On The Digital Life this week we explore storytelling, creativity, and artificial intelligence. Our cultural evolution is reflected in our ability to communicate through stories, creating shared experiences and meaning. Recent research from the University of Vermont and the University of Adelaide used an AI to classify the emotional arcs for 1,327 stories from Project Gutenberg’s fiction collection, identifying six types of narratives. Could these reverse-engineered storytelling components be used to build automated software tools for authors, or even to train machines to generate original works? Online streaming service Netflix already uses data generated from users’ movie and television preferences to help choose its next shows. What might happen when computers not only pick the shows, but also write the scripts for them?
The Six Main Arcs in Storytelling, as Identified by an A.I.
The strange world of computer-generated novels
A Japanese AI program just wrote a short novel, and it almost won a literary prize
Jon: Welcome to episode 203 of The Digital Life, a show about our insights into the future of design and technology. I’m your host Jon Follett, and with me is founder and co-host Dirk Knemeyer.
Dirk: Greetings, listeners.
Jon: For our podcast topic this week, we’re going to explore the ideas of storytelling and the algorithms behind it, its formulas. As creativity and artificial intelligence draw closer and closer to human outputs, this is an interesting question for us to explore.
What kicked off this idea for our discussion this week was some research that came from the University of Vermont and the University of Adelaide, in which they collected the story arcs from various works of fiction that were archived on Project Gutenberg. For folks unfamiliar with Project Gutenberg, it’s an open source repository of some of the great fiction works of the western world, and these researchers from the two universities took texts- I think there were about 1300 texts- and ran it through a series of analysis, and sort of derived from these texts the emotional arcs that make up stories in western literature. As you might expect, there aren’t a huge variety of story arcs. However, these can be combined and recombined into much more complex plots. While any particular piece of literature might have a couple of these building blocks, the core pieces are what this algorithm was identifying.
According to this research, there are six main building blocks here, and those are, number one, as many of us might be familiar with, is rags to riches. We are familiar with these. Charles Dickens is a huge … Has a huge body of work that a lot of the stories are rags to riches stories, and of course that maps well to sort of the American dream, right? You rise up from nothing, and you become something very important, so rags to riches. Of course the opposite of that is riches to rags. That’s where you start with a lot and you end up with nothing, which is somewhat less appealing I would think.
The third plot type, or emotional arc that they articulate, is- I like this one- man in a hole. That reminds me of Alice in Chains for some reason. Now I’m dating myself with my gen X music references. Man in a hole is where someone gets into a lot of trouble and then gets out of trouble again. I don’t know if that would map exactly to … Dirk, I don’t know if you’ve seen the movie Catch Me if You Can, about the counterfeiter, the fraudster.
Dirk: Frank Abagnale.
Dirk: Yes, a long time.
Jon: Right. That seems like a man in the hole type of story.
Dirk: Wouldn’t it be “man out of the hole”? Like, rags to riches is when you start here, you end there. Shouldn’t it be you start in the hole and you get out?
Jon: Yes. Yes. It’s a fall and then a rise again. Yeah, that’s sort of the core to that plot. Then the last three all have … They’re plots, but they’re all associated with a specific work. Icarus. You know, we all know the story of Icarus flying too close to the sun. Cinderella, which has a rise, then a fall, then a rise again, and the Oedipus, which is a fall, then rise, then fall. Those are the six core story arcs, and if you’re looking at a more complicated, longer work, like, say, some of the Harry Potter novels, those will have multiple components sort of integrated together to get these arcs sort of placed together. Not just like Lego bricks, but you kind of get the idea.
You can see, I mean, we’ve already started talking about how our culture in the United States is reflected through some of these stories, but now we’ve got artificial intelligence, which is more or less boiling down these texts to their essence. I think what’s interesting about this is there have been attempts to then take these pieces and then have the computer then reassemble those and create some new work out of it, right? You’ve boiled down, you’ve distilled the essence of these plot arcs. Now, can the AI sort of create from the components that it’s derived in the first place? From what I’ve seen, the resounding answer so far at least is, “No. No thank you.” AI is good at finding the patterns, but not so good at mimicking the outputs. At least for the time being, human authors, you might be safe so far.
Let’s just take that first part. Dirk, what was your reaction to this research? It feels a little bit new, but it’s kind of something we already knew, that there were these core plots here.
Dirk: Yeah. Yeah. I think the article mentioned that Kurt Vonnegut was initially doing some of those sort of breakdowns of different story arcs, you know, decades ago. The romance that we ascribe to humanity is largely just ignorance. What I mean by that is, things like writing a book, things like all the detail that we put into the description of a place or a character, those are the result of thought processes that are, at the basic level, as stripped down and brute force as what AI can be programmed to do. Aside from the fact that it’s a machine created by a person, as opposed to a machine that’s doing the work, I think it’s something of a false premise that the AI doing it, the machine doing it, is something that’s so different and foreign, and lacks the humanity. I think it’s getting down to the essence of the humanity. The humanity without the ignorance layered on top of romanticizing things simply because we can’t articulate the ways in which they actually happen.
