How Next-Gen Data Analytics Powers Your AI Strategy with Christina Stathopoulos, Founder at Dare to Data

[00:00:00] Richie Cotton: Hi Christina, welcome to the show. Thanks. I'm excited to be here. Wonderful. Just to kick off, I know you've been playing around with AI agents for data analysis and I'm curious, do they work yet? 

[00:00:11] Christina Stathopoulos: That's a good question to start off with. I have been playing around with these AI agent data analysts, data scientists.

[00:00:17] What I've been trying out a lot lately is Julius ai. And I think for what it could do, it's amazing. It can program in SQL, in Python. It will walk you through the steps they're taking to get to the conclusion. Like they don't just give you, you don't ask for an analysis and it just gives you the output.

[00:00:33] It'll actually take you step by step through what it's doing. It will output all of the code that it's generated to get to there. It even makes like recommendations of final, how the final insights can be applied to the business. Because, we're not just crunching numbers to crunch numbers.

[00:00:48] You've gotta generate some value for the business. So we'll even link it to the business at the end how you could use these insights. So I like that a lot. What I've also liked is that you can connect it natively to databases. That's a new thing that I've been trying as well. You can connect it directly to your databases.

[00:01:02] You don't have to, export the data and then re-upload it to the tool. So it works very nicely. But I would say that, up until a certain point, you still need the human data analyst or the human data scientist working together with it, because I know we, we hear a lot of people talking about that, AI's gonna replace us.

[00:01:21] Or even data professionals are saying, this job is going to be irrelevant in five, 10 years. I don't believe that at all. Because I'm seeing what it's like to test these tools. You use it as like a sidekick but you can't completely rely on it because it does make mistakes and it also doesn't have all the, the business context that you have and the experience that you have.

[00:01:41] So it's really like you have to combine what you know, your experience, your knowledge together with the tool. And and it can do amazing things, but you still need to be there to validate the output and make sure that it's staying on track. 

[00:01:53] Richie Cotton: Okay. I suppose that's reassuring in. In several ways.

[00:01:57] Firstly, the it's nice that there are some pain points being sold beyond just generating some SQL code automatically to crunch the numbers. It's also giving you business insights. It's also reassuring that we're not gonna be replaced by robots immediately. I'm curious then as to.

[00:02:12] How are your workflow changes? If you are doing data analysis with with one of these AI assistant does that change your workflow at all? 

[00:02:19] Christina Stathopoulos: I think it does, but in, in a good way. Like I said before, it's you should treat it like, like a sidekick, like it's supercharging your workflow.

[00:02:27] So the way that I'm seeing it changing my workflow is that I've got that sidekick working together with me, and it's helping me, it's helping me move along faster. It definitely helps, like with the brainstorming process, it can help with exploratory data analysis to, to generate things or to come up with ideas that maybe you didn't initially think about.

[00:02:46] It can help you move through that analysis much faster than you might do without the tool. And I've seen that it helps me even think outside of the box, like I mentioned before, the brainstorming part. Even when it's drawing conclusions too, it'll help me think outside of the box. It'll help me draw conclusions.

[00:03:01] But all of this, I'm saying like with a grain of salts, because you still need to be there to validate what it's doing and check it because they make mistakes. The tools that I'm using every day that have these AI features that I'm adding into my workflow, they still need to be, you still have to babysit them.

[00:03:17] 'cause it's like a, it's like an assistant, but they're not a senior. So you need to check what they're doing, but they can help you get. To your results, your outcome much faster. I think instead of, needing to double check things that you're doing or the debugging that you have to go through.

[00:03:32] They can help you get through that much faster. They can help you with brainstorming and so on. So I think it's more like it, it just gets added onto your workflow and it helps you move along faster. 

[00:03:40] Richie Cotton: Ah, that's interesting. 'Cause they're making mistakes, I guess it seems like it might be a really good fit for the stuff you mentioned, like the brainstorming, the exploratory data analysis.

[00:03:47] 'cause there you're just trying to ask a lot of questions, get a lot of answers, like really quickly. And if you make mistakes, it's fine. It's only you we're gonna see them. Whereas if you've got like a final report that you're showing to your boss. Does the stakes are higher there or even if you're showing to a customer, then, the stakes are very high in terms of getting it right.

[00:04:06] Do you find it's more helpful than in the earlier stages rather than the latest stages of an analysis? 

[00:04:11] Christina Stathopoulos: I think, it's helpful in both. It's right that like towards the end of the analysis when you're going to be presenting to stakeholders, like the pressure's on, so you really need to double check it there.

[00:04:20] But I think it's about even between both of them. It's helpful in both because at least in my case, like I'm doing the same level of double checking. Like I, I read through all of the work that it generates after it's done, all of the code that it's done. I double check it no matter if it's the beginning or the end, because I found that it makes mistakes in both leases and you would hate, for example, in the brainstorming process or the exploratory data analysis to just gives you an output that you really and you don't really double check it.

[00:04:45] And then you realize later on that it's made a mistake. Because then you're gonna, you know that mistake you're gonna pull through the entire analysis. It's true that you just don't have that as much pressure maybe at the beginning, but you wanna make sure that those big mistakes don't get drawn out till the end.

