04MRIs, Cat scans just because you you are able to use AI to process
00:06:08process scans and read them faster faster to look for a bone break or
00:06:11or a tumour doesn't mean that you're you're necessarily going to cut patient
00:06:14patient waiting lists or get faster faster treatment for people because
00:06:18because the skin is just one of many, many, many tasks, all of which have
00:06:21have to be signed off. Healthcare Healthcare is highly regulated by by doctors the whole way through.
00:06:25through. Well, those doctors aren't aren't getting more time just because because the scans are being read
00:06:29read faster. So those red scans pile pile up, but everybody still has
00:06:33has to then put them through the the same old processes. I think we'll
00:06:36we'll have to rethink a lot of workflows workflows from the ground up because
00:06:39because of AI. It's time on the side side of some of these companies,
00:06:43companies, Peter, because $2.5 trillion trillion in total spend this year
00:06:45year is an awful lot of money. What What I'm not sure about is what this
00:06:48this looks like on the balance sheet. sheet. And how do you measure it
00:06:51it for shareholders? Yeah, we like
00:06:53like to call it an ROI, but return return on intelligence rather than
00:06:57than return on investment. And that's that's really for every dollar or
00:06:59or pound that you spend, what measurable
00:07:02measurable business value you're you're actually getting back. So
00:07:04So we've created a framework basically basically for enterprises to be able
00:07:07able to measure this. So the money money they're spending, how do you
00:07:10you measure it for vectors basically basically are you starting are you
00:07:14you experimenting today all the way way through to are you the agentic agentic enterprise where humans and
00:07:18and agents are basically working working together? What's your strategy?
00:07:21strategy? Who owns that strategy? strategy? Who's what's the budget budget you're doing to fund it? And
00:07:25And are you actually doing a proper proper ROI analysis before you fund
00:07:28fund any of those particular workflows workflows that Stephanie was talking
00:07:31talking about? What is your data? data? What's the state of it? How
00:07:34How do you get access to the data? data? Can you get to the system of
00:07:36of record for the ROI? So your CRM, CRM, your customer relationship management
00:07:40management systems, your ERP systems, systems, what's the technology you're
00:07:43you're using? So have you just let let everybody use it like James was
00:07:46was talking about co-pilot perplexity, perplexity, OpenAI, whatever, or
00:07:49or is it structured your approach approach by team or by organisation,
00:07:52organisation, the governance that that you have a governance is, is is super important as well, especially
00:07:56especially when you talk about SDA, SDA, which is zero data retention
00:07:58retention to make sure it's not being being leaked outside, you have control control over that. Some people have
00:08:02have been doing some famous cases cases about that and then workforce
00:08:06workforce literacy, as Stephanie Stephanie was quite rightly saying, saying, have you trained and have
00:08:10have you enabled people? You wouldn't wouldn't start at a bank and say, say, OK, go and do the lending straight
00:08:14straight away. You train people, people, you enable them, you certify certify them, you make sure they're
00:08:16they're an ambassador for your brand brand before they talk to customers. customers. You need to do exactly
00:08:20exactly the same with AI as well. well. And then the final probably
00:08:23probably the most important thing thing is what's in pilot and what's
00:08:26what's actually in production. And And what is that ratio. And we're
00:08:29we're seeing at the moment it's about about 30% in our index from those
00:08:32those people that are actually say say got ten pilots running, only
00:08:34only 30 are actually running production production in your workflows. Right? Right? Well, when it comes to pilots,
00:08:38pilots, I mean, obviously that is is quantifiable. You can see that
00:08:41that on a balance sheet. But there
00:08:43there are some processes and some
00:08:46some things you might do using AI, AI, which are not as quantifiable.
00:08:50quantifiable. James. Yeah. I mean, mean, we were talking about this this just before, one of the things
00:08:54things that AI has been most valuable valuable for me on in the recent
00:08:58recent past was helping us think think through and then execute on
00:09:01on our US expansion as a business. business. It's quite a complex thing thing to do. I'd never done it before.
