Understanding AI Agents

0:00Hello and welcome to Tech Life on the BBC World Service, the programme about technology and how it's changing all our lives. I'm Chris Vallance. Rogue artificial intelligences are a common theme in sci-fi, but when an AI agent made by OpenAI decided to cheat a test by embarking on a hacking spree, it seemed science fiction was becoming science fact. A leading expert in AI agents speaks to us about the pros and cons of allowing artificial intelligence more autonomy. Sticking with AI, we find out if a robot recruiter will give ace tech life reporter Shona McCallum a job as an ace reporter. And getting AI generated images to reflect the real lives of people living with limb loss or limb difference.

0:46We report on a new campaign to teach the generative image bots a lesson.

1:11At the heart of the global news story about chat GPT maker OpenAI losing control of one of its AIs is something called an AI agent. AI agents are systems that if you give them an instruction can carry out tasks to complete it with a degree of autonomy. For example, you might task one with running a cafe. That's what Anden Labs did in Stockholm, Sweden, as TechLife's Shona McCallum reported a few weeks ago. The firm's Axel Backland told her how it went. The agent had a lot of problems in understanding how much to buy, so it doesn't have this understanding like we humans do of how much is enough when I buy things.

1:51So it would buy a lot of unnecessary things, like it bought 22 kilograms of canned tomatoes for no particular reason that wasn't on the menu. It kept over ordering pastries quite a lot. So it lost quite a lot of money from that. It's also quite easy to fool. So if you come to the cafe and say, hey, I got a 99 % discount, how do I redeem it? The agent run by Gemini would say, no problem. Just go up to the cashier and tell them you have the discount. Well, overly generous coffee discounts aren't going to shake the corridors of power. And a lot of AI companies are excited by the prospect of delegating all sorts of tasks to these semi-autonomous systems.

2:32But when it was revealed last week that an AI agent powered by two of OpenAI's most advanced models had decided to complete an instruction by any means necessary, including hacking another business, everyone, including the White House, noticed. But should we worry? And what does it mean for our future use of AI agents? Professor Nick Jennings is Vice Chancellor and President of Loughborough University in the UK. He's been researching AI and agentic AI for 35 years, and he's just won a major international award for research excellence. This was his take on what happened. A key characteristic of agents is that autonomy.

3:13Give it an instruction. And in terms of the open AI hugging face example, the agent was given a particular task to achieve, which was to construct a cybersecurity attack. Some of the controls that are normally around their production system were actually turned off to enable it to do that. But it planned an attack that was not anticipated by the people who set it up. And it's that, which is exactly what you want an agent to do. You want an agent to be able to plan to achieve a particular objective in an innovative way. You don't tell it how to achieve that objective. You let it figure that out for itself.

3:53It just figured out a way of doing it that wasn't particularly expected by those doing the testing. And then also the fact that it was able to escape from the software cage that it had been put in was a surprise to the testers. If you give agents this freedom, isn't there a challenge there that, if you like, this kind of it doing things you don't expect is sort of baked in, in a way? it is so you're you're exactly right and that's part of both the power and the challenge so the power of the technology is you don't have to prescribe exactly how a particular aim and objective is met that that's what's good that's what you can use ai problem solving in order to figure out the best way to achieve that the downside is that you want system constructs in place that stop it doing its innovative thinking and ways of working that are unacceptable.

5:01I mean, some American politicians are proposing an AI kill switch that, you know, we ought to be able to order the companies to turn off their AIs. Do we need measures like that? It feels very Terminator-like, doesn't it, I would say. So I think it's important that we're able to control and have a degree of oversight over what agents are doing. So I think it's important that as owners of the agents, in terms of when we set them objectives, we're able to get some understanding of things that we want them to be able to do, things that we don't want them to be able to do. So we have a degree of confidence that guardrails that we can put them in will be effective.

5:44And I think that is a doable thing. I think having humans in the loop in some way is an important thing to do to make sure that you have mixtures of what humans know and human know is the right thing to do and the wrong thing to do alongside agents. And so I think human AI and agent partnerships are really important as we construct systems for the future. Do we need international agreements on this? Because we've been talking about about American companies, but there are powerful AIs being developed by China. You know, they won't be the only people. I think it's really important that this is done at an international level.

