Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

Deep Learning with PolyAI

Deep Learning with PolyAI artwork

029: What is 'AI washing' and how you can (and should) avoid it

Apr 11, 202426 minAudio

Today's chat between Nikola and Damien delves into the concept of 'AI washing,' where companies exaggerate their level of investment in AI technology to capitalize on the current hype. The episode covers the evolution of AI, the challenge of distinguishing genuine AI applications from superficial ones, the potential need for regulation to prevent misuse of AI, and the implications for consumer experiences. Follow PolyAI on LinkedIn Watch this and other episodes of the Deep Learning pod on YouTube

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show. Tap a time to play from there.

NikolaI remember conversations back in 2019 2020 Where people tell me that they've got more people in that team than probably at the time, and, they're going to build it all. They have the resources, they have the data, and that was AI washing, right? That was completely, I watched it. Cause those people have, for the most part, fired most of those teams now. Cause the effort didn't amount to much.

DamienHello. Welcome. It's Deep Learning with PolyAI. And it's a new episode. I'm here with Nikola Merchik, CEO and co founder of PolyAI. Hey Nikola.

NikolaHey Damon, how's it going?

DamienNot too bad. I am in beautiful, sunny Florida, wrapping up a nice little conference. How about yourself?

NikolaNot bad. Your voice sounds like you've been uh,

DamienOh, yeah. I think it's a part of the mandate. Yeah. And when you're on a beach side conference, you have to go out and just keep back a couple of mind times. So thank you for noticing. I hope like maybe we tap into a whole new audience with this kind of scratchy voice.

NikolaYeah, no, it's good. I mean, I, I'd rotate my computer if I didn't fear the whole recording setup would go up, but we just moved to our new office in the city of London. It is as great as you'd expect. So, I envy. Oh, okay. Well, enjoy the new digs. I'll bring back some of the beach next time I'm there. Anyway well, so today's topic wanted to get into this concept. That's flitting around in headlines. It's called AI washing and it's essentially a modern take on green washing where companies historically back in the eco craze of being responsible to our home planet here with over inflate the level of investment resources that they were putting towards reducing their carbon footprint and light.

NikolaSo basically putting up a show that says, Hey we're taking care of our planet. Now with AI being ubiquitous with modern technology everyone wants to say that they have something going on or something AI in, in their roster. And so this phenomenon of AI washing is coming to be. So based on your understanding, You either want to use more AI or less AI, right? Because there's both concern of AI will destroy us and, we'll be uploaded when they're not. And then there's the whole other piece of are you innovative enough? Are you using AI? Like when you say greenwashing, I almost think that it's people claiming not to use it when they are.

NikolaWell, I would say that's that counter AI movement. The AI washers are the ones that are like, Hey kids, we're cool. We're hip. We're also AI powered, check out our AI. And they have no idea what they're talking about. It turns out it was a tool and like building some new solution, but really not a part of their functional product itself now, obviously. We have poly AI, you guys back in the Cambridge days, built your own language models. We built a product around this thing that ended up being the conversational AI. And we have a lot of deployments using generative as well. AI is a, it's not just our name.

NikolaIt's part of our core fiber, right? But a lot of companies out there aren't doing that. Maybe let's just start with AI is a concept right now, a lot of consumers, a lot of people who are laid to the idea, they're probably thinking of generative, right? Because their first introduction to AI in this modern concept is like chat GPT. So what is artificial intelligence? What is AI as a general construct? And how long has it really been around? Yeah, this is I'm sure we've done this in a podcast before, right? But it's irresponsible if we didn't, it's been around for a while, right?

NikolaLike it's really hard to look place where it exactly starts, like you had this famous Dartmouth conference in AI, I think 1966, where a bunch of professors gathered either for two weeks from the summer to solve AI. It was great. Very ambitious. Unfortunately they weren't as productive as they had hoped. So, we're still at it, right? The early days of AI were mostly about non statistical approaches, right? It was people trying to, and like the fascination has at the start and to this date, It's been very much about the computer, well, not computer games, but like games like chess and other ones, like showing you have a system that is intelligent by showing that it can, best the human, even the best of humans, right?