[00:04:58] Then you have to go all the way back to the beginning. 

[00:05:00] Richie Cotton: Yeah, that's fair. So if it suggests like a really. Inappropriate analytical technique, and you just take it like blindly okay, this is gonna work. Then your entire analysis ro and your whole project time is wasted. So I, yeah, I can certainly see how there are big problems even at the start of the year the workflow in terms of the different skills you need.

[00:05:18] Then if you are, if you're doing AI assisted data analysis or AI assisted data science do you need to learn different things? I guess proofreading is like a really important skill now. But beyond that talk me through what do you need to learn. 

[00:05:30] Christina Stathopoulos: That's a good point. You need good proofreading skills.

[00:05:32] Now we didn't need that as much before as analyst or data scientist. I think that as far as like the skills, I don't, at least in my opinion, I don't think the skills have changed too much. All the foundations remain and I think they always will. Nothing has changed in that sense. So it still means that as a data analyst or a data scientist, you still need the fundamentals you need.

[00:05:54] Statistics and probability you need sql. We mentioned you need data visualization. You need to learn how to program in Python, these types of things. I don't think they, they change. And also you need like this foundational knowledge, which would be like understanding how data works, understanding where it lives, how can you query it, how can you tell a story with it?

[00:06:14] These are things that you still need to learn. So those skills don't change. The only thing is that, like I mentioned earlier, you add on like this new layer of having this sidekick, this assistant right there with you, and it's supercharging what you're able to do. So you're able to get to your results faster.

[00:06:30] You're able to think, it's able to help you push and think outside of the box. So I think that's the thing that has changed. So in that sense, the new skills, I guess you would need is. To understand like how you can incorporate it into your workflow, like how can you work side by side with this tool and make sure that it's accelerating your process and not slowing you down?

[00:06:48] It might take some, it does. I do think there's like a learning curve with it and you need to practice. It's going to take some time at the beginning. Once you can like, get it natively into your workflow, the way that you work, it can accelerate what you're doing, what you're able to do. So really the main, new skill is how can you add this into your workflow?

[00:07:05] How does it work within your tool set? And then proofreading, like we talked about like you mentioned, that's a really good, a very important skill to have. 

[00:07:13] Richie Cotton: Yeah it's interesting the idea that there, it's basically like a second colleague. So it's like when a new person joins your team, always takes a bit of time to figure out how are you gonna work together efficiently and effectively.

[00:07:23] And yeah I can certainly imagine there's gonna be like, oh, you've gotta plan for having extra time just to work out how to make sure you're gonna work together. Okay. Okay. You mentioned data visualization as being one of the important skills data around is data scientists need.

[00:07:35] So this is like a perennial important skill. It was a bit of a scandal at the start of the month. So Open AI released GPT five there, this big flashy demo, which I have to say was very impressive. Demo mixture of reviews, G PT five. But one of the things was there was a an AI generated plot and it was terrible.

[00:07:53] Caused a minus scandal in data visualization circles. So a gen AI is still bad at creating plots. I'm wondering like. How do you do AI assistance for data visualization? In a good way? 

[00:08:04] Christina Stathopoulos: The whole scandal I was on top of it, it was crazy how bad the plots were that they generated.

[00:08:09] So when it comes to data visualization, this is one of my specialties, so I'm very picky about it and I do not think that AI tools are great at data visualization. We saw that proved by open ai, they had GPT five, which is supposed to be. So powerful, but it can't even do the most basic of data visualizations.

[00:08:28] I think they generate like ugly looking data visualizations. I think they make very basic mistakes. So you need to be very careful if you're using those to visualize data the way that I use them in my workflow. If I'm visualizing data, it's not necessarily to create that final plot. As you shouldn't do, because if that's what they did for GPT five, then we see the final results that it gives.

[00:08:51] So I don't use it necessarily to create that final plot that I'm gonna use, that I'm gonna present or, give to stakeholders. But I will use it like we talked about earlier, like brainstorming. I might use it for the brainstorming process, like to see what suggestions it has for how I.

[00:09:04] Visualize the data and it will give me examples and I can check out those examples, but then I would build my own plot with my own tools and my design specifications with the data. So I wouldn't only rely on it to do that. And then you can also see what interpretations it gives of the data visualization.

[00:09:22] So the ones that it's generated or the ones that you've created, you can provide it to it, and then you can see what interpretations does it take from that. Or what business conclusions would it take from this? And you can test that against what conclusions you are making. But again, like we've been talking about up until now, you had to take everything with a grain of salt because it can still make mistakes.

[00:09:42] I was using one the other day with data visualization and it made a really basic mistake because it was interpreting a graph that it had generated actually, and it was saying. It was like saying the opposite of what the chart said. Pretty much like it was comparing like male versus female, and let's to just to make it simple, like comparing male versus female and saying, the males have a higher spending rate than the females.

[00:10:03] But if you look at the plot, it was actually the opposite. The females had a higher spending rate than the males, but I don't know why it interpreted it the opposite way. So if you had taken that blindly and ran with it. You would be making a wrong interpretation, but it just takes, a little bit of experience and just the proofreading to recognize like, Hey, no this is wrong.