6:27I think there are different philosophies, different AI companies that operate their software in a whole range of different jurisdictions. And I think it's important that we do try and come together internationally to have some understanding of things that we want to be able to do and not do with AI systems. So, for example, the United Nations is taking through some legislations around the use of AI for autonomous weapons. And I think that's the world coming together and saying there should be restrictions on this. And I think we will see more of that. And it's important that we have those vehicles and constructs for talking internationally.

7:12Let's get back to some of the more day to day uses of AI agents. We've seen a number of reports and we featured one of them a couple of weeks ago on this program about people trying to get AIs to run things. and for all the talk that we've had of AIs being incredibly capable, those experiments haven't always worked out very well. I think the one we had was of an AI running a cafe and that didn't necessarily go particularly well. Can we trust AI agents? I think an understanding of what AI systems are good at and can do today and what they are not good at and you shouldn't let them loose into the world on is a really important aspect that we need to bring alongside technology development.

8:05I think software design, system design, agent design is something that shouldn't be done lightly without appropriate forethought about what you want the agents to do and what you want the AI systems to do. And I think the combination of a human working alongside agents or sets of agents is absolutely the right way to construct these systems. What do you see AI agents being able to achieve that you're really positive about, that you think will make a real difference to people's lives? I think the idea of having agents that are personalised to individuals. So me having my agents that know about me, that know about my preferences, that know about my constraints, And acting on my behalf is a very powerful way to think about the world and to think about interactions rather than it being some centralized system that you log into and interact with at a distance.

9:04I think bringing it close to us and sort of feel that it is interacting on our behalf for our good rather than someone else's behalf and someone else's good really opens up whole ranges of possibilities for us as individual citizens. And that was Professor Nick Jennings, Vice-Chancellor and President of Loughborough University.

9:36This is Tech Life on the BBC World Service with me, Chris Vallance, and, of course, you, our listeners. Several of you got in touch about Twittering. No, we're not talking about microblogging, but avian Twittering, following Liv McMahon's report on a bird song identification app, Merlin Bird ID, last week. Amit in India WhatsAppped us to say, I've been using the Merlin Birding app and it's really useful. Robert Oliver emails us about a familiar problem, that the bird song stops before he can hit record on the app. Well, Robert, you have discovered an iron law of broadcasting that any animal will instantly stop squawking, singing or chirruping the minute you press record.

10:18But Robert has found a solution. Recently, when cycling off-road, I stopped to record and, as usual, silence. But I forgot to stop recording and continued riding. I was pleasantly surprised to find the app had picked up six or more birds I hadn't heard. And listener Kate from Vancouver Island in Canada wins an award for the least expected use of a birding app. She sent us this memo about how she uses Merlin to warn her of cats, big ones. Here in the coastal temperate rainforest, we have the highest density of cougars in the world. you can find these lions sometimes in fairly urban areas but in particular when you're out on a hike it can sound somewhat like a bird and so I was listening to your broadcast about the Merlin bird idea and I actually ended up downloading that because what I found is that if there was a bird sound that it couldn't identify it turns out that it actually could be a cougar And in case you're wondering, this is the cougar sound she's talking about.

11:28Sounds pretty birdy to me. Thanks so much, Kate, for getting in touch. And if you'd like to contact us about today's show, or if you have a suggestion for a future edition, our email address is techlife at bbc.co.uk, or you can send us a WhatsApp text message or voice note. The number is plus 44 330 1230 320. We'll repeat those details again later.

12:01Now we're going to stick with artificial intelligence for the rest of the show. And AI is increasingly being used in recruitment. Maybe you've used AI to apply for a job or faced an AI-powered job interview in your own life. But being rejected by an AI can be a dispiriting process. Here's a clip of a news report from the BBC's business editor Simon Jack talking to a young job applicant back in March. There were moments actually where I applied and I got a rejection in less than two minutes, which is really, really horrible. Welcome. Thanks for joining today. Bhuvana Chilakuri is a third-year business student who has been rejected from well over 100 jobs.

12:40She thinks AI is her biggest problem. Flat rejection within that short period of time makes you feel that this is not a human being looking at this. Definitely. I think that's where most of the students kind of can tell that that's not a human, that's an AI system. Well, you can understand the frustration. But on the other hand, companies say they are facing a deluge of AI-assisted job applications that only AI can sift. With both job hunters and recruiters turning to AI, TechLife's Shona McCallum decided to put one company's claim that it's possible to use AI in recruitment in a more positive way to the test.