You mentioned, well, writers, and I’m misquoting you now, but you said, “Well, writers maybe don’t have much to worry about yet from AI.” I read another article this weekend about a writing competition in Japan, and for some number of years now they’ve invited non-human writers to participate. The implication from this article was this year was the first year that AI writers participated, and they’re welcome to do so, and it’s done anonymously, by the way, so that the people who are judging don’t know if someone’s a human or not, and it was under 1% of the submissions. It was roughly 1500 submissions and maybe 11 or 12 AI submissions to this contest, but at least one, and maybe some of them, got into the later rounds of the contest. Now, none of them won the big prizes. Humans won the big prizes, but the AIs were skillful enough to get through early rounds of a competition for the best writing in Japan.
Jon: Yeah. I saw that article as well, and I think the AI, it was a recombinant competition, so I don’t think the AI actually constructed sentences. I think it recombined sentences that were written by humans, which nonetheless, you have to, I mean …
Dirk: You have to start somewhere, right?
Jon: You have to start … Right. It doesn’t negate the fact that it made it to another round. I mean, you could take any two sentences. Any bad editor can mangle a piece. It’s probably a little bit more accurate to say that the AI edited or at least to that level of composition, which is impressive nonetheless. I do think you’ve articulated well sort of the point of having AI create from a rules-based set, create these pieces of fiction or what have you. I think there is a part of human input that is not rules-based, that is fundamentally, call it improvisational or chaotic, or does not necessarily follow a nice formula.
Dirk: Totally don’t agree. Make that case in a concrete way.
Dirk: Give me a concrete example.
Jon: For instance, if you’re a fan of jazz, which I am, there are elements of jazz which you can sort of create the elements that all players need to know to play together. There are certain rules, right, that allow a jazz group to play together. It’s not everybody improvising at the same time usually, but then you have someone like Ornette Coleman, who says, “Okay, what happens if I remove some of those rules and just see what happens when everybody is improvising at the same time?” Or you have somebody like Thelonious Monk, who is playing piano in such a way that no trained pianist, and very few untrained pianists would even think to do, because he’s looking at the instrument from a completely … I’m using “naïve” here, but not in a negative way. Like a perspective where he’s coming to the instrument without any of the rule sets associated with the usual … When you’re going to learn an instrument for the first time. He’s very much a self-trained person.
I think there are, when we’re talking about breaking down rules for machines to understand, there’s a human perspective that then sort of becomes something special I think, when you approach that same problem set, like a Thelonious Monk or Ornette Coleman, where the rule set is just sort of thrown out the window. That’s what I mean by that.
Dirk: Oh, Jon. That’s so quaint. Let me brutally break it down into how a machine-like process would deal with that. You have M songs out in the world, within which there are Y rules that are exhibited. Those songs could be broken down into constituent parts to illustrate the rules, can be broken down into constituent parts to illustrate what … I’m not a musician, so I’m now going to use terms that undermine my ability to make my point, but how this chord, it goes well with that chord, or this set of things goes well with that set of things. That’s all the easy stuff and the expected stuff. What the sort of jazz geniuses you were talking about do is take the unexpected things and make them work, and the way that they do that is by having an understanding of which unexpected things would work in the context of something else.
Now, they’re not doing that out of alchemy or magic. They’re doing that out of an understanding of the data, an understanding of all of these data sets, all of these combinatorials, all of these rules that they have internalized, and they’re plucking bits without consciously thinking about it of things that do work. But at the end of the day, it’s the same thing. It’s all data, and how it’s being used, and having a sense, in sort of the touchy-feely human way of when and how and why it should be used. The only barrier to AI being able to do that is having enough data that is structured correctly, that is tagged correctly, for lack of a better word, that the AI then is sophisticated enough in its construction to be able to use. I don’t have a sense of time horizons, but I think it’s years or decades, not centuries, before the best improvisational jazz music is being done by an AI.
Jon: Yeah. I’ll take that under advisement. I don’t agree, just because I’m sort of in the weeds here in terms of my sort of being soaked in this music. Is anything possible? Sure. Sure, it’s possible that the AI could have some huge total data set that makes it sort of apply to become Thelonious Monk, or whatever. Pick your avant garde jazz great. But I think there is something special about what the human being can do, because there’s also this … Who, then … I guess my followup question is, if AI Thelonious Monk is just jamming out thousands and thousands of jazz pieces, and most of them don’t work, where does that get vetted? I mean, there’s a lot of like sort of judgment and being able to get this music to resonate with people as well. Do you see that just as another data crunching problem, or is it like the AI is going to be able to sense the culture, and the environment, and the people who are listening to it, and all that, and that’s just a big massive data problem?