[00:10:22] That's completely opposite to what the graph is actually saying here. 

[00:10:26] Richie Cotton: Yeah, that sort of stuff is very dangerous. 'cause if you don't proofread it, then again, when you present a your stakeholders, you can look like an idiot. 'cause that's a really simple kind of plot. It's even someone who has just a rudimentary understanding of like how you interpret plots, they're gonna spot that.

[00:10:42] And yeah you have to be very careful there. And I agree. I'm the same. Like I get very picky about data visualizations. I'm like, I wouldn't have enjoyed it like that myself. So yeah. Enjoy. Do you have any tips on great data visualization? Always happy to hear that sort of thing.

[00:10:56] Christina Stathopoulos: Yeah, I could give, I could go on and on about tips. I'm very involved when it comes to data visualization. I even I'm an adjunct professor at a business school, at IE. Business school in Spain, and I teach a course there on data fluency where we focus on the last mile of analytics, which is data visualization, data storytelling.

[00:11:13] And I also lead that course across the business school. I'm always coming up with like new materials, new frameworks when it comes to data vz. I love it. But I think a couple of like very important tips for everyone, and I think AI should be listening to this and applying these tips as well, 

[00:11:30] Richie Cotton: is I hope episode of Data Frame gets scraped by the open AI and philanthropic bots.

[00:11:34] Yeah. 

[00:11:34] Christina Stathopoulos: I can only hope. I can only hope. So I think the number one tip that I always give is that. Simple is best and less is more. So you don't need to overcomplicate things. You don't need to make these overly exaggerated or trying to put too much like prettiness into the graph when you have to remember what's the purpose of this graph?

[00:11:53] What are you trying to show? And focus on that and do it in the simplest way possible so that your target audience can understand. They can get the message immediately. Another thing. Is to always design with your audience in mind, which kind of connects back to that. But just keeping your audience top of mind that you're not building the data visualization for yourself.

[00:12:12] You're building it for someone else to read it, to get whatever message you want to transmit to them. So always design with your audience in mind. If they are a more advanced audience, maybe if they are data practitioners or data engineers, then they have more technical knowledge and maybe they can understand a more elaborate, a more complex data visualization.

[00:12:30] But if. You're talking to like business stakeholders, directors, not technical people, then try to stick with those more basic graphs, what they know so that you don't confuse them and so they can get the message quickly. And then if I can add just one more tip, it would be that it's not just about choosing the right graphs, that's important, choosing the right graphs for your data, but it's also about choosing the right colors.

[00:12:53] I see colors are overlooked a lot or the importance of them. They're incredibly important. So I think just having, reviewing some of these basic things that we learned in kindergarten, but a color theory and understand how you can use those effectively. Your data visualizations. I wish I could show the screen right now.

[00:13:12] I have it on my phone. No, because right before we were recording this, I was watching the news here in Spain. I had my husband like, pause the TV because they were presenting like this map and it was like, those typical color coded maps and I had a legend on the side. It was the, it had made really big mistakes with the colors, and I was like, pause this for a second.

[00:13:31] I, I need to do a picture of this data visualization. My husband's probably what is wrong with you? Everything has to be like data focused. But I took a picture of it because I was like, this is a perfect example. I. Of them not thinking about the colors and not even thinking about like these simple design things.

[00:13:47] And we see this all around us all the time. Like I just said, this was on the news. You see it on websites, business, reports. It's really simple things that people, it's just because they don't take a step back and consider those foundational elements that you need for a data visualization.

[00:14:02] Richie Cotton: Absolutely. And the cool thing's incredibly important. Like I'm colorblind and I've had a lot of cases where I've been looking at a plot and people like, oh yeah look at the difference between those two. The red bar and the green bar. I'm like, those the same color. I, 

[00:14:17] Christina Stathopoulos: my, yeah. My husband is colorblind too.

[00:14:19] So I teach about this in my classes about colorblindness. Have you heard about the TIS color palette? 

[00:14:25] Richie Cotton: Ti Yeah. Yeah. Yes. Ti Yeah. Yeah. Gordon, do you wanna explain what's Virtus? 

[00:14:29] Christina Stathopoulos: Ah, yeah. Yeah. 

[00:14:29] Tis is it's a color palette optimized for colorblindness. So I think it's an one thing that I mentioned, like in my courses, my trainings, is that if you find that you have, colorblind people on your team, or a client who's colorblind.

[00:14:43] Be careful because you might be presenting a plot where they can't really tell the difference between, certain categories. So I recommend using this bees color PLA palette, which is optimized for colorblindness. So I'm sure you, maybe you're, you must be, if you're colorblind and you work in the data field, then you know it for sure.

[00:15:01] Richie Cotton: I have come across it. I've definitely used that so I can actually see what's going on in in the blood charade. So yeah, very useful. But, i'm wondering whether I've been pronouncing it wrong all the time. You pronounce it, I dunno. Vi vedis. 

[00:15:10] Christina Stathopoulos: I don't know how, I don't know how to pronounce it to be honest.

[00:15:13] Sdi, 

[00:15:14] Richie Cotton: we're not sure. It's some Latin name. Make it up as you go along. Yeah. Alright we're talking about, how generative AI isn't good for creating data visualizations. One thing it does work now for is generating SQL at least most of the time. So do you wanna talk through, like how you go about doing number crunching differently when you are using AI assistance?