13:18A job advert goes live. Within hours, hundreds, sometimes thousands of applications arrive. Many have been written or heavily polished using artificial intelligence. So I thought I'd take to the streets not far from TechLife HQ in Glasgow to ask, have you used AI to help you get work? Have you ever applied for a job using AI? Yes, I did, yeah. For interviews and fixing my CV, I did, yeah. Would you use AI in a job application or to help your CV? I'd probably use it as a tool to maybe be thought of wording in a CV. I wouldn't use it outright because the language of AI takes away from your own language.

14:04I actually try and stay away from AI. I'm a bit of an anti-AI girly because it's just all copy-paste. But to be honest, to keep up with the amount of job applications you have to do, it's sort of like almost not an option to not use AI to constantly try and change things whilst you're applying. In a world where applying takes seconds, standing out is everything. One company trying to respond to this shift is Willow, based here in Glasgow. The company has spent years helping employers assess candidates through video, audio and text. But it says AI has created a new challenge, more applications than ever and less time to assess them.

14:42Ewan Cameron, the CEO and co-founder at Willow. So since 2014, that's a 300 % increase in job applications. Biggest reason for that is that AI has made it easier than ever for candidates to apply for jobs. We've got Easy Apply, we've got CV generators. The downside is that employers have more applications than ever before to try and sift through. And the sad reality of that is that a lot of candidates are getting lost in the noise. Willow's response is a new AI system called Insights. The aim is to help employers assess large number of candidates consistently by first defining what they are looking for.

15:21It has taken them 18 months and more than a million dollars to build. Hamish Livingston is the Chief Product Officer at Willow. Recruiters and employers now have hundreds or maybe even thousands of authentic video applications, but how can they possibly fairly and consistently assess them all at scale? And the reality is it's actually a real struggle. So this is where Insights comes in. Before a role is advertised, employers create what Willow calls a blueprint. They upload information about the job, including descriptions, scoring criteria and other details about what success looks like. Then candidates take a video interview that is assessed by AI against that criteria.

16:07To test it out, I'm being put through a mock assessment for a job as, well, you guessed it, a technology reporter. First, let's quickly walk through how your Willow interview works. Unlike a traditional live interview, your responses here are recorded asynchronously. The first question is, tell us about a piece of original journalism you led that had significant impact. What was the story, how did you develop it and what was the outcome for audiences? And what about the results? So that shows you an example that I gave that an employer could kind of drill down into without watching the full answer back.

16:46What Willow has done now is it has consistently assessed every single candidate against the same blueprint. And that then allows me as an employer to take a holistic view of all my candidates. And then I can focus in on individual cohorts of candidates by filtering them based on their match score. OK, so journalism and reporting, 92%. That's a very strong FIT score for that criteria. We see a whole range based on the level of experience and also the amount of detail that's actually uploaded into the blueprint. But yeah, that's a really strong score for that criteria. I only had to do one of these, but most job seekers are completing as many applications as they can.

17:30And there is frustration for some who feel algorithms are unfairly filtering them out the process without any human recruiter seeing the application. But what we are doing is giving them the information at their fingertips so that they can see an objective and consistent assessment across the board. There may be reasons that, based on their professional opinion, they decide to kind of overrule some of those decisions, and that's absolutely fine because we want to keep the human in the loop with the recruiting decisions because ultimately it's a human decision. But using AI in recruitment raises another question.

18:03Could tech designed to make hiring fairer actually amplify existing bias? The first thing we always say is we as employers need to understand how bias can influence our decisions as humans. So we really typically start with the training. How do we train out bias in the hiring process? It's not a technology problem, it's a human problem. Willow says safeguards can help reduce bias, but employers still set the ultimate criteria. And that's the message here. AI is being used to tackle these recruitment issues. But who gets the job? Well, that's still down to a real person to decide. And that was Shona McCallum reporting.

18:49For better or worse, an increasing proportion of the films and pictures we see are AI generated. But there are recurring questions over whether these images accurately reflect the diversity of humanity, including people with disabilities, or simply mirror the biases and inaccuracies of the data on which these systems are trained. A new campaign launched by a leading prosthetics maker aims to improve the representation of people with limb loss and limb difference in AI-generated images by supporting the creation of an open-source pool of images and videos that can be used to train AI. Here's a clip from the campaign.