Dirk: Yeah. I think it’s a maturity model problem, where in early days, I mean, similar to … You mentioned that the book, the writing competition was just combinatorials of sentences. “Just” I’ll put in scare quotes. In the early days, there’s going to be a lot more bad than good. There’s going to be a lot more that’s not sophisticated, or doesn’t sound good, or doesn’t work out well, and there will be a lot of content generated. High noise, very little signal. As time goes on, that’s going to flip, and it’s going to get to the point where there’s not going to be need for any human or AI sort of editorial layer. The AI itself will be producing consistently good music that can be released in the same way as something that previously had gone through a long human vetting process that was a lot more expensive and time-consuming, and not necessarily offering an art, a music that was any better.
Jon: I think we’ll leave the music discussion there, because I know for a fact I can ramble on on this for a long time. I want to get back to the initial spark for that discussion. Of course, the tools around literature. I wanted to pick your brain about how … We have companies like Netflix, which are inherently mapping sort of content, much in the same way that the researchers did with the Project Gutenberg content. In the case of Netflix, they’re mapping this content to real users, and I just had this “aha” moment that you kind of wonder how far Netflix can take some of their algorithms when we begin to understand that, “Hey, there’s only so many stories that can be created.” I wonder where Netflix is going with their sort of next-level understanding of what human beings like in terms of stories, and what is going to be sellable, right? Be part of the culture as human beings consume more and more stories and movies.
Dirk, I know you’re a big consumer of Netflix. How is Netflix doing with that so far, and do you think their AI is likely to develop further in this vein successfully?
Dirk: Yeah. I mean, if we’re talking about Netflix, that starts to blend a couple of things together. I mean, I think their AI is doing okay. It’s tricky, because there’s four people in our family who are all on the same account. Even though we have our own logins, ostensibly there’s some pollution. My stream has been polluted by a few different people’s tastes, so I’m not going to impugn Netflix on that. What I will impugn Netflix on is choice, and the content that they’re offering has diminished. Now, I mean, we’re going to jump the tracks from AI and into a different topic, but I think it’s worth mentioning, you know, the cord cutting has been a big trend over the last decade. Less earlier, more recently. What I’ve noticed is that the services that enable cord cutting, you know, Netflix most primarily, but Amazon Prime as well as Hulu and others, the content is getting worse. It’s getting less. It’s now difficult when we want to watch something to flip around and find something that’s new and interesting. It’s sort of the same old stuff, or it’s a lot of bad stuff.
That sort of degradation in available content is something that I’ve noticed for some time, you know? One example of it, Hulu used to host the Criterion Collection, which is a collection of … I don’t know how many. I think it’s hundreds of sort of the great classic movies. Criterion took their ball and created their own thing now, so if you want to get into Criterion stuff, you have to pay another, you know, $10, $12, $15 a month. If you want to get into CBS shows, that’s $12, $13, $15 a month. Some of their stuff they have on Hulu. A lot of it they don’t. The bottom line is, now you have all of these different channels. They all cost money, and they’re not coordinated, and the content that any one of them shows is generally really limited and crappy. I mean, it makes me wish of the days of 500 channel cable TV, even though I only would use 10 to 12 channels. I was paying one bill, and I knew where I could always go for consistent content. Now it’s jumping around all of the different shards of content and having a hard time finding what we’re looking for.
I’m sorry to go away from Netflix and AI. I’m not currently qualified to talk about it, thanks to my family’s viewing habits, but I think Netflix and all of these other streaming services are doing an increasingly poor job of making available the right content, or at least surfacing it in a way that I can find it, because I try pretty hard, and I’m not finding a lot that’s new and interesting for me.
Jon: Yeah. I think that might be relevant in a way to the storytelling discussion that we’re having, because in … I feel like in theory, as Netflix gains this information about what you like, there should be an avenue for them to deliver content, whether it’s … You spoke about how a lot of the licensing deals have changed over time, and they’re not sort of available in the same way anymore, and that’s very understandable, but you would think that the content that is not necessarily mainstream but is still sort of mapping well to your tastes would improve over time if there are indeed sort of certain story types, certain genres, certain actors, whatever it is that appeals most to you, Dirk. I would think that over time, Netflix would be able to make up some of that ground with you. Whether they can or not remains to be seen, but it’s an interesting problem set, because certainly the licensing is an additional barrier or hurdle or them to have to get over as they try to find the right stories for Dirk, right?
Dirk: Yeah. Yeah, I’m sure I’m not alone in this. You know, it’s the typical … There’s certainly good parts to a free market and capitalism, but this is getting into some of the bad parts. Namely when this was new space, when it was sort of Netflix and nobody, or Netflix and one other company mucking around, there was an abundance of content. There was plenty of content. Now, you have a number of big players. You have all these smaller people seeing a cash grab and trying to get in on it, and the competition that’s created by the sort of abundance of possible revenues is fracturing the content and making it a lot more difficult to enjoy it, which is too bad.