NikolaLike Deep Blue versus Kasparov, then AlphaGo more recently. And that's all fine. Those systems are, well, Deep Blue. Was mostly a symbolic system, right? So it wasn't learning from patterns and trying to generalize knowledge from like reinforcement learning. It was really much more about people writing a bunch of good algorithms for, like they're basically creating a bot that will play you and they're encoding these rules that should hopefully give it the right next action in in a complex space. It helped not to get too mathematical, but chess go and similar games were good because They were these discrete systems that were fully visible and you could define what the next action is.

NikolaThe action space is an important concept, which is what do you do next, right? Do you have one of X moves? Cause if you do then, over time you can learn which one is best and hopefully recognize familiar and unfamiliar situations. And then for a while people really focused on those things and after some time it kind of like. It didn't really garner all that excitement because how many decades can you work on a better chess player before people Stop believing that you're about to, create an AI that will be like omnipresent and used for absolutely everything and whatnot You had a few starts in what is considered more narrowly machine learning, which is really using statistics To, uh, work on a problem to do a task of a particular kind, right?

NikolaIt could be image recognition. It could be facial recognition. It could be speech recognition, like all sorts of things that are maybe assistive things for humans all the way to like full end to end task completion of, I don't know, a bot that does something that you would have otherwise had to do on your computer for five minutes. And like machine learning really took off a few times and neural network approaches. Had a few kind of different phases in the nineties. People showed that, the multi layer perceptron could do some stuff, but it really wasn't even that competitive compared to other statistical approaches, like I don't know, a random forest or other classification approaches.

NikolaAnd then really 2012 onwards we entered the era of deep learning, and it hasn't really stopped since where through various different optimization methods, we got to the point that. These systems became state of the art for a vast number of different tasks. And performance improved and we got more and more confidence that the whole thing has a future and can do some things better than a human. Right. And that all kept progressing and we were all excited, right? That's the era where I started my PhD. And that was like, two years a year and a half after some of the first seminal papers on deep learning came about mostly Jeff Hinton's group.

DamienOkay.

NikolaBut then, you basically the final flip is LLM showing that Model that is fairly generic can actually do far more than we ever initially expected it would do. it was built to predict the next word, right? And to tell you which sentence is more likely than another sentence by looking at the perplexity, how likely are these words to go together, but taken to an extreme and consuming all the data. And then there's a bunch of other tricks and training and creating synthetic data sets where they see. Examples of how to ask us to perform how questions can be answered and stuff like that.

NikolaThat's where then you get to an interesting point where They start generalizing and generalizing in ways that you did not necessarily, you hope that they would pick it up from data, but you didn't explicitly give them a mechanism. They inferred that themselves. And that's the point where I think usually it's

Damienis that the progression? Cause like you said, going back to the early foundational applications of AI, it was around games because it was a very finite data set. It's sitting in an narrow space like with go and chess, but now with Natural increase in computing power and access to information. That's that was the expected evolution to what we see today.

NikolaI think what's interesting is that we all expected that these things would get better. But these like emerging properties, like the ability to generalize that well, I think people hope for it. And then for a while it wasn't really happening. Then, you had a few people like Helene Sutzkever who believed that they would do far more far more like that. There would be a moment where the increase in capacity would not just be like some thing that follows mathematically the amount of data, but it really gets to a level of capability of something that you did not intend. And that's right.

NikolaAnd that's why we're so obsessed with them right now, because it really feels like we. Something unexpected happened, right? Like it's like the, GPT 3. 5 is really, a bit of a, like an Oppenheimer moment, right? Where like Trinity test, poof, okay. The world is now different, right? We have a singular system that by consuming all this data has become capable of being this almost utility function for development. They can do so much more right now. Creativity is exploited and people claim that, well, they're trying to do all sorts of fascinating things.