[00:15:35] Christina Stathopoulos: Yeah. So I think it's similar to what I mentioned for the AI agents earlier, for for typical data analysis, data science, it's still, you're still treating it like a sidekick to help you with those SQL queries, right? I think that it's very helpful, especially like when you know what you want to do and it can help you format that SQL query.

[00:15:54] Especially when you get into like more complex ones or you forgot, how a certain function works, it can help you generate that and create that for you. You still always need to validate it, of course, make sure that it's going along correctly and it's generating what you wanted. And so it's not like you're never like off the hook with it, but it does help you.

[00:16:12] Again, it's helping you just move along faster and helping you when you forget, the exact syntax for something or how to, for how to do some sort of formula and then. I was gonna add that another thing that I've found it really helpful for when it comes to SQL is like it's amazing at doing it the other way around from what I just explained.

[00:16:30] So to, to give you like an example, I use this tool called Sherlock, which is like a repository for SQL queries. It's very similar to something that we had at Google. I was at Google before and we used something called Plex, and it reminds me of this environment that we had of Plex, but it's called Sherlock and.

[00:16:49] So it helps, it's like a repository for saving SQL code and it has like these embedded AI features. So one of them that I found that's quite useful, especially for a team where everybody's generating and sharing their SQL code base is that when you write you write out your SQL code your query, and you save it.

[00:17:05] It will then automatically describe what it's doing in natural language and it will save that with the code. So if someone on your team accesses it, they can read through that description and understand what it's doing before they dive into the code itself. So give this just an English natural language explanation.

[00:17:25] Somebody even like a business user would be able to read it and be like, oh, that's what this code is doing. Or somebody on the team, they can read it really quickly with the manager and get a quick idea of what that code is doing before you actually, have to analyze the code itself. So I found that also really useful.

[00:17:42] Interesting. 

[00:17:42] Richie Cotton: Yeah, so going from, I've got some sequel to, I've got a description of it. There's so many use cases of this 'cause first of all, nobody likes documenting their own code. It's tediously boring and yeah, definitely better outsource to a machine if at least if it gets the answer right.

[00:17:55] But then also you've got a prompt then to reuse next time you wanna write some sequel that's a bit like that. 'Cause yeah, you've got the natural language explanation of what you want to do. Throw that into your Ative ai and then yeah, that's the prompt for. Next time. 

[00:18:08] Christina Stathopoulos: Yeah, that's a really good point.

[00:18:10] Exactly. By the way, is there anything else you would add to this? Are you, I know you're a data practitioner too, so are you using AI in any other way when it comes to SQL or database and your analytics workflow? 

[00:18:20] Richie Cotton: I say, yeah it took a while to change my approach. 'cause a lot of the time I was like, okay, I would start off writing code then if I made a mistake I wanted to ask AI to fix it.

[00:18:30] But now a lot of the times I'm like, I'm trying to force myself more to start. Prompting and giving the AI to generate code and seeing how well it works. I'm trying to, I'm trying to be AI first and go, okay, let's get the AI to do everything and see where it breaks rather than trying to write the code myself and then get the AI to fix it.

[00:18:46] So I've been trying to change my workflows. I'm getting old. It's harder to change workflow than it was 20 years ago. But yeah I am trying to use AI a lot more for data analysis and data science stuff. 

[00:18:56] Christina Stathopoulos: I would say that, sticking with those old ways has its benefits too.

[00:19:00] Because if you rely too much on it, you can forget how to do that. Like you said, you wanna, you wanna just start with it from the beginning, which is good too. But I like to go like back and forth because sometimes I like to challenge myself to do it so I don't forget. 'cause that's, I think that's one risk that a lot of us data practitioners are going to have.

[00:19:18] Is that you become like over reliant on it, that you're just, generating the natural language. You're not coding it from the beginning, and then what if you get like rusty? Then it gets harder to validate it. So I think that's something that like all of us are gonna have to figure out is like how much AI first is.

[00:19:33] Okay. Because if we start to forget that fundamental knowledge, we can't even write the code without that AI tool, then maybe we're becoming like, you don't wanna get too reliant on your assistant if you can't do your job. So I think, I don't know how we can manage this, but we'll have to find a better balance.

[00:19:47] So we don't know. Forget what we're doing. 

[00:19:49] Richie Cotton: Yeah. No that's a definite risk. Oh, I saw a study, it was on Polish doctors or something recently, so they were given AI assistance for. Doing diagnoses for a few months, and then they tested their performance afterwards without the ai. And basically everyone had been using the ai, their performance decreased, they got worse than they were at the start at diagnosing conditions.

[00:20:12] So yeah, there is a different problem where you become reliant on AI and you become unable. Function yourself. I guess it's probably like relatively straightforward. If you don't have AI for a while, you will relearn the skills a lot faster. But yeah it's a certain, certainly a worry that your skills will degrade over time.

[00:20:27] I think so. Alright. We've talked a lot about what happens on an individual level as a data analyst or data scientist. I know you also do a lot of strategic work. Let's talk about what happens at the sort of team level, the executive level. Do you think data strategy has to change? Now that we have AI assistance.

