Erika Hall

Jam Mayer

Resources

Mule Design Studio

Conversational Design

Just Enough Research

ChatGPT

Jam

With ChatGPT's popularity and rise of AI tools, UX and UI design might just accelerate towards a more human-centred approach even more. What would conversational design look like in an AI future? Welcome to the Conversologist podcast, where we talk about the art and science of conversation in the digital space. We know that technology can be a powerful enabler, but communication and emotional connection still need to be at the core. I'm your host, Jam Mayer, and I invite you to converse with us. It's a great pleasure to have Erika Hall in our show today. She is the co-founder of Mule Design, author of Just Enough Research and Conversational Design, where you'll find the art of communication can elevate technology, and why conversation is the best model for creating a more human-centred design. Hello, Erika, and thank you so much again for spending your time with me today.

Erika

Hi, Jam. It's a pleasure to be here.

Jam

Now, I'm just going to dive into it because I know we had our first initial talk when we first met online talking about the show, and we were talking about ChatGPT and prompts, if you remember. I would like to actually continue our conversation around that topic because you did mention that, look, actually prompts versus just talking to ChatGPT, there might be a difference, but prompts might actually produce better results. Or would it matter? What are your thoughts?

Erika

I think it really depends on what you mean by 'better results'. All of these tools are very early and they're changing all the time. The models are learning and they're releasing more and more sophisticated models into the world. The tools are training humans just as humans are training the tools. I think that's an interesting space we're in right now because we've all learned, through web search, how to form really good search queries. I think from a user's perspective, right now, everybody's testing out or just been working with them like this, with a tremendous amount of excitement. I think right now there are so many questions swirling around about how much of using these tools is hype, how much is reality, and how do I, as a person interacting with them, get the best results out of them.

Jam

Sometimes it's better. I saw that you actually posted it on LinkedIn. But it's sometimes better and it's more efficient, probably not producing better results. But it could be more efficient in terms of talking to ChatGPT. So prompts come in. That's why it's all over the internet. Interesting that you mentioned. Well, sometimes people just want to go right straight to it, and the prompt can do that, versus spending, I don't know, just having a long conversation with ChatGPT. It's really more of efficiency.

Erika

Yeah. Again, going back to people learning from using these systems, sometimes people just want the thing that they want. It's like they want direct navigation as opposed to a conversational interaction.

Jam

We're still experimenting on it. Are we actually going to get something else or do we focus on the time spent talking to ChatGPT so that we can get to our end goal? I guess that's the whole point, right? Or do we just go for efficiency, but do we actually get the thing we really want through the prompts? Focusing on the industry, just wondering if you've seen any actually using the tool in their work?

Erika

I haven't, but like I said, everyone is trying these things right now and playing with them and seeing. I'm starting to hear stories from, say, marketing copywriters who are losing freelance gigs because of it. But I feel like that's not totally a short term phenomenon, but I think it's a false economy to say that you can just fire your human and replace it with a system like ChatGPT because a lot of human intervention will be required. I don't know if you saw the Ted Chiang piece in the New Yorker, maybe a couple of months ago now. He's an amazing science fiction writer and a programmer. He wrote the story that the movie Arrival was based on. I highly recommend all his work. But his analogy was that ChatGPT was like a blurry JPEG. It's lossy. It's like a blurry JPEG of the web. It puts sentences together based on this predictive algorithm, but there's information missing. The difference is if you look at a blurry JPEG, you can tell that some detail is missing. You can tell it's not accurate. You see weird artefacts because of the compression algorithm. But the problem with using these language-based systems is you can't tell what's made up unless you know.

The trouble is that sometimes the answers are pretty good. I've heard it described as okay, like a substance. If you ask a question, it's an okay Wikipedia article. But the problem is Wikipedia articles, good ones cite their actual sources. And we're starting to hear a lot of examples of these large language models making up citations wholesale, because they sound good. The systems give an answer that sounds like a real answer, but they don't fact check. They're making things up. They're so called hallucinating. The problem is if you're asking a basic question, you'll probably get something right. One of the best examples that I've heard of is someone who took an elderly relative in for some tests for health care, and they got the results back, but the doctor wasn't able to meet with them yet. And they were really concerned. They're like, What's going on with my relative? I'm not a doctor. And they fed the results into ChatGPT, and they were able to get a plain language interpretation of the results to say, Oh, these results mean this. That's all they needed. They weren't going to be doing treatment, but they just wanted to know how bad is it? What is the issue? What are the related treatments or prognosis I might be looking at? For something like that, it seemed really useful where it's like, I need an answer that's 60% dependable, but not the details.