NikolaSo I think that's where it the whole stuff around AI washing is interesting because I think we've compared it to at least if you look at the life of a CIO, Like this moment is similar, except for now, even the CEO is far more interested. To the moment where cloud computing came to be and, those first became like the dominant companies because they had the versatility and ability to do far more and to use other services and applications. And they could pivot much faster than someone who had servers that, and people deploying software off of like CDs onto those servers.

NikolaThere was a moment where people bet on it. And then, the conservative ones said, Hey, safety is going to be an issue. All like safety again, different kind of safety. This is more like cyber security. But those that like made the bets, profited generously and others are still talking about how they're navigating their. Transition to the cloud. So I think like having seen that,

DamienCloudwashing had its moment, then it seems.

NikolaI don't think so. I don't think so. I think there, it wasn't seen as immediately as beneficial. Now it is. And look, I think with AI, because people don't want to miss another moment in the evolution of technology like this. Everyone almost wants to like, pump themselves up and say, Hey, I'm actually doing a lot with AI. What are you doing with AI? And that's really where now we're just like in the, the peak of the current hype cycle where like people can do no wrong when they try AI initiatives, what they're actually doing and, it may be just that they don't want to look stupid in front of their friends.

DamienBut that's the thing is that, first of all, this is a longstanding concept. It's going into a massive evolution with generatives capabilities. All the big tech companies are creating their own LLMs. It's something where you and I is well, not a lot more about this stuff, but me as a consumer, I'm like, Wow, this must be important technology if it's literally what every single big tech company is putting out and touting as far as their entrance their player in the big overarching consumer tech race. So again, this goes back to when people conceptualize AI, their thought is, Oh it's that chat GPT thing.

DamienSo they're talking about generative AI and when companies say, Hey, we're AI powered, a lot of times it is. That open source, LLM that is just doing something and they put their wrapper on it and then, like you said, it was, they gave it a UX, right? And then tried to brand it as a product, but it isn't, it's really it's a data set. It's a tool. So I understand the nature of why companies would want to AI wash, right? It's like tapping into the zeitgeist. Everyone's talking about it anyway. So let's, Try and find a way to put it into the fold. Where is the line between actually utilizing artificial intelligence in product solutions in delivering services versus it's just a functional tool, but it's not actually powering the end result.

NikolaIt depends. I think that like people would try to build all sorts of things, like whether, you've got like a friend of us to say that, C3 AI probably has. A much larger market kept us because, they got thicker and people want to buy AI, right? So you're going to buy, even though they're not really much of a company. But really I don't think it matters frankly, cause the products that are powered with the right kind of Architecture and the right kind of model. I think now we're talking just about like very tactical bets, right?

NikolaIt's as old as any other software iteration, right? Where some people want to build it all themselves. Some people will my package products from others. Some people would buy configurable platforms and some people will want to build their own LLM and then build on top. I think it's really just what are your actual problems and like, where's the right place to start? And, people have done at least in our space, People have done absolutely everything already. I have clients and prospects that have done everything from, get a voice assistant from us all the way to try to build their own and work with us without a template of systems all the way to people have Such well funded teams that they're training their own models.

NikolaThey're gonna do it all themselves and by and large, you know I've been at it for a while now. I remember conversations back in 2019 2020 Where people tell me that they've got more people in that team than probably at the time, and, they're going to build it all. They have the resources, they have the data, and that was AI washing, right? That was completely, I watched it. Cause those people have, for the most part, fired most of those teams now. Cause the effort didn't amount to much. Now, of course the technology is more powerful unless you can do more with less, uh, the interesting bit with I don't know if this falls under your AI washing bit, but it's we'll see a demo, they get excited.