[00:20:43] Christina Stathopoulos: I think data, like data strategy, data teams have had to evolve very quickly with everything going on in the AI space. I think the main thing though, that changes is that like it raises the stakes for data strategy. So the data teams involved, it reshapes what they need to deliver. They might have a bit more pressure now.

[00:21:04] I think we've seen this big movement happening now for years is that the data and the AI teams, analytics teams, whatever it may be, they are moving away from being like a cost center. To now being like they need to generate and prove their value. Like you really need to show what are you doing for the business?

[00:21:20] How are you generating value for the business? And I think it's putting more pressure on data teams to now it's like they've always needed to generate like accurate data that's important for the business. And now it goes like beyond that. They need to make this data that's ready for. AI systems. So it needs to be like machine readable.

[00:21:37] It needs to be labeled, it needs to be model ready, whatever it may be. But they need to make that data now ready for whatever AI systems the company is using. And then I also think that like the data team scope. Is shifting a lot too. Depends on the company the exact organizations. But now they might need to have a wider scope where they need to manage like realtime data.

[00:21:58] They need to manage more unstructured data. Whereas before, it was very like structured database focused. But now you need to go beyond that. You might need to be preparing, more unstructured data that requires. All sorts of new skills, real time data, et cetera. So I think it's just been like this shift and more pressure on the data teams when it comes to everything happening in the AI space, 

[00:22:19] Richie Cotton: I feel like there's a consistent story with every team in every organization.

[00:22:23] Oh yeah. There, there's an increasing amount of pressure you've gotta take on more new tasks and you've gotta figure it out now and deliver more value. But yeah. So it, it's a problem there that's pretty universal but. Have you seen how like the organizational changes 'cause you've got a data strategy, you've got an AI strategy, you've got a business strategy.

[00:22:41] How do you keep these things aligned? 

[00:22:43] Christina Stathopoulos: Yeah, you gotta keep 'em aligned because they need to work. They gotta work like hand in hand. I think you mentioned three things, right? So you mentioned business strategy, a data strategy, and you have AI strategy, right? 

[00:22:53] Richie Cotton: I mean there's probably even more, but Yeah.

[00:22:56] Christina Stathopoulos: Then there's more strategy. Yeah. Bi strategy. We can like, keep going on and just adding on little pieces, but I think it's like they need to work hand in hand. And you also need to have very clear like the roles, responsibilities of each one. Who's responsible for each thing the teams, and understand what exactly they're responsible for.

[00:23:12] So I think the business strategy should be pretty clear because that hasn't really changed. So the business strategy is the why and the what. So you're defining the vision, the goals. Identifying the value creation opportunities, those sorts of things. Then you have the data strategy, which again, like the data strategy itself hasn't changed so much.

[00:23:31] That's still the foundations, that's the enablement, right? They're the ones responsible for the data infrastructure, the governance, the quality, the accessibility, those types of things. That's data strategy. And then you have on, you add on now like the AI strategy, so that's more of I guess the, how.

[00:23:50] So defining how you're going to use the data and the tools at your disposal to deliver those business outcomes, you need to link back to the business strategy that I talked about first, how are you gonna deliver on these goals, the vision of the business, using the data that you have on hand, the tools that you have, and you're gonna apply AI to all of it.

[00:24:10] So I think like the, they all connect together and they need to work hand in hand. And when it even comes to the roles, the people working there, there should be, people dedicated to each type of strategy. But you're also gonna have a lot of like overlap too, where people are responsible for multiple things.

[00:24:25] You'll have a lot of like data and AI professionals who kind of work on both sides because it's very hard to separate. The data from the ai. So you'll have a lot of these overlapping responsibilities, roles between the strategies too. 

[00:24:36] Richie Cotton: Yeah. That's interesting. And I like the idea actually business strategy is probably not changing that much.

[00:24:41] The way you make money, you still gonna make money that way. Yeah it's mostly the AI strategy that's new on there. But you talk about shifting roles. Do you wanna talk me through like how data roles are and responsibilities are changing then? You mentioned you've gotta find ways to get more value.

[00:24:57] How do you make this change? 

[00:24:58] Christina Stathopoulos: Yeah, I think like we've said, it's about like value creation. You're expected now, you're not just a cost center, like you're like off the hook. You need to really be able to prove how you can deliver for the business. I think this is tough because a lot of data people, myself included, I can get like very wrapped up in, I love like solving problems, but like in the code and the data, I get very like hands down in the project that I'm doing and you get very disconnected from the business itself.

[00:25:26] This is risky for you, like for your career because in the end, if you want to, if you want to grow in your career as a data or AI professional. No matter what, you still need to prove to the business, like what value you bring to the business. And now more than ever, we talked about it, like the pressure is growing.

[00:25:42] So now more than ever, you've got to be able to prove what value you are delivering. What exactly, the work that you're doing. How is this helping the business? I don't think this has actually changed because this has always been that way, but I think it's become even more like important, even more fundamental to data and AI professionals is.

[00:26:01] You don't wanna just get pushed into a back office if you wanna grow in your role, you need to be able to prove how you're delivering for the business. And I think a way to like work on this then is never forgetting what's the end goal? The end goal is not to write this nice SQL query that the rest of your team can understand or whatever.

