087 – Chatbots and AI

[00:00:00] Lily: Hello, and welcome to the IDEMS podcast. I’m Lily Clements, a Data Scientist, and I’m here with David Stern, a founding director of IDEMS. Hi, David.

[00:00:14] David: Hi, Lily. Looking forward to this discussion.

[00:00:17] Lily: Yes, I thought today we could discuss about chatbots and AI.

[00:00:20] David: Yes. And this is particularly relevant for us because we’ve got very sucked in to building chatbots. But most of our work is building old fashioned chatbots, whereas everybody now is excited about AI based chatbots.

[00:00:35] Lily: So could you first just distinguish that difference for me between old fashioned chatbots and…

[00:00:41] David: Well, for a long time now, computer systems have been set up so that you can choose from a set of answers, and you have this with your phone, whether it’s an audio sort of chatbot type thing, where it says choose one to do this, choose two to do that.

[00:00:58] Lily: Sure. So those flow charts you see when you’re younger and it’s, or maybe even these days, to make a decision.

[00:01:06] David: Yeah. And it’s taking you through a set of deterministic, and I think I’m using that word a bit carefully, information that in a way which is consistent, reliable, and from a user’s perspective, sometimes really frustrating.

[00:01:22] Lily: Yes, I know, in fact a lot of the time when I start with chatbots, if I’m using a kind of online ones I will just straight up say put me through to the human.

[00:01:33] David: Exactly. And this is where when they are designed in certain ways, if you are looking for a particular type of information, if it doesn’t have the exact thing you’re looking for, it can just take you around in circles in ways which are really frustrating. Online banking is where I find I continually get very frustrated with the way some of these chatbots are set up.

[00:01:59] Lily: That is the one I had in particular in my head, yes.

[00:02:03] David: But it’s used in all sorts of service industries, airlines use them in all sorts of ways, it saves, in theory, it is better than a call centre because in a call centre you need humans to have those structured systems anyway. And so there should only be a set of things you can do anyway. Done right, it can be more efficient and more effective.

[00:02:26] Lily: And don’t get me wrong, I’ve definitely come across some which are done excellently. Where you go through and you get your answer straight away and it’s perfect.

[00:02:34] David: Exactly, and I think one of the key things is what are you using this for? And one of those key cases is that as with the banking or the airline, you have a specific need at a given time and this is set up to help you meet that need. And the way in which a sort of deterministic system can do that compared to a human, if it’s well designed, at times it can be very effective. But it does depend on the nature of the need, and there are just ways in which humans are better listeners. And this is where the AI chatbots come in.

[00:03:18] Lily: Okay, and I just want to clarify, you’re saying deterministic systems, are these your kind of traditional systems?

[00:03:23] David: Yes, so a traditional deterministic system, you design it, so it will always do the same. Whoever’s going through, if you do it twice, you’ll get the same result every time for doing the same set of things.

[00:03:34] Lily: Okay.

[00:03:35] David: It’s always the same.

[00:03:37] Lily: And so now we’re having kind of AI chatbots. I guess one of my questions is, how’s this different from, say, something like generated AI like ChatGPT? Is that not some kind of chatbot?

[00:03:49] David: I would argue that is. ChatGPT, generative AI in that context is a AI chatbot. You chat with it. You can have a discussion. Now that discussion you can design your generative AI to be more specific on a particular topic. You can train it on specific information and data so that it learns how to get that data out, and to give that data back out to respond to more specific questions.

This is, I think that you’re using the word generative here. This is what a lot of people think of when they think of now AI chatbots. They think of generative AI chatbots.

[00:04:33] Lily: Okay.

[00:04:33] David: Which are linguistically able to interact with you and maybe even pass the Turing test. We’ve discussed the Turing test before on previous episodes, where an external observer may not be able to tell the difference between you chatting with a human, and you chatting with the generative AI chatbot. And this is exactly the sort of scenario that people are wanting in some sense. That they feel that, when you get really frustrated with the deterministic system, it’s because you’re trying to explain something, and it doesn’t have that in its vocabulary.

Whereas if you have a generative AI system, the hope is that this is more like just chatting with a human explaining your problem and then trying to understand what your problem is and having a big enough database to be able to do this.

[00:05:28] Lily: But I know that the downside with things like ChatGPT is it’s good on that style, but not the substance, something we’ve spoken about before, but does it say accurate things?