And so in cases where you're looking for something like some sample marketing copy, because marketing copy is usually high level. It doesn't need to be totally accurate a lot of the time. It's to give a feeling, to give a flavour, to give a little pitch. So for something like that, maybe. But anything where you need up to date information and you're not going to get the proof because you're not going to know which pieces it's making up. But the question is, how much is that a time saver if you then go have a fact check? And that's the other thing I've been hearing about working with me is to say, I tried using it and then I had to go back and correct all the errors. That was more time than just writing it from the beginning. It's early days, though, and so I don't want to say anything like there's nothing like a best practice right now because these are experimental tools, right?

How are they going to replace or augment certain human tasks? Let's find out.

Jam

I've seen a lot of videos out there or have seen, well, Google search results saying, hey, here's an article. If you want to create a website from scratch, just ask ChatGPT to do it. There are other AI tools. Obviously, that's just the hype right now. It's the most popular. But any thoughts on designers or web designers, to be specific, who are using this tool? Is it a good start?

Erika

I think it gives you a certain consensus reality. If you ask it for something, it will tell you probably what an average of what people would say something like that would look like. And so maybe if you're looking... Because I know sometimes it's hard to find good examples of things, especially when you're starting out. And then if you step back, this has always been an issue for us. If we have something in mind and then you go out on the web and you're like, Give me a good example of that, you start to see that the way things are in reality. Because even if you thought an organisation had, say, a really well designed website, over time, they really start to degrade for use a lot of the time because different things get added on and the systems get modified and things get deprecated and a new CMO comes in and has thoughts. Or maybe they did a big strategic design and then acquired another company. And then they're like, oh, well, we get our web strategy from the ground up, so let's just bolt that on. And so it can be hard to look for good examples in reality.

People go and they look at Behance and Dribble or whatever to see examples of things. Before, if you're just looking for examples or looking for inspiration, fine. But as far as making a website is one of these tools, is using it easier than using Squarespace or Wix? Because there's so many kits out there that are set up, or tools that really have everything built in and you have to put your content in anyway. And so especially if you need the content management system behind it, I suppose it could give you a starting layout and you're like, oh, I want it to look like that. If you're feeling your way around, which is often a part of the beginning of the creative process, you're looking for inspiration, great. But when people talk about the way it will replace certain jobs, there are already tools out there. If you're a small business, you can use Wix or Squarespace to knock together a perfectly fine site really fast. And maybe you have somebody help you out with it. Things probably seem easier to me because I've been in the business and maybe it's less easy if you have a pet store and you don't know anything about the internet. Maybe you need somebody to help you out with that. But you couldn't just use AI and totally eliminate an expert because you wouldn't know it after. So you need an expert anyway. So it's like, are you getting a WordPress expert or a Squarespace expert, or are you getting an AI and some other back-end expert?

So I don't know. But like I said, it's super early. None of these things are totally ready for prime time. We're still finding out capabilities. One of the things to keep in mind is the systems that we're working on and designing are all integrated with other systems. So it's not like just because you use AI for one part of it, that doesn't make it interoperable with everything else it has to be interoperable with. I've always talked about the fact that it's the things under the surface, the design you don't see that determines whether or not the design is good. What makes a system human-centered? Well, if it's got a human-centered, ethical business model, if the underlying business model isn't good, you might put something beautiful and elegant on the front-end, but that's just going to be a way to involve people in an exploitative or extractive or otherwise unethical system.

Jam

And this is a great segue to, well, the principles. I mean, it's in your book, the Principles of Conversational Design. Just for the listeners who may have not or they may not know what conversational design is, probably you can give us a quick summary or definition of what it is and maybe just some of the principles in your book. Then we can see how the principles might actually be impacted. Maybe it may change, maybe not because of this AI... I don't want to say hype because I know it's real. We're in that AI experimentation stage. Let me say that. What is conversational design anyway?

