Laura Gast - Uniquely Human: Data Storytelling in the Age of AI

Let's start with a brief introduction of who I am. So my name is Laura Gast and some of you may remember me from last year where I talked about cognitive science and information design, but I am someone who spends a lot of time thinking about how we think and learning about how we learn. So my background has been a little crazy. I started in machine learning on image analysis for NASA satellite images. I went through some political science, modeling for elections. I did some infectious disease epidemiology. I worked on WhatsApp chatbots with the WHO during COVID, and now I work on data strategy at a military family-related NGO. So it sounds very wide and different, but at the core of it is the idea that I need to communicate something very technical and specific to an audience that might not totally know 100% what I'm talking about, right?

So I want to talk a little bit about how we can communicate across this technical AI idea and get it to the humans who we want to actually make and take action, right? So let's get into this by saying, congratulations, you're amazing. You just created the most amazing AI-based workflow model analysis pipeline, what have you. Well done, congratulations. Maybe you figured out a way to sell widgets better, or you unlocked a secret about life, the universe, and everything, or maybe you went the other way on that. You do you. No matter what you did, you're going to have to walk into a room and sell it to your C-suite, right? Now you, the person who spent months, maybe years, working on this thing, your best thing ever, you want to walk into that room and just be like, hey guys, data speaks for itself. We got to do the thing. Out of curiosity, people in this room, how many of you have heard some version of the data speaks for itself, the model speaks for itself? Yeah, we've all some version of that, right?

I am a firm believer that the data can speak for itself when the data collects itself, cleans itself, and maintains itself. Until then, you are the one that has to speak for your data. So if we are tasked with getting our C-suite on board, fully understanding, trusting, believing us, how can we act as the best ambassador for our great thing, right?

I am a firm believer that the data can speak for itself when the data collects itself, cleans itself, and maintains itself. Until then, you are the one that has to speak for your data.

Modes of persuasion

So I want to talk a little bit about modes of persuasion, just briefly, all right? If you've ever taken a public speaking class, or philosophy 101, or speech writing, you may have heard of Aristotle's rhetoric. Sounds really fancy. It's the way philosophers like to, you know, sound fancy. Again, a technical thing that I need to communicate to some people who maybe don't do Aristotelian philosophy, whatever. So what you need to do to persuade is first, you're going to appear to logic. You're going to appear to logos, right? This is your data. This is your methods. One plus one equals two, you know, ipso facto QED conclusion, right? That's one thing you have to appeal to. You need to make sure that you've got logic under it. You also need to apply, appeal to ethos, right? To trust. Not only can you trust my model, but I'm also a credible speaker for this thing. You trust me and you trust my work. You're also going to need to appeal to pathos, or the emotion. And so it sounds funny, especially for us data people who want it, everything's logical. But you can think about the phrases, the heart of the matter, right? Or the end of your pitch, your new strategy, dear C-suite, we need to take this path either because look how exciting and hopeful we are, or look how bad this could be. Aren't you afraid of it being bad? We need to take this path, right?

So you can kind of think about that in the narrative that you create, right? Then the fourth part, which is maybe is not Aristotelian, but it's also Kairos, so the appeal to time. You can think about this in audience and timing. You may have the best argument and presentation and solution ever, but you're not going to pitch that at the meeting to say thank you to the departing CEO, right? Timing, things that are outside of your control. So AI and ML and deep learning are absolutely brilliantly powerful tools that you can use for pattern recognition, prediction analysis, everything else you've heard this week, right? The methods, you're the one that has to teach the computer what to do. You have to set your prompts, right? You have to work through it correctly. All of the rest of this is still pretty squarely the domain of the human, right? You are the person that has to bridge this brilliant logical thing and bring it to the humans.

So I'm not going to go through all of these things. There's no point to that. We only have like 17 minutes left or something. What I want to do is distill this down to three ideas. One, you are responsible for setting context because context is the thing that feeds your story. And then you need to make sure you're telling your story well so that the people you're telling it to can do something, right? You're going to have some impact. You want an action, right?