[00:26:21] It's not this type of, that's not the end goal. The end goal is like understanding what is this gonna do for the business. How are you gonna use this and apply this in your actual business to generate more sales or save costs or whatever it may be. So I think it's always it's a, it's like a change of mindset, or at least an evol evolving mindset where you cannot forget, like what's the end goal?

[00:26:41] You've gotta be, you've gotta be focused on that. I don't know if that makes sense or if you could, if you would add on anything else to it. But I think that's like the biggest thing that I've seen and something that a lot of us struggle with. I think a lot of data practitioners struggle with it, myself included.

[00:26:55] Like I mentioned. 

[00:26:55] Richie Cotton: Yeah, absolutely. Definitely something I've fallen over in the past in my career, it was like, oh yeah, going for that perfect model, that perfect analysis, and then actually project deadlines come way past and no one gets anymore. You missed the opportunity to do something useful for your business.

[00:27:10] Yeah I agree that keeping the end goal in mind is very important. One thing we'll say though, sometimes it is just very satisfying to, write some good code, do some, do a good analysis, and. If you're purely focused on let's grind out an analysis as fast as you can to get a result, sometimes you don't leave it in good state for if you wanna reuse the components of that project later on.

[00:27:31] So there are definite trade off stream, like short-term goals and long-term goals that sometimes get forgotten about in the chaos of business. 

[00:27:38] Christina Stathopoulos: That's true. That's a very good point. Yeah. 

[00:27:40] Richie Cotton: Alright I know one other thing you are passionate about is trustworthy ai. Now, I've heard of responsible ai.

[00:27:46] So is trustworthy ai, is that the same thing or something different? 

[00:27:49] Christina Stathopoulos: No, I think they're different. So I would say that responsible AI is it's like an umbrella term. So it includes everything. It includes like ethical, safe, sustainable ai. And by doing all of that, you generate trustworthy ai.

[00:28:05] So I think that's you. You create the trustworthy AI by creating responsible ai. So responsible AI comes first, umbrella term, and then trustworthy AI would fall underneath that. 

[00:28:14] Richie Cotton: Okay. Alright. So subset of responsible ai. Cool. Yeah, talk me through the relationship between data and then trust in ai.

[00:28:21] Christina Stathopoulos: I think the link between data and trust and ai I mean it's fundamental. So the AI learns and it makes decisions from the data rights. And if you're working with bad data, whether it's like biased or it's, full of errors or it's incomplete data, for example then you're gonna generate unreliable AI and you're going to lose trust.

[00:28:42] So the trust in your ai, it all comes back to the data. Actually, most things that happen with ai, it comes back to the data, but the trust very fundamental. If you wanna be able to trust the ai, you better have good data. 

[00:28:53] Richie Cotton: Absolutely. I think with all the hyper around ai, people often forget that data is still incredibly important.

[00:28:58] It's central. If you want good ai, you gotta have good data to, let's talk about what can go wrong. I live a good disaster story yeah. What are the sort of common mistakes you make if you're trying to. Create trustworthy ai. 

[00:29:08] Christina Stathopoulos: There's a lot of mistakes that you can make. So I think, eh I'll give like a couple of examples.

[00:29:14] So one common mistake is like black box decisions. So you're not transparent about the rationale behind your decision making, or you simply cannot explain how the AI is arriving to its conclusions. So it's these black box algorithms, right? This is common. This is common all, all over the place, but if we give an example, like a bank, so let's say like a bank is using an algorithm to make decisions on who they're gonna give a loan to or a mortgage to, and then if you get denied and the bank can explain why exactly you got denied, then you're gonna create like this distrust and this frustration with your customers, right?

[00:29:53] So I think black box algorithms, black box ai, that's a really common mistake. Another thing is, this is one I love. I could go like down a rabbit hole into, but it's biased or unfair. Outcomes. So that's a really common mistake as well. To give an example, so there's this really big lawsuit that got opened quite recently in the US against, I'm not gonna say names, but against a well-known SaaS company, a software as a service company.

[00:30:18] And the lawsuit is alleging that they were using the, this AI driven process, AI driven, like applicant screening for people applying for a job there that it was discriminating against people over 40 years old. And it was preventing them from being hired. And 40 is not even, it's not even that old, but it was preventing people.

[00:30:37] Richie Cotton: I'm right there with you. 40 is not old. 

[00:30:40] Christina Stathopoulos: I don't even, I can't even, I can't even believe it. It shows that age, but okay. So supposedly people over 40 were not being hired because the AI was discriminating against them. It was biased. Ai. And then so that's one thing bias. Another one as well that I'm seeing a lot, especially like recently.

[00:30:59] Is over promising on AI capabilities. And you probably, you've probably seen this a lot too. So you you have companies that are like marketing that their AI is a hundred percent accurate, for example, which is practically impossible. Or you have or their marketing that like they have these fully autonomous systems.

[00:31:17] When the truth is actually very far from it, like they're not they're setting these very unrealistic expectations. So when people find out you're gonna, you're gonna have problems if you're over promising. Then if I could add one more mistake but this links to like what we were talking about before, it's just the mistake of having poor data quality.