Erika

Well, the way I think about it, because I know people think about it in different ways, is using how humans converse with each other as a model for how to design computer systems so that you make them easier to work with. This is at a deeper level than just the mode of something that... I'm not talking about voice interfaces, I'm not talking about chat text interfaces, but I'm talking about thinking about what really makes conversation work. The book, I drew a lot on the linguist and philosopher Paul Grice in his work, and he worked in the field of pragmatic. It's language in use and the context of use and turn taking is a big part of it. And what's implied in conversation, not necessarily just the syntax, but really fundamentally, what makes the conversation work at a deeper level. And some of the key principles, he articulated a cooperative principle, which means fundamentally, what makes the conversation work is that the people involved have a goal and they want to cooperate with each other. so I'm sure you've noticed this. If you've had a conversation with someone and the conversation feels really hard because the other person is fighting with you a lot or calling out your mistakes, or they just talk too much and they don't let you get a word in, or they cut you off. The goal of the conversation is that it feels irrelevant or they're not being truthful, they're avoiding answering the question, or they're outright lying to you. Those principles of conversation can be applied to systems, too. It's like a conversation should be cooperative. It should be turn-based, which means, yeah, it's just like this, back-and-forth. That interaction, that makes something conversational. It's not like a monologue. It's not like I'm handing you a document. It needs to be to the point and proofful and just enough information. Not too much information, not too little information.

And then another linguist came along, Robin Lakoff, and she added politeness. A conversation should be polite and respectful and supportive and all these things. So when you're designing an interaction between a human and a computer, it should have these same principles. It should feel, it should be turn-based. It should respect your time. It should be polite, all of these things, cooperative and error tolerant. That's another key principle. This is the one where you can really see how some of the systems that were called conversational maybe five years ago work, really.

If you've used a voice assistant or a tap bot thing, you have to get your input just right. Say you're talking to the Amazon Echo using Alexa, you have to really make a request that it will understand. And if you say something that isn't a request and it recognizes or can deal with it, it's just like, Sorry, can't help you. Sorry, can't help you. The same thing with a lot of chat interfaces. You have to be so precise in giving it an input that it can handle. And you contrast this with Google search. And Google is very good because Google has done trillions of queries over the years and has learned over time. But you could type misspellings into Google and it could figure out what you mean. And you can tell this in conversation with other people if you use the wrong word and they know what you mean, right? And they don't stop you and say, I'm sorry, that's the wrong word. What do you mean? And put you on the spot. And that's really an odd feeling because we all misspeak. But if you're in conversation with somebody and you say something wrong, but they're really cooperating with you, they're really on your side and they answer your question and they continue the conversation while getting your meaning, that's a huge part of it.

Being really context aware is another part of it. And that's the hardest thing for a computer system is context. People are great at context. If you're talking to somebody in person or even on the phone, you can tell, Oh, this person is in a hurry. This person is dressed. Oh, this person seems really happy. Oh, they look like they need help or they're struggling. And you can adjust. Computers can't adjust to that. And so a website can't tell, Oh, this person is really in a hurry, so I'd better not show them all these pop ups. But just going back to error tolerate. They either prevent errors or really help you recover. And that's the area where I think any system can be improved to make it easier to deal with. If you can anticipate, what are all the ways that the person using it can mess up the input? The worst is if you have an address, like a phone number field, for example. There's no excuse for a phone number field just being one input and then yelling at you. I don't know if you have this experience where it wants you to put in the dashes or something, and you're like...

Jam

Oh, yes.

Erika

Did you ever have to try or the worst if a form is really poorly designed and the tip for the instruction disappears when you're typing and then it yells at you? It's being a bad conversationalist.

Jam

There's the desktop version and obviously the mobile version. Sometimes the desktop is perfect and then when you move to mobile, you go, I can't even see the fields. I mean, the text or even I'm not even so sure where I'm going to put the text because it's so small or maybe it's just too big so it just goes out of the screen. So it gets frustrating.