Jon: Yeah. That is sort of too bad, and also probably a reason why they’re investing more and more into their originals, so they don’t get cut off from content on the backend. The final piece to this discussion, I am very interested in sort of this combination of human and AI to generate news stories. We talked about the computer generating stories, and we’ve also talked about the computer identifying the right stories to deliver to you. In this last segment, I wanted to explore the idea that human writers could use AI to enhance their output, right? We all sort of work faster because maybe we use Google Docs, or in the past used Microsoft Word, made things easier, right?
I could see a step in the future where the software, you know, you’re a fiction writer, and you’re creating your architecture for your story, and this piece of software can provide for you certain elements of story that you’re like, “Okay, I’m doing some science fiction. I’m going to do, you know, the rags to riches kind of story. I’m going to use dialogue in this style. I’m going to lay out the parameters, and I’m going to do the plotting and create the characters, but I wonder if there isn’t potential, especially in an era where there’s so many opportunities for content now. You just pointed out like five or six different services that all need content now. I wonder if there’s this human plus AI storytelling software that’s going to come along, because I sure would love to experiment with something like that.
Dirk: Yeah. Yeah. I mean, there’s so much that could be done, right? The problem is that it appears that corporate interests are pursuing AI as human replacement as opposed to AI as human enhancement, right? When you said, “Hey, there could be this cool new tool.” My brain started going in different directions of, like, “What could there be?” And what would be really neat and easy to do with the kind of data they have at their disposal is, if I’m a fiction writer, I’m writing, and I don’t know. I’m not going to even put a line out there, because it would be horrible coming off the top of my head, but I’m writing about somebody smoking a cigar, for example. It could pop up a little thing and say “words commonly associated with cigars,” that are taken from just some curated group of like the best 100,000 books ever written, or whatever that looks like. I don’t know what the right chalk lines to draw are, but then it could give me a list of 10 words. It could show me sort of how frequently they’re used.
The one with the highest rating might be “unctuous,” for example, and I might say, “Oh, ‘unctuous.’ It’s a little too distinctive. It’s so popular. Probably I don’t want another unctuous cigar.” So I start to tick further down the list, and I see, “Oh, interesting. This word … I don’t remember seeing that, and I like it. It kind of fits my style. Let me go with that.” Like, that would be fantastic, right? Because it’s doing a couple of things. It’s sort of putting the dictionary-thesaurus right at our fingertips. It’s doing it in context, but then it’s also giving metadata around frequency, and, “How accustomed are readers to seeing that word? To seeing that thing in that way that does work?” That software really could be done with today’s technology. Like, that’s not some futuristic AI thing, but it’s … It’s not a problem that’s interesting to people with money and will to do things around AI, because there’s not a lot of money to be made there. The money to be made is, “Kick Dirk out of the writing process, and have the damn computer write about unctuous cigars until the cows come home.”
Jon: Right. Yeah, I’d love to see a list of, you know, “Here are the sort of top 10 cigars.” Or, “Here’s how cigar descriptions are handled based on sort of size of cigar, and the type of leaf that’s rolled.” Et cetera, et cetera.
Dirk: Yeah. Amen.
Jon: Yeah. There’s a startup idea for somebody, but unfortunately since it would be targeted at authors, you probably aren’t going to make any money.
Dirk: You’ll probably lose a lot of money, but a lot of people will be happy.
Jon: That’s right. Listeners, hope you enjoyed this episode on storytelling and AI, and if you did, please Tweet at us and let us know what we did right and what we can do better. For now, remember that while you’re listening to the show, you can follow along with the things that we’re mentioning here in realtime. Just head over to TheDigitaLife.com. That’s just one L in TheDigitaLife, and go to the page for the show. We’ve included links to pretty much everything mentioned by everybody, so it’s a rich information resource to take advantage of while you’re listening, or afterward if you’re trying to remember something that you liked.
You can find The Digital Life on iTunes, SoundCloud, Stitcher, PlayerFM, and Google Play. If you follow us outside of the show, you can follow me on Twitter @JonFollett. That’s J-O-N F-O-L-L-E-T-T. Of course the whole show is brought to you by Involution Studios, which you can check out at GoInvo.com. That’s G-O-I-N-V-O dot com. Dirk?
Dirk: You can follow me on Twitter @DKnemeyer. That’s at D-K-N-E-M-E-Y-E-R. I promise I don’t write about unctuous cigars, because that would be gross, but thank you so much for listening.
Jon: That’s it for episode 203 of The Digital Life. For Dirk Knemeyer, I’m Jon Follett, and we’ll see you next time.