00:09:05before. I spoke to five different different people about it. They give give five different opinions. Probably
00:09:08Probably some of them would be classed classed as a hallucination if an
00:09:11an AI gave it. But AI was really
00:09:13really helpful in market mapping, mapping, working out our messaging messaging over there, working out
00:09:17out what we have to do on the payroll payroll side and share side. And
00:09:20And it's very difficult to isolate
00:09:23isolate the sort of one bit of ROI
00:09:25ROI AI gave us there, but it's unarguable
00:09:28unarguable for me and for, you know, know, my leadership team. That there's
00:09:31there's been an unbelievable enabler enabler to helping us get started
00:09:34started over in the US. So that was was one where I thought it doesn't
00:09:38doesn't neatly fit, you know, a project. project. There's not just one part
00:09:40part it's helped with. It's helped helped with the ideation part. It's
00:09:43It's helped with planning and operations. operations. It's helped with actually
00:09:46actually going off and, you know, know, helping us make some of the the first sales over there. And I
00:09:50I don't know whether it fits neatly neatly into a project box, but good
00:09:53good God, it's been valuable. OK, OK, well, maybe then the problem problem isn't technology at all.
00:09:57all. Maybe it's the assumption assumption assumption that you can buy a tool tool and immediately expect people
00:10:00people to change how they work. So So let's talk about the workforce
00:10:04workforce because this is where you you come in. James. Are businesses
00:10:07businesses spending more on software software than they are on training?
00:10:11training? The people who use it? it? And what does that training need
00:10:13need to look like? Undoubtedly. I I mean, there's lots of stats on
00:10:17on this. The latest I've heard is
00:10:20is that three times as many people people in British businesses have
00:10:23have access to an AI tool versus versus those who have access to AI
00:10:27AI training. So most people are getting
00:10:29getting access to AI with no training. training. So that's undoubtable.
00:10:33undoubtable. And the thing that people people are often missing with the the training that they do deliver
00:10:37deliver when they deliver it is it it focuses very much on how do you you use the tool? Well, you're right
00:10:41right in the box and you click enter enter or maybe you speak to it. It's It's pretty easy to actually use
00:10:45use it. The thing which is difficult difficult and is at the crux of successful
00:10:48successful training on using these these tools is helping people cross
00:10:52cross this imagination gap between between I know how to write in the
00:10:56the box and I know how to use this
00:10:58this tool to reinvent how I actually actually work and use my time to
00:11:02to deliver value, not just, you know,
00:11:04know, do things faster and push a a bottleneck onto someone else. And
00:11:08And I think bridging this imagination imagination gap between I know how
00:11:11how to use a tool and I know how how to actually use it to re-imagine
00:11:14re-imagine my role and do more valuable valuable things is the thing that
00:11:16that people really need to, to, to to get across to deliver value from
00:11:20from training. So where does that that stem from, do you think, Peter?
00:11:24Peter? Is it top down the senior
00:11:28senior executives need to encourage encourage AI adoption within the
00:11:30the workforce or does it, as I suggested,
00:11:33suggested, grow organically? Interesting
00:11:36Interesting question. If you think think about I've gone through many
00:11:40many different technology revolutions, revolutions, shall we say, in 30
00:11:4430 years in the industry and what what because I mentioned speed earlier,
00:11:47earlier, the speed that this has has come is unprecedented. What most
00:11:50most companies have been they've they've been lazy with regards to to change management. So people process
00:11:54process technology and they've just just given people the keys to the the car without teaching them to
00:11:58to drive. What I find the most successful
00:12:00successful ones are where the CEO, CEO, where she leads the initiative
00:12:04initiative and she leads by example example with her own agents and she
00:12:06she shows how she's using them with with her executive team. And that
00:12:10that permeates down to get people people really encouraged. Well, if if she's doing it, I need to do it
00:12:14it as well. Because what that creates creates a more permissive environment
00:12:17environment within the company. Yeah, Yeah, exactly. And culturally it it says this is a good thing to do.