Erika

So what I would say is that even if you're designing a GUI or a system with a visual component, it's exactly what you're saying, that the interaction is interrupted in these ways that violate the principles of a good conversation. And that's why conversational design is applicable to all interaction design. Because if you start not by thinking of the layout, because we're not doing... Even if there is a graphic element, it's not graphic design. There are visual elements, there are graphical elements, but it could be really easy to think of it as a static page and design it as a static page. We even have that metaphor. Those visual aspects of the design should follow what you want the interaction to feel like because it's an interaction in time. There is a spatial component, but the most important part of interactive systems is thinking about how you're using somebody's time because that's what people feel. They feel like my life, the sand dripping down out of the hourglass. I'm losing life force by interacting with this system where I'm just trying to fill out a form. I'm trying to purchase something, I'm trying to get some information. And if it feels like, oh, there's a really good back and forth. It's giving me enough information. It's asking for enough information to keep it going. It's not making me feel stupid. It's making me feel smart and respected. These are qualities that should be a part of any interaction. And then you get to the point of, once you have a sense of, Okay, in what moment or context in somebody's life are they going to be having this interaction with the system? What do they need? What's the timing going to be like for these different moments? Once you have that, then you can think about putting the visual design on that. But so often still, design starts with, Oh, we're going to make some boxes, and then we're going to make the boxes hop around. As opposed to, We're going to design the conversation. The interaction between the person and the system first, and then we're going to think, How does the visual aspect of that support it?

Jam

I love that. Knowing the content as well on input and output, is that something that we also need to consider?

Erika

Absolutely, because that's the meaning. There's the meaning underlying it, and you layer things onto that. So you can't design an interactive system without meaning unless it's just a tool that you're using that doesn't have any verbal information. There are things like that, like a remote for your TV is just a tool. So even if it were a purely digital remote, maybe there's no content there. But for anything that's a service, you need to have the language there first. And sometimes it helps to think about it, like, if this were a person, you wouldn't care what the person looked like if it was somebody trying to perform a role for you. For example, say a lawyer, you wouldn't, and maybe people are going to disagree with me, you wouldn't choose your lawyer based on what they looked like, right? You wouldn't say, Okay, the most important part is I want my lawyer to look like this. And then I hope that they have a good knowledge of the law I need help with.

Jam

I love that. Yep.

Erika

You would start with the content. You'd start with, oh, I need a business lawyer. I need an IP lawyer. I need a real estate lawyer. You'd start with, what law do I need help with? And then you'd layer on top of that, Oh, it'd be nice if the lawyer's got to go to court, so the lawyer's got to put on a suit or something. If you think about an application or a website fulfilling a role, the important thing, you start from what does that system need to know? How does it need to sound under what circumstances is somebody going to interact with that system? And then you get up to the point of, well, given all this, then what should it look like? We look at Craigslist. Craigslist is over 20 years old. It's still going. There are a lot of other websites that people use now, so I don't think it's not as dominant as it once was. But it was a page of links. Then at some point, they added photos and they have cleaned up and refined the look of it over time. But it's still fundamentally a page of links and some photos. But underneath the way that Craig and his team designed the business rules, that was the important design.

The things that they decided to charge for and the things that were free, that made it work. Those choices, those design choices, made the company very successful and last for decades now, where other websites and apps have come and gone. Other internet businesses have disappeared. And that one retained its utility because of the way they thought about what was important in the interaction. What was important was speed. I think that's the thing we were talking about, keeping the conversation going back and forth. I think that the perception of something feeling very fast and feeling very easy to use and error tolerate and not like you've got to sit there and weigh your choices. That's something Craigslist, that real right because the lack of a lot of visual polish makes it feel really fast. You can scan down. If you're looking for an apartment or you're looking for a chair or something, you just scan down this list of links and if there's new stuff, it just pops up to the top and it feels really fast to deal with and it doesn't feel like, Oh, there's a lot of polish.