[00:05:38] David: And this is where if you train it on specific enough substance, then the hope is that actually for a narrow role that you might be wanting, think about this now, is what if your bank, instead of going through a sort of choice system, what if that was done by generative AI? What if you could rely on it to give you that information in a way which was reliable?

This is where exciting work is happening. And this is still early days, you don’t have many service industries actually building this out, because at the moment it’s quite, well, it’s expensive to do. And arguably what you’re actually getting is third parties trying to learn how to do this on behalf of clients, who would then contract somebody to do it.

And the hope is that’s a huge business model, which is just opening up, that you get your tailor made, generative AI interaction for your specific need, be it related to banking or your flight or all sorts of other areas. Now, of course, there’s danger in this, because with a deterministic system, you know what information is going out and what information is not going out.

With generative AI, at the moment, I do not believe it is possible to always be certain that the information going out is going to represent what you are happy as an organisation going out. So for a bank, to use such a system could be really not desirable. If you get a customer that gets told something which is incorrect by the system, and they now publicise this, this could get you into trouble.

So there’s good reasons why, particularly these sorts of sensitive areas, shouldn’t probably be rushing to have that better user experience because putting in place those structures is difficult. Now I do believe there’s ways around this, that generative AI could be used to improve what is currently available without losing some of those criteria, such as being certain of what is going out.

[00:08:22] David: This is where I think it will be a while before you’ll do your online banking using generative AI chatbots. But that’s not to say there aren’t other areas which may be less sensitive than online banking where you might find that you can do this.

[00:08:42] Lily: Yeah. And so in terms of us and the chatbots we do, we do this deterministic approach.

[00:08:50] David: We’re being pushed to work towards using AI systems in different ways and being challenged to think about how do we do so responsibly. And that’s a really big deal. And so very simply, yes, at the moment we use deterministic systems. We try to build them for very different purpose, not so much for individual information when you need on demand, but actually a number of the chatbots that we’re about, it’s more about a learning, almost a learning pathway, actually sharing, going through a program of information which is shared with you and which engages you.

And that’s a different way of using these chatbots, and there are lots of different ways, we’re not saying that one is better than the other. I think there’s a lot of misunderstandings around chatbots in different contexts and how they are used and how they can be used and maybe how they should be used. There’s some very interesting research about some of the values that you could get with these more AI based chatbots related to things like mental health.

And so there is some fascinating work happening, around positive uses of the more generative AI chatbots related to specific areas, which we’re trying to stay abreast of and understand what is possible. And there are also in the international development spaces, there’s elements of using the more deterministic systems for education, which are very powerful, as well as the sort of things that we do related to reducing violence against children and sexual reproductive health. These sorts of things.

And really, I would argue that a lot of effort has gone into this. And there is a lot of misunderstanding. And we find in so many places that we’re working with people who know that chatbots could really help them in certain ways, but where the uses of them and how they use and what they put together are really not thought through.

Even just things like safeguarding within chatbots is something which is actually quite difficult to achieve. We put quite a lot of time into this, because, actually, UNICEF wrote a wonderful document about safeguarding and chatbots. We took it very seriously. And so we’ve got elements of catching if people type random words which could mean that they might be suicidal or need help in some way, then the chatbot actually tries to interrupt what it’s doing and recognize that. And that’s what it recommended in this UNICEF document.

We’ve since found out that we’re about the only group who actually does this, that document exists but nobody else follows it because it’s too hard. Of course, we didn’t realise that it was too hard and we just did it. But it is, it’s been a lot of work and it’s totally changed the way we build chatbots.

But it is the sort of thing where if I think about now doing that with a generative AI chatbot, it makes me very nervous. That I don’t know that I would be able to put that safeguarding in place in the same way. And it’s created huge complexity which we hide within our systems, that is just there generically with almost all the chatbots we create. This is part of what’s in the background.

But if we were to now be moving towards using generative AI directly to users, then we can’t guarantee that in the same way. I don’t know that we could guarantee the safeguarding in the same way as we can now. And that’s not to say that everything that should be safeguarded against will always be guaranteed. But it is something where this can improve over time. If ever there are things which are missed, it can be identified and built into the system because it’s deterministic.

And so maybe I should switch to saying part of our thinking about how we think generative AI could be really powerfully integrated into deterministic systems, rather than necessarily replacing some of these chatbots.

Do you have any other questions you’d like to dig into first before I go in that direction?