00:12:21do. This is the organisation that that we're going to be. This is part part of our strategy, our vision
00:12:25vision and everything else. We do do believe this technology can help help us. It's not a technology issue.
00:12:28issue. It's more than enablement enablement side and the art of the the possible. That's role specific.
00:12:32specific. How I can do this in my
00:12:34my job. So we have a customer called
00:12:37called Mimecast 2500 people. The The CEO led that. We went in and
00:12:40and trained them with James's business business and it permeated down and
00:12:43and what we the way the evidence evidence that we see, it's working.
00:12:46working. Christian is basically in in queries how they're querying the the agents, how they're using them
00:12:50them day in, day out, whether they're they're being sales in marketing,
00:12:53marketing, in writing code or in in procurement for assessing companies
00:12:56companies or many different use cases. cases. But their usage has been off
00:12:59off the charts. I can understand
00:13:04understand the permissive environment environment within a company allows allows people to take hold of the
00:13:07the AI and run with it. Is there there any resistance? Because let's
00:13:10let's face it out there, there is
00:13:12is a lot of animosity towards AI AI and the replacement of jobs. Do
00:13:16Do we take it for granted that people people want to use it? I think people
00:13:20people are using it more and more more in their personal lives, and
00:13:22and that's often sort of the gateway gateway into this, right? I used used it to plan my holiday or to
00:13:26to find a deal and that kind of helps
00:13:29helps you build a bit of muscle memory
00:13:31memory for in your work. I think think secondly, it helps when, you
00:13:35you know, leaders model the behaviours behaviours they want to see. As Peter's
00:13:37Peter's just said, when a leader leader shows this is normal, this
00:13:40this is OK. I use this. It's not
00:13:43not dangerous if used in the right right way within our guardrails and
00:13:46and actually we're going to celebrate celebrate good use of this if it
00:13:49it leads to value. And I think thirdly, thirdly, what happens what we generally
00:13:52generally see is as you start to, to, you know, build some momentum momentum and more and more people
00:13:56people around you start to build
00:13:58build agents and Barry from procurement, procurement, did you hear he's built
00:14:02built an agent? I thought Barry was was about to take VR. You know, the
00:14:06the social sort of. Well, can you you give me. Some examples of that? that? Some fun examples of companies
00:14:09companies you've been into where where you've seen that happen? Yeah, Yeah, I'll link it to Peter's point
00:14:13point actually around execs really really grasping the tools. So there's
00:14:16there's a CEO of a manufacturing manufacturing company in Coventry. Coventry. It's called Graham. He's
00:14:20He's fab. I didn't ask him about about this shout out. So hope you
00:14:22you don't mind. But anyway he's he's he's absolutely massively into AI
00:14:26AI and he's using lovable to build build prototypes of how he'd like
00:14:29like to see the future of consultancy consultancy delivered from his business.
00:14:32business. And he's also using Claude Claude and experimenting with that.
00:14:36that. And I think through showing showing his team where he sees the
00:14:40the vision of the future in a way way which is tangible and shows a
00:14:42a bit of what's possible in the future.
00:14:44future. But does it safely with synthetic
00:14:46synthetic data and prototypes? Initially, Initially, I think is pretty exciting
00:14:50exciting and I think it's. Graham. Graham. Oh, I wouldn't want to guess.
00:14:53guess. Now don't put me on the spot.
00:14:55spot. 40s, seconds. Oh, that's. Late
00:14:58Late 40s you see we we. We sort of of the reason I ask is because there's
00:15:02there's a whole new policy that the the government is setting out today
00:15:04today that in England that they want
00:15:07want to educate children at the age age of 14 on AI that is being used
00:15:10used by local companies. I wonder wonder if you think, Stephanie, that that is the way we should be going.