NikolaThey give a presentation to the wider company about like how they're using generative AI. And long before, you're off to the races to build something, but you don't know how to productize it fully, how to package it up, how to feed analytics from it, how to create the right safety conditions. And then, you might be in trouble, right? So we'll see what the right approach is for the different, application areas where people are using generative AI and like there will be winners and there'll be losers. I don't think that the collective wave will disappoint. I think there might be people disappointed when, as a band, they start training their own LLM and Two years later, they've done what, five foundational startups have the resources to do and, the chips, the data, the whole infrastructure around it, right?

NikolaSo like those guys are probably disappointed, right? They may or may not admit that to themselves because they probably won't, like a bank building its own LLM will not build something comparable to open AI or Anthropic or Mistral, others, right? But, I think that those using LLMs to power products will, achieve efficiencies. Right now, I think it really depends on who you're buying it from, what you're buying, are you buying it because I told you to buy AI in which case you, you may not do that well, or do you have a business problem that you're solving from?

NikolaAnd then you look at okay, what am I actually doing? So, I spoke to, to, to a prospect recently and they're doing everything from chart automation to. Assistive tools with generative and, we've all seen co pilot demos, right? So that's very interesting. If you need to make a slide and you go look at this presentation now, create a summarization slide on a way, turn that into a table. That's fantastic. That's like making so many investment banking analysts redundant. Because a lot of sleeveless nights will be solved with one prompt into GPT, right?

NikolaBut if we're talking about like a live problem, say you're a contact center agent near your job, and you've been asked a question you don't know the answer to. Like either the answer shows up in front of you straight away and you start reading it or more likely, right? Like it takes a bit of time for it to show up. You're mumbling in the meantime You're reading the right to understand you're panicking. You started saying one thing. Did you see another? Listening to you, I mean I've seen this right like the system is listening to you and yeah It starts listening to what you said that feeds the problem differently and then the whole thing changes in front of you And I've been fascinated with the problem because you know I'd love to have PolyEye with a complete product offering that offers everything from full automation to assistive tools, both chat and voice.

NikolaWe do both chat and voice, but we only do end to end automation. And the debate around should we be doing assistive tools is as old as the company itself, right? Yeah. And what we found to be very interesting is, While the same model, the same LLM in the background can generalize and support multiple use cases in end to end fulfillment scenarios, that's not the case for an assistive tool, right? If you're quoting me different, like insurance things and you're like, Hey, with this success payment, you'll pay me 17 a month.

NikolaAnd with a much larger access, you'll pay me 12 a month. Like that needs to pop up and show all these differences and things. It needs to be. Like very clear to me very quickly if I'm going to communicate that to you efficiently, right? And to build all that you can't really go and um, it's not that easy to create that interface that would then be reusable for all these different things. So the U S

DamienWith that real time application without those, without those moments of latency, where you have to sit there and mumble until the systems can catch up to the pace of conversation I really like this concept that you mentioned about the early days of AI washing organizations would say, Hey, we've staffed up. So it's based on, it's based on the human resource element, right? It's, Oh, we've got the team that can definitely build this. And that's the early premise. We don't have any of the technology, not a whisper, but we've got the team that does so more to come. I think that's fascinating and an interesting take on. It's really cool.

DamienWhat's going

Nikolaon? The technology is new. No one's ever built anything functional with it. And then you're sitting there believing that you have it all, right? At least have a healthy dose of skepticism. It's great that you're that confident, but come on, be realistic, right? It's not that easy. You just can't be that sure.

DamienIt's not proven yet. Okay. So be confident, be clear, be transparent. Don't try and put the cart before the horse. These are ways that you.

NikolaIn these environments, it's difficult, right? You have to get the buy in from people. And then you have to be a bit extra confident and then, either you're a hero or you're fired, but that's cool. That's what I always really appreciate, especially, when you look at like an American versus European context. Always, and not just us, I think every like enterprise software company has found that, American companies empower stakeholders to take some risks, make some decisions. Hell, you might not even end up fired if you're transparent and you're cap the, the investment at the right time. In fact, the best people in a lot of mistakes and the few really impactful.