[00:13:00] Lily: No, this is incredibly interesting. I work on the data side of the chatbots, but I didn’t know that there was this aspect. So if I’m understanding correctly, if you type in certain words into the chatbots, for the ones we work on, it will get out of that like deterministic algorithmic state of the chain that’s going down and go, no, we are now going to this part and overrides it.

[00:13:20] David: Yeah, if you were to put in something where you say suicide or whatever in the middle of what you’re typing into the chatbot, then it will flag that as being a safeguarding word, take you out of the flow you’re in, puts you into a state where it says, it sounds like you might need help. Do you need help?

[00:13:40] Lily: Sure.

[00:13:40] David: And then it takes you into another part. It’s all deterministic still.

[00:13:43] Lily: Yes.

[00:13:43] David: It’s taking you out of what you were doing. And we’ve done this sort of programmatically throughout chatbot. And this is sort of a huge system which was built on top. And we put this in so when people are authoring chatbots with us they don’t worry about this so much but this is then added on top of what they then do worry about. So this is then put on as sort of part of the safeguarding.

Now, that has caused its own problems in different ways, but these are exactly the sort of things that can get ironed out over time as you figure out what you want to happen deterministically. And as I say, we can’t take credit for thinking of this. This has come out of this really powerful UNICEF document which really had guidelines on this. Which, apparently, we seem to be the only ones who really took it seriously. Well, we’ve not yet come across another sort of chatbot builder working in that sphere, where they’ve taken this document as seriously as we have and really tried to implement this.

But this is something which is now baked into the chatbots that we build in many cases, because we understand why, given a deterministic chatbot, you can get into these loops which are just so frustrating, where you could be doing harm. We want to build chatbots that don’t do any harm. And so it’s been a very big part of the work. It’s been a lot of effort that’s gone in.

But let me take a step back and say okay how do we see really responsible use of AI in these concepts. And one of the things which is absolutely clear is that one of the reasons these deterministic ones don’t work well when you have something specific you want is because they’re not comprehensive enough. It’s a lot of time and effort to author and to think of everything people might want to ask.

As I say, let me just make this distinction again between the sort of chatbot where we’re trying to take people on a journey through a sort of almost learning material, which is really the heart of a lot of the chatbots we build, and these chatbots which you approach when you have a problem or a question.

[00:15:53] Lily: Sure, okay.

[00:15:54] David: It’s that latter which really needs you to be comprehensive. And this is where you get so frustrated. I’m going to give one example because I don’t think anyone’s going to take offence at this. But I, for whatever reason, I needed on a document the postcode which corresponded to the sort code I had for the branch. My sort code was from years and years ago, so the branch it associated with didn’t exist anymore. And so I didn’t know the current postcode that I should have associated to my sort code. And I thought, oh, at last, something which the chatbot would be able to help me with.

And it was so frustrating, I couldn’t get the information. I eventually did, but not using the bank’s chatbot. It was something where it was a very specific need and I needed it because there was a form which for some reason wanted the postcode associated to the sort code. And I had to put that in, otherwise I couldn’t complete the form. But I couldn’t get that postcode because the branch that used to exist 20 years ago, I am quite old, it was over 20 years ago, but when I started, that didn’t exist anymore. And so I didn’t know how to get the postcode that was now appropriate. I thought this is an easy question, I just go through it and it wasn’t there, for whatever reason it wasn’t part of what the system was set up to do.

And when you have that sort of system, which is so frustrating, when you can’t do the little thing you need it to do, that’s when if you talk to a human, they can say, no, actually, sorry, we can’t do that. We can’t give you that information because.

[00:17:35] Lily: Yeah.

[00:17:35] David: Or, yes, here’s the information. I understand what you need. But with these chatbots, I just kept going round in circles in different ways. It was horrible. And that’s exactly where, it is possible that with an AI based system, that could be improved and that could be done better, that you could actually get this.

And one of the things is you could have a much wider variety, you could actually feed in the interactions that have happened in the past and actually understand what are the things that people were wanting to ask as questions, and this is the key thing, you could use that to generate more expansive deterministic systems, which just have more options essentially open to you, which it recognizes, and which could be humanly verifiable.

[00:18:20] Lily: Yep.

[00:18:20] David: Yes, this is sensible to do. So you could use these AI systems to actually trawl through all the conversations that have happened and identify questions that should be answered by the system, which are not currently answered by the system and generate proposals for what would be sensible. which could then be human verified and you could then expand the system.

So that’s one way in which you could build better deterministic systems because you could use AI to be able to take the data that exists and identify new pathways through your deterministic system.