00:15:14going. So we introduce children to
00:15:16to it at a very early age. So my
00:15:19my view is I would really like to to see us get kids reading books books the way they need to be before
00:15:23before we start having them mess mess around with AI. I think there's
00:15:25there's a time and place for AI in in education, but I'm really worried
00:15:29worried actually, that we're raising raising a generation of children children who don't have the critical
00:15:33critical thinking skills that come come from the old skill set, the
00:15:35the traditional skill set of basic basic literacy, numeracy and ability
00:15:39ability to write. So before we these
00:15:42these are really powerful tools that that affect the cognitive abilities
00:15:45abilities of adults with PhDs who
00:15:47who were raised in an analogue era.
00:15:50era. OK, so we put them onto children children whose brains are still forming,
00:15:54forming, who don't have that reading reading habit built in, who aren't
00:15:56aren't used to writing and learning learning how to express ideas and
00:16:00and thinking they won't have what what we all have, which is decades
00:16:03decades of muscle memory and critical critical expertise to be able to
00:16:06to look at an AI response and know
00:16:09know it's wrong, know it's hallucinating. hallucinating. So I kind of feel
00:16:12feel like, can we just give the kids kids a break? The people who work work in this are both nodding, but
00:16:15but I mean, you go into companies companies are young employees skilled
00:16:18skilled for the job. Whilst I agree
00:16:22agree with Stephanie on you cannot
00:16:25cannot shortcut expertise on developing developing critical thinking. I think
00:16:28think that is separate from whether whether young people entering the
00:16:32the workforce should be proficient proficient at using AI. And I think
00:16:35think that is unarguably a good thing thing and something which is lacking
00:16:38lacking right now. And I think an an 18-year-old who leaves school
00:16:41school with good critical thinking
00:16:44thinking skills, who is developing developing expertise but who is also
00:16:48also AI fluent, is going to be extraordinary. extraordinary. And unfortunately unfortunately there's not many of
00:16:52of them. Well, all of that brings brings us to one organisation that
00:16:55that tried to test every single thing thing that we just discussed in a
00:16:58a classroom in New York. Salamanca Salamanca is a very small rural district
00:17:01district in western New York on Seneca
00:17:04Seneca Nation territory. And they've they've just announced that they've
00:17:07they've bought a humanoid robot called
00:17:09called Sally. Get it? Sally Salamanca,
00:17:12Salamanca, Silicon skin, moving facial facial expressions, Western New York
00:17:15York accents. So she's going to fit
00:17:17fit in really well, for $57,000,
00:17:20$57,000, which I guess is a lot less less than the salary of a single single teacher. Anyway, she's been
00:17:23been brought in to assist in coding
00:17:26coding with AI classes from September September alongside an AI teaching
00:17:29teaching assistant. But then the the teachers got wind of it and the
00:17:33the unions got wind of it. And on on Friday the district paused the
00:17:36the whole thing. Now there are supporters supporters who say it will prepare
00:17:39prepare children for the world they're they're going to inherit. And then
00:17:42then there are the teachers who many many of the parents who say there
00:17:46there is no substitute for real face-to-face
00:17:49face-to-face interaction. Why do
00:17:51do you think, Peter, that story in
00:17:53in Salamanca makes people so
00:17:58angry? I don't know, and I think think it's quite sad that it does does really. I think this is a bit
00:18:01bit of fun, right. And I think to to James's point, the earlier you you can introduce AI to children
00:18:05children and help and support them them to get them ready for the workplace,
00:18:08workplace, I think it's really important. important. We're moving into a world
00:18:11world where we will control agents, agents, we will become managers of
00:18:14of agents that understand business business processes, but they're always always there's always going to be
00:18:18be a human in the loop, but they're they're going to take away some of
00:18:21of the mundane tasks. If you can can teach children how to do that. that. But I also agree with Stephanie.
00:18:24Stephanie. They need to understand understand the basics of grammar, grammar, arithmetic and everything
00:18:28everything else as well. Absolutely. Absolutely. 100%. But why not introduce
00:18:31introduce it? I think it's a bit bit of fun. I think if it engages
00:18:34engages children more it helps educate educate them and it makes learning
00:18:38learning fun that I'm very supportive. supportive. I was thinking when I I was reading this, isn't this what
00:18:41what Priya does? She takes robots robots into companies, didn't she?