NikolaGood decisions. And they're like, end up in C level roles. Because, they're like visionary leaders rather than people who hedge, I think You should be experimenting liberally. You should be experimenting internally. You should be experimenting with external vendors. And you know what? Funding is sloshing around everywhere. There are a lot of companies doing a lot of work at every level of the implementation you might want.

DamienWell, that's a key. That's a key contributor to this concept of AI washing and gaining predominance is from a funding level. Like it's just, that's where dollars want to go. And so there's a little bit of that trying to capitalize on that market. Momentum and say, Oh, all we have to do is like you were saying earlier about buying C3 AI, it's Oh, you just have the you're, you have the right ticker. And so you just happen to step into it. So now people are basically slapping an AI sticker on whatever they're pitching for to try and secure that funding.

NikolaOh, absolutely. I mean, Look, I talk to investors all the time and it's very interesting. I would actually tell you that it's more interesting to speak to investors who have been AI washed. Cause you'll speak to someone who's never in an interesting company before. And look, when we started, I was a bit more humble than I am right now. But right now, when I go into a conversation, then, someone's like, how are you going to compete with open AI? I'm like, you haven't looked at the space much, have you? It's like me going and, talking about American football, even though all I know is that the coach and his tactics matter a lot more than they do for real football.

NikolaBut it's I'm sorry for offending you. But I'm right. It's called cold soccer. Okay. Yeah. Those that have been investing in it for a while, like they are pretty savvy about these things. Although I think for them as well, this day and age is difficult to make the right bets around differentiation.

DamienYeah. I think that's fair. What about the other side of it? The future of it? There's a very thin line between what could be considered what we're talking about. This like AI washing light, which is just, I don't really know what I'm saying. I don't really know what you expect, but I want to make sure that I'm a part of the ad conversation and going into things like puffery or bait and switch, which have real world applications and real world finds. If you, if, yeah. There's a potential for looking into legislating. I know there's already some active legislation as of mid 2000 23. Was it last year where something was passed that basically prevented conflicts of interest for finance companies and AI solutions within their infrastructure?

DamienSo it's, we're starting to trickle into responsible use of and responsible transparency of utilization of things like AI tools. So what do you think is the potential future for trying to regulate this to make sure that it doesn't get to an extent of violating FTC or SEC regulatory requirements?

NikolaI don't know. That's that's a whole different question. I'm not sure about that one. I think that in terms of using AI and regulating it it's the wild west right now. I think the EU is always at the frontier of regulation, if not innovation. So, The EU AI act is probably the most comprehensive piece of legislation with opinions and all sorts of things. I think it's really to regulate it effectively, have to know where the buck will stop and that's really difficult right now. I think for instance, recently you saw the FTC. Talk about like banning robocalling for like a bunch of new use cases and stuff, because it is getting difficult for people to know whether something's automated or not.

NikolaIt's a debate that we in our products and deployment teams have had for a long time around like whether you should be saying you're automated or not, or at least in the first instance. And if you do how do you say it? It has real repercussions on how people speak to the system, because even if they know that it's AI or not, like then you, when you build the first layer of the onion, the second one is, well, okay. If I think it's AI, do I think it's going to work or do I just start screaming for the agent? So, the first level is ethics slash maybe a bit of grandstanding where you're like, you should be honest, transparent, et cetera.

NikolaIt's okay, but you should also make it a seamless experience. If you're going to use it, you need to make it functional. So you know how for example, when you log into a zoom call, that like recording bits, right? But then after some time there, they started adding that this code is being recorded, right? And every time someone new logs in, you hear this code is being recorded, right? To make sure that they haven't missed it. Honestly, that doesn't improve user experience, but it ensures that people see that the code is being recorded, that dot like flashing and telling is being recorded. You might miss it, but it's there. So I think as we grow with these systems my hope is that for automated voice.

NikolaOkay. We just have a different dial tone. And then we're just able to basically no, it's automated. Like we teach the general population. This is not an automated experience, but we don't make a fuss with it. We don't need a 10 second warning because when you hear a 10 second warning, your experience is worse. Cause you wasted 10 seconds.