And the other thing, which is really frustrating within a deterministic system is, in a good deterministic system, you never want more than, let’s say, nine options because otherwise, you can’t fit them on a screen, you can’t interact with them in easy ways, and so on. But that means that you often get deeply nested things where you need to do press one, press two, press three for this, you’ve got to go really deep and you’ve got to listen to it all, or read everything to get that deep, and you just wanted to get this one precise bit of information.

And so the other thing that I believe that you could do through using, generative AI technology is do essentially a mapping where you use generative AI to interpret, what do you want to do today? And you type something in, and there are systems that do this now, or that exist, you see them. I’ve understood this, am I correct to understand, do you want to do this, or do you want to do this, or do you want to do this? These are the things that I think you might be trying to do, do any of those match?

And I think that actually, again, if I think of that system, and I tie that together with the other system, I think you could build, use generative AI to make these deterministic systems so much better, and they would just improve over time as there are things you identify people are trying to do that the AI system can’t do because it doesn’t have that option yet. You could then say, okay, these are extra options we should be adding because people are asking to do this, and so on. So I think there’s ways that you could build AI supported systems, which are still deterministic, that would far out compete anything that I’ve used recently.

[00:20:42] Lily: And this would have in it then, very easily and naturally, these feedback loops of this is where it’s going this is where it’s going wrong, then it can continually improve over time, like other successful AI systems that we’ve spoken about in previous podcasts.

[00:20:59] David: That’s absolutely correct. And this is what’s so powerful. So this has all the ingredients, I would argue, of a really successful way of developing an AI enhanced tool.

[00:21:15] Lily: Yeah.

[00:21:15] David: And it’s entirely responsible because it’s still deterministic.

[00:21:18] Lily: Yeah.

[00:21:18] David: You still have the sort of finite tree of things you can do. It’s just you can navigate the tree better because you can use your sort of linguistic ability to navigate it, hopefully. And you can actually get improvements suggested over the time which can then be human verified and human entered.

I really believe strongly this is the route that we want to go to build what I would be calling responsible AI enhanced chatbots, which are not what I would consider generative AI chatbots, because you’re never generating a response on demand. The responses are all existing within this deterministic structure.

[00:22:01] David: And so that approach is one which we’re really keen on. We’re looking to pursue and yeah, it’s something we haven’t done yet. And part of the problem to do this really well, from our perspective, it’s not that expensive but it is expensive compared to a lot of the things we’re currently doing. And so we’d actually be needing to get investment to really pursue this aggressively as a route to getting powerful AI enhanced chatbots.

We’re not the only people thinking about this. There’s other people who are pursuing these sorts of things. I would argue that most of the others I know pursuing these sorts of directions are maybe less cautious than us. We are maybe a little overly cautious. We’re really rather worried about some of the potential pitfalls.

[00:22:55] Lily: Yeah.

[00:22:56] David: And of course, in the speed in which AI development is happening, being overly cautious is maybe not a great strategy, because things are moving fast. So we’re likely to be scooped in this and other people will run away with it. But if they’re doing it responsibly, I don’t mind if other people take the idea and run with it and do good, responsible stuff. It doesn’t have to be us, I wouldn’t argue this is our intellectual property in any shape or form. This is just how I believe responsible chatbots could be enhanced using AI.

[00:23:28] Lily: Excellent. No, that’s really, really interesting and very insightful. Thank you very much. Is there any kind of final words you want to say?

[00:23:36] David: We should clarify that really neither of us are the best person within IDEMS to talk about this. So, I do have overall oversight of what’s happening in these areas, but I’m not the one getting my hands dirty at the moment actually making these things work.

And as you say, you’re on the data side of this, and so you know enough to ask some good questions. This has been a nice discussion, but we do have other members of the team who are deeply ingrained and deeply involved in actually implementing this. And so it’s an exciting piece of work. It’s something which we’re, I believe there’s real needs around it and real opportunities for the future.

[00:24:14] Lily: Yeah, it sounds very exciting. And it’s a really positive way of how, often AI is seen as doom and, oh, it’s going to do this and that, and it’s just a very positive way of how it can enhance these systems.

[00:24:30] David: Absolutely, and these AI enhanced chatbots I think they could be transformative in terms of how they build up the different things, the different capabilities in a way that would be totally responsible. Anyway, thank you, this has been fun, it’s nice, good fun discussing.

[00:24:47] Lily: Yes, thank you.