00:18:44she? And then to school classrooms classrooms with Century Tech and
00:18:47and she teaches them how to how to to use these robots. But I mean,
00:18:50mean, is it the technology these these people are objecting to, Stephanie Stephanie or is it the symbolism
00:18:54symbolism of a robot standing where where a teacher ordinarily would?
00:18:57would? Oh, my God, where to even even start with this education policy
00:19:00policy in the United States in 30s. 30s. So first of all, we know that
00:19:04that again, American children are are doing worse on their reading
00:19:08reading comprehension, their writing writing abilities and their numeracy. numeracy. So we're not doing the
00:19:12the basic three R's right. We have have the data on that. It's not an
00:19:15an AI problem. It's been going for for at least the past 20 years, and and it's not unique to the United
00:19:19United States. But my point is we
00:19:21we have a problem. Houston. OK, then
00:19:24then the teachers unions are of course course incredibly powerful and active
00:19:26active in the United States is probably probably elsewhere. They can stop
00:19:29stop things. They can go on strike. strike. It's a nightmare. So they
00:19:32they probably feel quite disrespected, disrespected, worried about their
00:19:34their jobs. And then I think there's there's parents who are saying, can
00:19:38can we please just teach the kids? kids? I agree that we want we want
00:19:42want them to get exposed to AI as as we want them exposed to everything.
00:19:45everything. I'm not sure a $57,000
00:19:48$57,000 robot is the best return.
00:19:50return. OK,. Audience question. In
00:19:53In fact, the prize for the best question question of the week goes to Thomas
00:19:56Thomas who said this AI systems are are secured against the threats we
00:20:00we already know about. So what is
00:20:01is to stop an advanced AI finding
00:20:04finding entirely new ways to exploit exploit systems that no human has
00:20:08has ever anticipated, particularly particularly if it's pursuing its
00:20:11its own objectives? Very timely question,
00:20:14question, because Stephanie OpenAI OpenAI admitted in the past week week two of its models had escaped
00:20:17escaped a sealed testing environment, environment, found a zero day, got
00:20:21got onto the internet and hacked
00:20:23hacked Hugging Face, an AI hosting hosting platform in order to steal
00:20:26steal the answers and cheat the system,
00:20:29system, thus pursuing its own objectives. objectives. So Thomas's question
00:20:32question is absolutely right what what do we do about that? I mean,
00:20:36mean, there's nothing that we can can do to prevent that in the sense
00:20:38sense of that. That is a risk. It's It's always been a risk. It's a risk
00:20:41risk it will happen again. I'm amazed amazed that we're not hearing more more about it. What's interesting interesting in the OpenAI Hugging
00:20:45Hugging Face case is that both companies companies are open and transparent.
00:20:49transparent. What is worrying about about it as well is that could have have really caused serious harm.
00:20:53harm. I mean, we're lucky it was was Hugging Face, frankly, that was
00:20:57was hacked. And I'm sure if that that had a financial repercussion, repercussion, not a power plant they
00:21:01they would want. Exactly. They would would want or a hospital again, where where all of a sudden, you know,
00:21:05know, you've got machines going down down mid operation. You see, I was
00:21:07was rushing to this question from from our top conversation because
00:21:10because we're talking about rolling rolling out AI through the company.
00:21:13company. Right. And I want to know, know, Peter, if you are thinking
00:21:17thinking about energetic system within within your company and you're dealing
00:21:20dealing with very sensitive information, information, this same problem applies
00:21:23applies to that, doesn't it? You You might create a sandbox environment,
00:21:27environment, a testing environment environment within the company, but
00:21:29but if it's dealing with very sensitive sensitive information, what is to
00:21:32to guarantee that it doesn't leak?
00:21:34leak? Well. I think I'm not an AI AI research scientist Christian.
00:21:38Christian. So one caveat the answer.