DamienYeah, no, it's true. And it makes a lot of sense when you're saying, if I'm hearing you correctly, that just. When it comes to regulating against nefarious use of this sort of concept of AI washing, right? That the focus would be on customer consumer facing applications of the technology or the systems, like something's going on in the backend. You're using that as part of marketing material. It's teetering on puffery. That's not really the highest priority because it doesn't involve the consumer experience, which that touch point between a business or brand and their users is probably

Nikolaanother AI system. This is like pre generative, right? But you've had computer vision systems that ended up like doing racial profiling because of the data they were trained. This is not a new debate. It's an important debate that, like it really depends on the use case. And then, the more we make these systems data driven and complicated, the harder it will be to actually regulate their behavior. So you can expect basically have like stringent testing and definitions for like an AI system for a specific task, right? They need to define stringent benchmarks of how it works.

NikolaAnd I think that's the only way that you get to something like fair in the end. And it's also really hard to regulate whether AI is used or not, because that's really hard. We try in the first 10 minutes, like there isn't really a strict definition. It's all just technology.

DamienYeah. You'd have to, you'd have to basically get your doctorate in order to understand, interpret, and create that sort of legislation. How many people would

Nikolaactually enforce it? I don't know, really. Cause like I, I have a PhD in AI, and honestly. Like 80 percent of it became fully obsolete the moment that, GVC 3. 5 was out. What do I have to say? That's well, thanks for your

Damientuition.

NikolaI had a scholarship, thank God. But yeah, but hopefully we produce some value on top of it for society, but it's stuff. I think it's the age of an expert in this space. Like the most creative people working with this stuff now are not the academics. In fact, people are increasingly not doing PhDs because all the people you could have done PhDs with are paid, multi million dollar salaries in open AI or Google or whatever. So

Damienyeah, they're sitting in the private sector at this stage.

NikolaAcademic like distinction and titles and everything is worth diddly squat. So.

DamienCool. Cool. Projection for the future. All right. That's a great point to add on too. Nicola, appreciate your time. Appreciate like being able to tackle the next great like phrase that's in the headlines about AI washing and how to decode it and how to get past it. Can't wait to talk to you next week. Awesome. Have fun in Orlando. Thanks. Take care.

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. What if your contact center became a product team?

    Nov 17, 202635 minAudio

    Three years ago, Simplyhealth's PolyAI deployment was only resolving simple FAQs over the phone. Today, they run voice, in-app, email, chat, and WhatsApp conversations on the platform. Nikola Mrkšić talks with Paul...

    Credit the people on this episode

  2. What happens if your AI agents get 1% better every day?

    Sep 10, 202639 minAudio

    An agent that gets better every day compounds into a very different agent within months. That's the idea behind Wren, PolyAI's new release: an agent whose job is to make your other agents better. Some of what it finds...

    Credit the people on this episode

  3. Who's coordinating your army of AI agents?

    Sep 3, 202644 minAudio

    Companies are pouring money into AI, but not everyone is getting the returns. Ian Jacobs, VP and Lead Analyst at Opus Research, thinks he knows why. In this episode, your host Nikola Mrkšić sits down with Ian to work...

    Credit the people on this episode

  4. Why should CX leaders care about MCP?

    Aug 20, 202636 minAudio

    When PolyAI built its Builder MCP, one client helped shape it: Vixxo Facility Solutions, one of the largest facilities providers in the US. In this episode, guest host Damian Sasso talks with Vixxo's CTO Derek Neighbors...

    Credit the people on this episode

  5. Can AI really hear a call the way a person does?

    Aug 6, 202622 minAudio

    PolyAI just launched Dialog-RSN-1, its first audio-native model, and your host Nikola Mrkšić sat down with its builder, Matt Henderson, to unpack why it’s a game-changer for building voice agents. Most voice AI either...

    Credit the people on this episode