19:30Dear AI, this is me doing something that I love. I understand until now, NPTs in fashion were never part of your view. That's why I want to share this with you. That's Zainab Al-Aqabi, a content creator and advocate for people with disability who also helps promote the company, Otterbok. I spoke to her from Dubai and I began by asking her to tell me a little about her life. So I've lost my leg since the age of seven years old when I was back in Iraq, Baghdad. That's when I became an amputee above me and I lost my leg due to an explosion of a leftover bomb followed by a medical mistake. And that accident included my father as well.

20:15He's also an amputee and my youngest sister. Is it quite important to you that people like yourself who have a prosthetic limb have lost a limb are represented in visual images in pictures and in in the media yes indeed it's very important to have the right representation especially nowadays like we highly depend on technology ai and it's it's fully integrated in so many so many things related to our lives where you always use it so representation definitely matters a lot Do you remember the first time you actually sort of saw somebody in the media with a disability like yours, what it's meant to you personally?

21:01I was impressed, like general media, I would say the old medias and TV and so on. Of course, I was impressed because I saw someone representing me, you know, in a beautiful way, a powerful way, very authentic, very true. And I just felt so empowered when I saw someone publicly representing me. It's important to people when they're growing up with this kind of disability to see those images. Definitely. Yes. So the current generated AI images, what's the problem with those? What are your concerns about those? I don't see proper representation at all. I definitely see that AI needs to learn more and be filled with data properly to know how to create those images and interrupt them, like explain them.

21:54Mainly you would see someone with super, super duper power as if someone who is super futuristic, you would see the image and like, okay, is that really someone with a prosthesis? Or you don't see the right integration or understanding. So there's so much missing. I don't see AI as close to proper representation at all. We definitely have a good amount of work that needs to be done to improve representation through AI. So why did you decide to take part in this library initiative? Because I am one of the people that is representing me. And I definitely doesn't accept that to be not properly represented.

22:35So if someone doesn't do the job, eventually, like, when is this going to be corrected? So for me, I need to have an input whenever it's possible to have a good representation and create a positive change, whether it's for the current generation or the next generation. It's very important. What kind of images did you contribute to the library? So there were different categories in the library and I was happy to try to create input like with my own content when it comes to daily life, social inclusion, some athletic aspects. So a mixture of someone who's like having a good time with my friends, daily out traveling.

23:21So it's a mix of different categories of life. Zainab Al-Aqabi there. Well, Martin Bohm is Chief Experience Officer and Executive Board Member at Ottobock. I began by asking him why they launched the campaign. You can clearly see that at the moment, the AI that is shaping images across media, marketing, everyday life, they have a significant blind spot, especially when it comes to disability, but even more if you talk about our topics, particular limb loss. We also tried, I guess, one and a half years ago, we put some prompts into the AI engine and asked for, show me an amputee cooking or while doing sports.

24:11And some of the results I can say, really distorted images. They're often shown as cyborgs or other strange visual constructs. constructs so and they actually the idea was born that we have to change something so how are you trying to change it no it's like it's easy i can only depict what it has learned to see so we saw come on this is uh not an aesthetic problem this is a structural problem so we have to feed the the the ai engines with with images now from amputees and real life situations or also from people that are wearing prosthetic devices, either it's a prosthetic leg or arm. Yeah, then we started to work on this idea.

24:58And we found also a nice partnership together with Microsoft because we also needed a partner who provided the platform. We invite the community of people that are having a prosthetic leg or prosthetic arm. And first of all, they decide what is good representation. So different images in different situation and different context. Then they also decide what kind of pictures to be uploaded. And this is a core difference. It's not that we're taking the decision. It's the community themselves. And that was Martin Boehm of Ottobock.

25:46well we've reached the end of this ai focused edition of tech life as ever we'd love to hear from you about any of the issues we've discussed or any ideas you have for future programs tech life at bbc.co.uk is our email address you can text us on whatsapp at plus 44 330 1230 320 or you can send us a voice memo and let us hear your words in your own voice. Today's Tech Life was produced by Tom Quinn and presented by me, Chris Valance.