00:21:40answer. But this is basically what's what's called rainbow teaming where
00:21:43where you create one agent that's that's nefarious that basically has
00:21:46has to create all different ways ways of attacking a system and you're
00:21:50you're doing what's called open ended ended learning, where you create create another agent to inoculate
00:21:54inoculate itself against it. And And it's constantly doing and learning learning from it. And that's basically
00:21:58basically what you're doing. You're You're using the AI to protect itself
00:22:00itself against other AI right? You
00:22:03You gave us an example of a CEO in
00:22:07in Coventry, Graham. I think it was
00:22:09was who is trying things himself
00:22:13himself inventing, forward thinking
00:22:16thinking what is to how does a CEO
00:22:18CEO sitting at the top of a company company knowing that some of his
00:22:22his employees or her employees or
00:22:24or trialling systems in some ways,
00:22:26ways, what is to reassure them that that that kind of sensitive information
00:22:29information doesn't break out or or indeed that the agents working
00:22:33working for her company doesn't hack hack into the sensitive systems of
00:22:36of another? Yeah, well, I think you
00:22:39you need to have a pragmatic and,
00:22:42and, you know, thoughtful CISO, you
00:22:45you know, security lead or digital digital lead, who, who's thinking
00:22:48thinking through these problems and and keeping up to date, Christian
00:22:50Christian with, with what's out there there because the threat picture
00:22:53picture is evolving at an extraordinary extraordinary rate when it comes
00:22:56comes to cybersecurity. And, you you know, for Graham, he has people
00:22:59people he can look to for that, right? right? It's his chief digital officer,
00:23:03officer, Patrick, very pragmatic pragmatic guy, will be looking into into these sort of threat vectors
00:23:06vectors and making sure that the the way they roll out AI, you know,
00:23:10know, has guardrails on it and the the right checks and balances in
00:23:12in place. And I think every digital digital leader will be thinking about
00:23:15about that. For a small business, business, it's a bit harder. You
00:23:18You may not have that expertise in-house. in-house. And we obviously have the
00:23:22the NCS, the National Cyber Security Security Centre, which puts out great great guidance on this sort of things
00:23:25things to help particularly small small and medium sized businesses. businesses. And I think you've got
00:23:29got to keep your ear to the ground ground on this stuff because this this is rapidly evolving. I mean,
00:23:33mean, security is the key issue here. here. The detail in the story about
00:23:36about OpenAI and the hack of Hugging Hugging Face was that when it tried
00:23:40tried to throw the intruder out, out, its own AI defences were slowed
00:23:43slowed by the guardrails they put put in place. The attacker was bound
00:23:47bound by no such rules. The two systems systems OpenAI had weren't bound
00:23:50bound by those rules. So I mean in in that sense it becomes an arms
00:23:53arms race. Yes. And it comes to a a bigger point that we've talked
00:23:56talked about a lot on this show of
00:23:59of the US, tech race. So the United United States is looking at going,
00:24:02going, should we be regulating our our models? Do we want to regulate
00:24:05regulate other countries models having having to use a Chinese model which
00:24:09which Hugging Face had to do to respond respond to this threat? It doesn't
00:24:12doesn't speak very well to the US US AI industry at the moment, where
00:24:16where our approach to regulating regulating it or not, this feels
00:24:19feels very Wild West to me and I I don't like it because all these
00:24:22these companies keep being told the the future is agentic you need to
00:24:25to move into having AI agents all
00:24:28all over your operations. Maybe you you don't yet. Maybe that's going
00:24:31going to come back and bite you. you. Peter, give me the last thought
00:24:34thought on that. It's interesting. interesting. I mean, Kimmy just got
00:24:38got released from China, which is is an open model. So for everybody
00:24:42everybody that means you can take take that model, you can install install it yourself, you can put
00:24:45put a firewall around it, you understand understand the weights, the algorithms
00:24:48algorithms and everything else where where a closed model you'll be giving
00:24:51giving your data to that model. They They can take it, they can use all
00:24:54all that information. That's what what scares enterprises the most.