[00:31:33] It always comes back to the data. So we always say, garbage in, garbage out. I think that remains true. It's always going to remain true if you train your AI or even like your basic analytics, your basic business inte. Systems. If you train it outdated, inaccurate data, then you're gonna, you're gonna produce bad results, unreliable results.

[00:31:53] So another mistake is just relying on bad quality data. 

[00:31:57] Richie Cotton: Okay. Many different things can go wrong. Then certainly I can see, okay. Yeah. Your model or your AI is not explainable. That's gonna give you problems if you, particularly if you're in a regulated industry. Big, no-no in, in finance, for sure.

[00:32:09] And then, yeah certainly data quality. It's a well known problem that yeah, you've got bad data go, that you're not gonna get anything good outta it. Presumably we've just scared people now about how none AI is gonna be trustworthy. You have any advice for people who are now terrified?

[00:32:20] Okay, how do you go about making sure that any work you do, any AI you are making yourself, that's gonna be trustworthy. 

[00:32:27] Christina Stathopoulos: I don't wanna, I don't wanna scare people. There, there's good and bad. We don't have time to get into it, but there's like the other positive side of responsible ai, right?

[00:32:34] Because I even teach courses on the subject of responsible ai, so I love it. There is a lot of focus on the negative side because there's a lot of risks and a lot of risks that people are not aware of and people are becoming more aware of, thankfully. But it's taken some time. There's the other side too, so I don't wanna be focused just on the negative, but there is a good side to responsible ai, like for example.

[00:32:53] I talk about in my classes the good side of responsible ai, like using it for very good purposes. So AI for good or data for good things like, for example, using AI to help people like with disabilities. Like the blind or deaf people, there are incredible applications of how they're using AI now to help, for example, deaf people, especially deaf children.

[00:33:16] I found out through my research that deaf children have a much higher much higher literacy rate so that they don't learn to read, and they're using AI to tackle this, that there's a problem with teaching them how to read because they can't hear. So the parents oftentimes don't know how to teach them to read and they fall behind.

[00:33:31] They're using AI to tackle this. There's other applications for blind. So there's these incredible applications of like good good responsible ai. So I know this is getting a little bit off track, but I did wanna highlight the good side. So you had asked about what can we do to make this trustworthy ai I think there's a lot of different things you can do, and it depends on the problem, because if we're talking about like bad data, or if we're talking about bias data or black box algorithms, like each thing has specific things you can do to tackle it. There's different, frameworks, there's different ways you can tackle it.

[00:34:00] Like when it comes to bias, for example, you have these fairness metrics. Metrics where you can literally measure the fairness of your systems and try to improve on that fairness with black box algorithms as well. There, there are there are techniques for explainable AI or X AI explainable AI to break down these black boxes and create glass boxes, is what I say.

[00:34:21] So it depends on what exactly you are trying to tackle. There are ways to go about it. And there's lots of work being done just because the responsible AI space I think is exploding. It's growing a lot because companies have started to realize like it's not just about jumping on onto the AI wave.

[00:34:40] 'cause if you jump too fast. You can create some serious risks, and if you do it more calmly and you do it in a responsible way, you can avoid those risks, especially in the long term. You might get ahead in the short term, but you can face some serious risks in the long term. And what I always say is that the company should tackle this proactively because if you tackle it proactively, you avoid what I call the trifecta of risk, which is reputational risk, regulatory risk, and operational risk.

[00:35:05] If you put responsible ai. Into the strategy from the start. You can avoid, like I said, the reputational risk, regulatory risk and operational risk. 

[00:35:15] Richie Cotton: Oh, man. I think we, we've gone through like the pep talk back to scaring people again, but, okay. Yeah. So talk me through do you wanna talk me through the three things in more detail?

[00:35:24] Reputational what was it? Regulatory and operational risk. Talk me through Yeah. What's the difference and. Obvious things. 

[00:35:29] Christina Stathopoulos: So reputational I think is pretty obvious. That's more like exactly the company that I mentioned earlier, the SaaS company that's going through a lawsuit because they are, discriminating against people over 40.

[00:35:40] I'm not saying the name here, but if you Google it, you can find what company it is. This is going to hurt their reputation. Because now they're getting called out for discriminatory practices and other, there's been other controversies where companies have been found to be like discriminating based on race or based on gender, and you don't want something like that next to your name as a company.

[00:35:59] That's horrible for the reputation of the company. So that's for reputational risk. I think that's pretty obvious. For regulatory risk as well. That's that comes back to, there's changing regulations now. It's not it's not like you can just do everything as you want blindly.

[00:36:13] There are certain regulations depending on the region the place in the world that you are. Europe, for example, is a bit more strict. It's stricter than other, the European union's stricter than other places, other regions in the world. But I think every region is starting to develop their own data and AI regulations.

[00:36:30] So that's the regulatory risk, like being able to be prepared and develop things in accordance with that. And then operational risk as well. Operational is just problems that go wrong operational wise and it can come back to responsible practices. So not having lag. Even like connecting back to bias.

[00:36:46] If you're marketing like a product that is supposed to be for the whole population, let's say, but you're biased against, males or biased against females, you're missing out on half of the population, right? And so it can change you operational wise that you are not able to deliver on what you're even trying to do from the start.