Jam

I also think it's simple. I just want to focus on the time element because you've mentioned this a few times already. It's the time element. Just moving a little bit on the AI side of things, a lot of people are saying that, Well, you should go and use AI tools now, regardless of what they're for. Because it's faster. We're time squeezed professionals, so we should use AI. Use AI, AI, AI, AI. What are your thoughts on that in terms of when you start thinking of the principles of conversational design, putting in time in there and then add all of it and sprinkle it with AI?

Erika

It's important to keep in mind that there's been machine learning, so called AI, in a lot of systems for a long time. I just think it's just with ChatGPT, all this and we have these. The large language models, people are like, Wow. But there's been a certain amount of AI in a lot of systems. But now it's like, oh, it's got... It's like everyone, oat bran was in everything. Now it's like, oh, it's got AI. Ai is in everything. It's it. Hell. But it's not necessarily better. So I'd say, try these things, but it's all about what you're optimising for and what's important. If accuracy is really important. A great use case I heard was for academic researchers who do not speak English as their first language, but want to submit to some of the more prestigious journals like Nature or Cell. There are some really big journals that if you get published in that journal, it gets your paper out to a much wider audience and it's very good on your CV. It's very good for your professional reputation and it's very good for sharing your research. But historically, researchers who didn't speak English as a first language were at a disadvantage. And so a use case I've heard is that you can use these tools to help adjust the language in your paper into the language of the reviewer so that your language style doesn't get in the way of the science.

And that's a great use case. So there are places where it can help you reduce the barriers. And there are places where it can help reduce bias. But the problem is there are also places where these tools can exacerbate bias because they've been trained on biassed data because humans are biassed. And so the case I heard about with the researchers who were able to submit, have more advantages submitting their papers so that the language didn't get in the way of science in the eyes of the reviewers, that's a great use case for removing bias. But there are other cases where it's very easy to end up in a bad place. And so, write these things, but always keep the principles in mind. Like Ezra Klein, who's an American journalist, he wrote a piece in the New York Times a while back talking about the business models for AI and saying, That's the place where we need to really look at it. Because if you have these systems that are just trying to get people to look at more ads or trying to manipulate attention because the underlying business model is, Oh, steal this person's attention away from their primary tasks.

If you take that and you use AI to put it on rails, that's bad. But you've got to, again, fix that at the business model level. So much of what's important about this technology or about how it's going to influence design, again, it's the underneath stuff in terms of what are the principles? Are you really trying to, again, cooperate with somebody and help them meet their goal? Or are you trying to take their attention away from them and divert it to something that's good for your business but not good for that person or not good for society? That's the part to really pay attention to. I mean, it's fun because it feels like these things feel magical and easy to use. I just type in a few words and I get something. But then you have to say, but did I get the thing I needed? That's the question. Just because you got something really fast and you got something that looks good and looks like something that is really well made. Again, it's going to that fuzzy JPEG. Is it something that just looks true? We have so many problems with misinformation already.

And again, I think so many of the dangers that people are worried about are their dangers we've already been not managing very well. I know there was an open letter calling for a six month pause on developing these systems so that we could figure it out. But that's not going to... I think it was good... It's good to put these concerns out there. But stuff's happening, and I think the conversations we have to have are being really clear on what's your goal, what are the underlying principles, what's your code of ethics, how does the business make money? Are your goals really in alignment? Because this is what recline was saying again, that we have alignment problems. We have the problem that human wellbeing and corporate interests are out of alignment.

Jam

Oh, gosh. Yes.

Erika

And that's what we have to fix. So there are all these places where the amount of power, say, an employer has versus the amount of power that somebody working for them has. Those could be really out of alignment. And those are the things that we have to work on and fix, and we always have to ask ourselves with any new tool or technology, what is it promising to do? Is it really doing that or is it just the new shiny thing? Because there are so many things that websites are great for. I always use my bank as an example because I use Bank of America and they were one of the first banks with a chatbot out there. I am particularly irritated because it's Bank of America. It's called Erika. And I'm like, and then I tap on Erika. I'm like, no, I don't need to have that conversation with what I need from my bank is really straightforward, and I just want to click on things. I want to click on what's my balance. Oh, I want to click on “transfer money” from this account to this account or send money to this person.