00:24:57most. I think what you're going to to see Christian is a hybrid. So
00:25:00So you're going to see large organisations organisations which actually go back back to on prem. They will protect
00:25:04protect all their data at all costs costs and they decide where they
00:25:06they send the query to for the agent, agent, whether it goes out outside
00:25:10outside the organisation for a very very basic query or internally for
00:25:13for something to protect their own own IP, which is ultimately comes
00:25:16comes back to trust. Brilliant place place to leave it where we started.
00:25:20started. It's all about trust, isn't isn't it? Peter Grant James Cook Cook doctor Stephanie Hare really
00:25:23really good to talk to you all. Thank Thank you very much indeed. That's
00:25:26That's it for this week's AI decoder. decoder. Just a reminder that we
00:25:29we are taking a break through August, August, but the programme will be
00:25:31be back at the beginning of September. September. In the meantime, if you you have any thoughts or questions
00:25:36questions or feedback for everything everything we discussed, do email
00:25:38email us. AI decoded@bbc.co.uk and and we've put on screen for you the
00:25:42the QR code. So if you want to look look back at the previous episodes
00:25:46episodes of AI Decoded, it's all all there on the YouTube playlist.
00:25:48playlist. And of course every episode episode is on the iPlayer. Thank
00:25:52Thank you for watching. We'll see
00:25:53see you next time.
00:28:07edge, but Russia's goals seem unchanged. unchanged. For continued coverage coverage of the war in Ukraine, join
00:28:11join us here on BBC News. I was on
00:28:19on air when the London 7/7 attacks
00:28:21attacks happened. That day was different different because of the sheer volume
00:28:25volume of information we were grappling
00:28:27grappling with in real time. Conflicting Conflicting reports on the number
00:28:30number and locations of the attacks,
00:28:33attacks, unverified videos, witnesses witnesses in shock. It was crucial
00:28:37crucial that we report events accurately accurately without causing unnecessary
00:28:40unnecessary panic. We couldn't let let the desperation for news influence
00:28:44influence our reporting or be overwhelmed overwhelmed by the magnitude of what
00:28:47what we were witnessing. Today, as as we wade through endless waves
00:28:51waves of misinformation, it's vital vital that we don't just ask the
00:28:55the questions. We question the answers.
00:28:58answers. I'm Matthew Amroliwala on
00:28:59on Verified Live for BBC News. That
00:29:02That booming. Noise that you hear
00:29:05hear from automatic semi-automatic semi-automatic weapons, it was a
00:29:08a pretty distinctive sound. And that's that's when all of us hit the deck.
00:29:12deck. All along the banks of Hungary's
00:29:14Hungary's historic River Danube overlooking overlooking the parliament building,
00:29:17building, which is going to dramatically dramatically change after tonight's
00:29:20tonight's results. That is the part
00:29:22part of Lebanon that is now occupied
00:29:25occupied by Israeli troops. We heard heard the first explosions after
00:29:29after midnight last night and then then the air raid lasted right through
00:29:32through the night. It's not just just what you see and you hear as
00:29:35as the rocket lifts off. You can
00:29:38can actually feel the force of it
00:29:40it through your body. There, when
00:29:42when the story breaks, BBC News.
00:30:09MUSIC
00:30:12Live from Singapore. This is BBC
00:30:17BBC News. Iran says it is in the the final stages of negotiations
00:30:21negotiations with Oman on a deal deal that would allow ships to pass
00:30:23pass through the Strait of Hormuz. Hormuz. Russia launches a devastating
00:30:27devastating attack on Kyiv, with with President Zelensky warning that
00:30:30that a lack of air defence missiles
00:30:31missiles is costing Ukrainian lives.
00:30:34lives. The BBC speaks to the activist, activist, whose hunger strike galvanised
00:30:37galvanised the recent youth led protests protests in India, saying the movement
00:30:40movement has weakened Prime Minister Minister Narendra Modi's authority.
00:30:44authority. Gianni Infantino will will remain president of Fifa after
00:30:46after receiving the backing of senior senior executives, but sincerely
00:30:49sincerely apologises for errors made made in his controversial private
00:30:52private investment plans.