[00:37:05] So hopefully that, that makes sense between the three. I could talk and talk forever about responsible AI and the risks, but are good. There are good things too. There's a good side too. 

[00:37:14] Richie Cotton: Absolutely. And I think just being aware of all these risks is a good start towards being able to deal with those risks, to mitigate those risks.

[00:37:20] But that that hiring example you gave, it's interesting because AI hiring tools there's a lot of them about now. But yeah it's one of those things in the EUI act, I think all these sort of HR use cases, they're classified as high risk just 'cause there are these problems with with bias and discrimination.

[00:37:36] Yeah definitely something to cautiously approach rather than dive headlong into. Alright. Since you mentioned AI for good, I do love that. Like certainly it's cool, like you're doing some data science project to like optimize ad clicks or something, but like actually helping people with real problems.

[00:37:55] That's a really wonderful use case. Have you done any more working in space or do you know any more sort of good examples of AI for good? We need a cheery story to finish up, I think. 

[00:38:04] Christina Stathopoulos: Yeah. Yeah. Let's finish like on a, let's finish on a more positive note. So AI for Good. Yeah. I love to explore this in my course and responsible ai.

[00:38:13] So I already mentioned like the di the disabilities one, I think that I think there's so much possibility for AI in the healthcare space medicine, but particularly helping people with any sort of disability, experience the world, be a part of the world in a better way. I mentioned the deaf example.

[00:38:28] They're helping kids that are bored deaf because they have a very low literacy rate blind. There are a couple of products out now that are like glasses for the blind that help them, that help describe like their surroundings help them read text. Like for example, these are things that a lot of us, if you're not blind or you're not deaf, you don't think about.

[00:38:49] The problems that somebody with these disabilities would face on a regular day-to-day basis. So if you take somebody who's blind that goes to the supermarket, and maybe they have their guide dog with them, but what if they wanna buy, milk and they can't read the cover of the milk?

[00:39:03] They can't tell what, which one is whole milk. Which one is skim milk? You have these glasses now that can help read the text out loud to the blind person so they can understand what they're looking for. Even the the glasses have embedded into them, hands free. So they can call Be My Eyes, I think it is.

[00:39:19] There's a couple apps like Be My Eyes and other ones where you can Live, call somebody as a blind person to help you. Navigate a situation they can see with you. So that's where Deaf blind. Another really interesting example for AI for Good, I wish I could remember the name of it, but I don't remember right now.

[00:39:36] It's an AI company that is helping detect wildfires. So right now I, I live in the south of Spain and Spain is experiencing these massive wildfires. It's just killing it's burning acres and acres all over Spain right now. They're trying to get it under control, but there is this company that does like early detection, wildfires.

[00:39:59] Because they, I can't remember the exact statistics, but they said that just most wildfires, they don't get, they don't get found or detected early enough, so when somebody finally notices it, they call. Then you have to get people dispatched out there. You lose a lot of time, and in that time, the wildfire is spreading.

[00:40:15] So it's these systems that they have, I can't remember if it's through drones or maybe even like the towers as well, the electrical towers, but they're watching around and looking for early signs of a possible wildfire, like little smoke going up. They automatically report this to the local authorities and they're able to go out there and check it immediately and so they can stop it from spreading and getting out of control.

[00:40:37] I think that's another useful use case. 

[00:40:40] Richie Cotton: Oh all these examples are wonderful. Certainly I like the sort of the. The idea of using technology to help with your sort of sensory defects. I wear glasses. I've got my regular iPhones at the moment. Usually I wear hearing aids as well.

[00:40:52] But yeah technology to help with these sensory things. Brilliant. I also like the idea of wildfire detection. I love backpacking and yeah, I like to not. Burn to death in wilderness. That's a really good use case. It's nice when data and AI are solving these societal problems.

[00:41:05] Alright. Okay. Just to wrap up I always love to find new people to follow. Is there anyone whose work or research you are interested in at the moment? 

[00:41:13] Christina Stathopoulos: Yeah, there's lots of people, but I think one person who I've been following for quite a long time now, and I love her work, I recommend it for everybody to follow, is Dr.

[00:41:23] Joy Guam Weenie. So she's based out of MIT in the us. She works a lot in responsible ai. No, no surprise there, probably that I'm following somebody closely in responsible ai. But she's done a lot of work, especially on when it comes to like bias, algorithmic bias discrimination. She does a lot of championing or campaigning for more diversity in the teams that are working on these systems as well as the data that we're using to train these systems.

[00:41:52] So I highly recommend people follow her. She has a fairly new book out as well called Unmasking ai. So I recommend that all about her work. And I think her website is called poet of code, poet of code.com. So you can learn more about her, follow her on LinkedIn. That's, that would be like my number one recommendation for others to follow.

[00:42:10] Richie Cotton: That's wonderful. So Joyce actually previous data frame guest as well. So yeah, she came on the show talking about masking ai. Yes definitely recommend checking out her work. Wonderful. Alright thank you so much for your time, Christina. Great chat. 

[00:42:22] Christina Stathopoulos: Of course. Thank you for having me.