I don't want to necessarily… Autopay is great. But then it's a question of these agentive systems, the systems that act on your behalf. How often would you have just a person acting on your behalf? How often do you need concierge versus self service where you're doing the thing and making sure it's being done right? Just because something is so called AI doesn't mean it's better at making the choices that you want made. Say if you book your own travel, you know you're like, here are all the flights I could take, here are all the prices, and you can see all of your options and you can see all the rules. But if you got on the phone with somebody, which is what everybody used to do to book travel, imagine being on the phone and saying, find me the best priced ticket to Hawaii. And you'd have to trust, right? You'd have to trust. If they came back and said, yeah, this is the best ticket at this time, I feel like a lot of people still want to see all the options, right?

Jam

Oh, truly. Yeah, definitely. Like myself, in terms of options, let's just go back to the chatbot and stuff. I would prefer because we don't know the different users. We can always have a perfect system, but we don't know the different users that's going to land on the website or chat with a chatbot, whatever it is. There are some people who would rather actually talk to a person or an AI and just keep on talking. They've got all the time in the world. We can't take that away from them. But that's why phones still exist, I think. And as you said, there are people who are just in a hurry. Give me the pricing. I need the date, this date. Done. Book it.

Erika

Yeah, totally. I think there are a lot of assumptions about what fast means and what efficient means. Yeah, it's going to mean different things to different people. Getting good service means different things to different people. I think this is going to be, hopefully, adding in this machine learning will make systems function better. But again, it all has to do with what the systems are optimising for. Are they cooperating? Going back to that cooperative principle, are they in alignment? Is the business goal in alignment with the user goal? Or is it trying to divert you to something else? Oh, gosh, Bill Gates had a really, a quote that I've used before about automating systems. If you have a process and it's working well and you automate it, that's great. But if you have a process that's not working well and you automate that, it's going to be terrible. That's essentially what a zombie apocalypse is. You've got zombies and you automate them and they just eat brains, brains, brains, brains, brains.

Jam

Just to wrap up the conversation, I'm absolutely loving this. I have more questions. Maybe we can do another episode. I don't think AI, at least today, will change the principles that you've mentioned. I think it is going to still hold firm. And I love the fact that it should still hold firm. I don't think it's going to change even in the near future. AI can help, but not necessarily the principles, the alignment. You already mentioned it.

Erika

Yeah.

Jam

That's good news because a lot of people are afraid and things are going to change.

Erika

No, things could totally change because we haven't really touched on... There's going to be emergence phenomena out of this. There are going to be things that the designers didn't intend. Once the machines are all talking to each other, we're going to see new properties and new things we have to deal with happen, and they're going to be unintended consequences. I'm not saying don't worry. I'm saying worry, but always go back to we've got to think about what we're optimising for. We've got to think about what the... Because the whole point of any technology or any software is to help people do something. It's all about making sure you have good goals and good guardrails in place. If you unleash these systems and you just tell AI to maximise profit at all costs or something, no matter what, bad things could happen. The most important thing is to want to not destroy the earth and want to not destroy human society. You start with that, and then you can start to see where to put in the checks and the guardrails and the regulations to make sure that this technology doesn't destroy things too bad.

Jam

Totally agree 100%. Well, thanks so much, Erika, for letting me pick your brain today. I mean, there's still so much to talk about around this topic. As you said, it's very early days. And hopefully when you have time and you've got new info, we could do another episode in the near future.

Erika

Everything will be different. Six months from now, everything's changed.

Jam

Definitely.

Erika

The robot overlords are ruling the Earth. Yeah, we'll be like, Oh, well, gift.

Jam

I know, exactly. Anyway, for the listeners who are listening to this podcast, thank you for your time and check out the show notes for any links and resources that Erika has discussed today, including where you can get Erika's books. I highly recommend you get them, especially if you're UI and UX. And I have a few friends who are already saying, Oh, my God, I want to see that.

But hit that follow button or bell wherever you're seeing this to be notified of our next episode on your preferred social media platform. There's heaps of them out there. We are primarily, of course, on Spotify and you can subscribe to our show. Go ahead, choose your favourite podcast app and hit that Follow button. Thanks for listening and spending your time with us. Remember to keep the conversation going.