Context

So let's talk about each of these three things together. Okay. So the first thing, context. So I'm kind of describing this as background or framing or circumstances, right? So humans love giving things scores, right? 92% on an exam or five stars on Uber or I don't know, a KPI, an OKR, 10 out of 10, no notes, et cetera. So why do we score things? We score things because we enjoy ranking them and we rank everything. All of these are real, by the way. We rank things so we can determine what is good, what is better, and what is best because given the options, we want the best thing, right? We want the better thing if we can't get the best thing and we'll take good at the end of the day. So you'd think with all this scoring and ranking, we would understand a score.

So let's play a game. Is this a good score? Is it a good score? No clue. What happens if I start adding context and I tell you that this is a women's gymnastic score on the uneven bars? I see some knots. Some people have some subject matter expertise in this. What happens if I give you the components that went into my KPI? What if I give you a little indicator? What if I tell you this is the Olympics? What if I tell you this was the athlete and this was the face they made when they got that score, right? So all of the context that's on this screen, along with the context that the announcers are delivering to you while you're watching this, is setting the stage so you can understand that for some athletes, this might be a great score. But for this athlete, it's not a great score. And given the context of this competition, this is really not a great score, right?

But maybe we have too much subject matter expertise in the room on women's gymnastics and the Goat Miss Simone Biles. Let's look at things maybe we don't have subject matter expertise on. Are these good scores? 50.37 in the 400-meter hurdles, or 4.74 seconds in speed climbing, or 48 in trap shooting, or 694 in 72-arrow, 70-meter round archery. What about these? Are these good scores? Devoid of context, it's really hard to know. But if I add a little more context and tell you that each one of these was a world record at this year's Olympics, you go, oh yeah, those are good. But how do you compare across these? Are any of these better than good? Are these the best of the best, even though they're world records? Well, you might get more information if you know that that 400-meter hurdles world record holder, Sydney McLaughlin Leverone, has broken her own record seven times in a row over the last three years. Wow, that sounds like a really good score, right? Like, that's amazing. Or you think of speed climbing. We've only been collecting records since like the late 90s. But for archery, we've been collecting those for over 100 years. So breaking a record that's newer might contextually seem better, right?

This is also why data visualization is so important. So let me give you a data visualization. DataViz. And if I add a little context and a little annotation, you see that all but one of these swimmers is swimming in the opposite direction. And then I give you a little more context and I tell you this is another goat, Katie Ledecky, winning a race. And here you can see the scale of the win. You can see just how best that best thing is, right?

So this idea of context is sometimes easy and sometimes hard, and sometimes an input and sometimes an output. So when it comes to an input, sometimes these are easy to build in, right? Or sorry, yeah, it's easy to build in. It's easy to give you a score and say this is pretty average given this situation. And sometimes they're harder to build in. Like, how do you build in sarcasm and parody into an LLM? How do you teach a computer what sarcasm is, right? And then for outputs, and again, this is your responsibility, sometimes there's oh-duh results, right? You get something out and you need to contextualize it. Here we're looking at an LLM cloud, I think, misunderstanding the context of 9.11 versus 9.9. Numerically, 9.9 is larger, but for version controls, 9.11 is larger, right? So you're missing that extra component of context. And then what I have, I call the oh-no results, and I apologize ahead of time for this, but video AI, the major problem it has is context, right? It's predicting the pixel next to it in time and space. It has no ability, most of them for right now, have no ability to put in real-world context that a human should have two legs and a beam should not wibble in space, right? So this horror show, there are models that are working on this, but they still fail more often than they succeed. They're fancy, they're compute-intensive, but the idea of getting context in is still really difficult.

Story as access

So we're going to move away from this horror show, and we're going to move into the idea of story, which most people will talk about as the arc and the narrative, right? I want you to think about it differently. I want you to think about story as access, right? So what do I mean by story as access? So we're going to go back to another Olympic KPI and talk about Cole Hawker winning the men's 1,500-meter race in Paris with a time of 3.2765. It was good enough for an Olympic record and a gold medal. Cool. All right.

Now I got my data viz here of the finish line, and you can see Cole Hawker with his arms wide, excited, and then you can see the other American, Jared Degusse on the far left, leaning forward, working really hard. You can see Kerr and Ingebrigtsen just behind, looking a little dissuaded because they didn't win. And you go, okay, cool. That's neat, Laura. Thanks.

But what I just gave you, it's really hard to remember that this is the way sports broadcasting and sports journalism used to work until the 1960s, when this guy, Rune Ulrich, in 1960 reimagined the idea of sports broadcasting and sports journalism. He, in 1960, had a famous memo, we are going to add show business to sports. He brought out the wide world of sports where it introduced Americans to new and exciting new sports, and they got invested and started moving in, moving in and enjoying these sports. And then his protege, Dick Abersole, took charge in the late 90s, or sorry, 1989, when he switched to the priority of sports broadcasting being storytelling. Prior to that, it was one camera on a football game, you know, some sets. And he wanted to change it. He said, every game is an epic story to tell. The way to capture audiences for unfamiliar sports is to tell the stories of athletes competing, to pique their curiosity and give them a rooting interest. The audience is not passive. Your C-suite is not a passive recipient of the information you're trying to sell to them. You want them to be excited and rooting for whatever outcome you want them to be rooting, right?

So if we go back to my data viz here of this finale, I don't have time to tell you the whole story of this race. But if you know anything about the Drake-Kendrick-Lamar like rap battles, where they've been talking trash on each other for a very long time, they have nothing on Kerr and Ingebrigtsen. For years, they've been talking trash on each other about this race and about this distance. And if you watch that race, you see them fighting each other. The announcers are talking about this great story. And here comes Cole Hawker, ranked number six in the world, sneaking up behind him, taking that gold, getting the Olympic record. ESPN and many other places are calling this a showdown for the ages. This was an amazing story. Trust me, go watch it.

Ethical storytelling and the limits of AI

So the problem with storytelling is that fake stories are being sold to us as real stories, and as humans, we're not sure how to feel about that. Sometimes it feels okay. Sometimes we like the fake story better. Sometimes we like the computer story better. We think it's better than the human. We're very uncomfortable with it, in the best case scenario. In the worst case scenario, we're angry. There's backlash. Google had to pull their Gemini ad from the Olympics this year because everybody was so uncomfortable and upset with the idea that this father would use Gemini to write a letter from his daughter to his favorite athlete. And it was like, this is gross. Just write your own letter. It should come from the heart. Remember, talking about that emotion comes from the heart. So we care about it. They pulled this ad after like four days of the Olympics because the backlash was so bad. We've also got authors who are putting out books that maybe were aided by AI, not wholly by AI. But the backlash is so strong because as humans, we're like, but that's a computer. That's not meant for me. Why are you trying to get at my emotions when that's a computer? Right?

So this is a complex idea that basically there is danger in listening to a great storyteller in a world where great stories and things that are true aren't the same thing. Right? Remember, I had to say you had to appeal to that credibility. You had to appeal to that trust. You don't want to lose the trust of your audience. You want to be transparent of, I used this great chatbot or this great chat GPT to help me write this code. And then we move from there. Right? Hallucinations, deep fakes. You don't want to turn over the idea of telling this story of giving your audience access to your story to a computer. Right? Story is access. Access is making someone an active participant. It's too important.

Story is access. Access is making someone an active participant. It's too important.

Impact and the human in the loop

So moving on to the last thing, those two items together are the thing that create impact. And again, this is where it's uniquely human because you know what? Data is not information. Information is not knowledge and knowledge is not wisdom. That is to say that humans are in the loop because data doesn't act. A KPI doesn't act. Maybe we would put a trigger for something to happen, but the human decided that trigger. AI doesn't have intent. AI is still solving the problems we ask of it in very careful and iterative ways. It's not hopefully creating problems that then it's solving its own problems. Right? So, you know, the brilliant thing is that humans aren't and can't be cut out even in this age of wonderful, amazing generative AI and AI tools and all of these wonderful things. And I think that's pretty neat.

So, you know, my call to action at the end of this, my impact that I hope to have on you is for you to remember that you are the one speaking for your DAB. You are the one speaking for your model. You're speaking for your output. You are the one communicating this. So make sure that you set that context right because no one knows your work better than you. Make sure that you're telling that story. So you're giving people access to be excited or fearful or cautious or wrapped up in cheering you on for your new work. And then make sure that you're telling people what impact you want at the end of the day. And remember that the humans are the ones, the humans in your audience are the ones that are going to create that impact. So thank you very much. This is where you can find me. The blog is coming back to life. So thank you very much.