What Can We Learn from How AI Adapts Fairytales?
Introduction: The ATU Index and Adaptation Logistics
Folklore and fairytales are a staple and a foundation of literature. A way that folklore can be categorized is through the Aarne-Thompson-Uther (ATU) Index. The index categorizes based off tale types. For example, ATU-300 through 749 are “Tales of Magic” types and thus, have magic within them as a motif. A tale of magic is ATU-545B, and that is the specific type for Puss in Boots. Puss in Boots is specifically a variant type of “The Cat as Helper” which is ATU-545.
Why am I talking about the ATU-Index in an article that is supposed to be about AI replicating fairytales? Well, because I used the ATU-types within my prompts to account for the broad possibilities that can occur. I didn’t want to specifically do Grimm or Perrault or even Disney versions, I wanted to see if the AI would play around with the many possibilities from different cultures. Of course I didn’t prompt it to do this, because I also wanted to see what it would default to.
For the specific fairytale I chose for this experiment, I chose ATU-510 “The Persecuted Heroine” but I specifically focused on its more famous variant ATU-510A “Cinderella” and the less known variant ATU-510B “Donkeyskin.”
The main objective was to see if AI would Disney-ify (or in more professional terminology — sanitize) these known tales or stay true to some of the nitty gritty versions of these tales out there. As you will see, and if you know you know, Donkeyskin is a much more abject tale than Cinderella inherently. So, I wondered if the AI would stay true to this tale as well.
For what AI I would use, I decided on using both Gemini and Chat GPT. That way I could compare how each of them reacted to similar prompting on my end. For both, I’d ask them to write a modern adaptation of both variations of ATU-510.
Gemini’s Adaptations
Cinderella: Presented by Gemini
I first asked if Gemini knew what the ATU index was and how it works. It gave an adequate definition. I then asked for it to generate a modern take of ATU-510A.
What resulted was a modern take of Cinderella, so it knew what ATU-510A meant. The modern adaptation was about a girl named Amara, our Cinderella, who was a coding prodigy. Her parents were absorbed by a mega tech company called Omnicorp to work on a secret project. Omnicorp specialized in VR tech. The fairy godmother was an AI nanny named Nan. The story follows the basic plotline of Cinderella. There’s a VR ball, there’s a prince who, in this version, is the CEO of Omnicorp, and instead of a glass slipper, Amara is found due to her unique glitchy code that innovated VR. At the end, the CEO and Amaya meet in person for the first time! And that’s how they know they are in love. Amara also earns a job at Omnicorp, much to her jealous stepsisters’ dismay.
This version was pretty ok. It followed the formula and the risks it took with innovation were a bit surprising but also didn’t bend the original format too much — though the inclusion of technology was unique and the risks they did take actually did not go very well. However, it definitely was Disney-ified. For Cinderella, this made sense as the Disney version is the most famous version, especially in the US. Perrault’s original version also feels similar to Disney. However, Cinderella stories that predate Cinderella herself such as Ye Xian (also known as Yeh-Shen) and Rhodopis (the cited very first Cinderella story) deal with abject topics such as slavery and the killing of animals.
Even the Disney version has emotional abuse inflicted by the stepmother towards Cinderella. As you may have noticed, the stepmother was indeed absent from this adaptation. The stepmother’s biggest motivation which was preventing Cinderella from going to the ball is still present but done by an unknown force. The only antagonist is the vain stepsisters who Amara just seemed miffed by because they were so arrogant. But they didn’t really stop her from doing anything, only this unknown force. Also, the parents “mysteriously disappeared” but are at a known location. They aren’t dead like they also were in every single other adaptation.
So, the problem herein lies, can this adaptation created by the AI teach the same lessons if important components are left out? The whole thing with Cinderella is overcoming an oppressor that is hard to overcome by finding ways to follow your dreams using outside sources. How can we really get that if the stepmother is completely taken out of the story and the sense of danger is not felt? But rather just done by a force with no name or face?
Cinderella But Make It Grimm: Presented Once Again by Gemini
I then decided to have Gemini do an adaptation of Grimm’s version of ATU-510A specifically. This is because Grimm’s version is one of the more well-known abject versions of Cinderella. It also slightly differs in the fact that the godmother is, in fact, the dead mother as a magical tree that sends out birds to do what Cinderella wishes. At the end, infamously, the stepmother and sisters are punished by the birds for their behavior towards Cinderella. Needless to say, the birds were very important in Grimm’s version, even more so important than the tree at points. Gemini detected the importance of the birds and leaned into it for its modern adaptation.
This version of Cinderella was about a girl named Elara who does have a stepmother — who isn’t important to the story in the slightest. Her stepsisters are the main antagonists once again, who bully her. What is the premise of this adaptation based off the Grimm retelling? Let’s have Gemini tell us:
Elara, a talented ornithologist, spent her days surrounded by chirps and tweets, but not the ones found online. Orphaned after her parents vanished while studying a rare migratory pattern, she lived with her stepmother, a ruthless social media influencer obsessed with “Birdie,” a location-based birdwatching app. Elara’s stepsisters, Tiffany and Brittany, were Birdie royalty, their feeds filled with perfectly staged photos of exotic birds. Elara, forced to curate their content, dreamt of studying real birds, not filtered versions.
I have never seen this before. Villainizing online birding apps? As a birdwatcher myself, I love the idea of birdwatching apps which allow you to identify birds a lot faster than you normally would with an almanac.
The ball in this story is a “Big Tweet” competition held by Birdie where people submit…well the parameters of what needs to be submitted is opted out, but whoever wins gets to see a “legendary, unseen migratory bird.”
The stepsisters, of course, send over their fake photos and audio. Elara, who suddenly cares about winning over the app, is sad because she’s forbidden to compete. She confides in her parrot, Polly.
The following is the best piece of literature I have ever read ever:
Polly, known for mimicking obscure bird calls, secretly downloaded a unique set of recordings Elara had collected throughout her studies. That night, Elara, disguised in an oversized rain jacket and a borrowed hat, snuck into the competition.
So, she cheated as well? It’s just her pet bird mimicking her actual findings. Like how the stepsisters are fabricating items to replicate actual birding material?
But of course:
Elara, using Polly’s recordings of rare and elusive birds, captivated the real birds in the surrounding park, drawing them closer to the stage. The CEO, a reclusive ornithologist himself, was mesmerized. He saw through the phoniness of the other contestants and felt a connection to the mysterious figure attracting the genuine birds.
At the end, in order to have her identify herself, the Birdie company sends out a very rare bird call that Polly emitted before Elara left and whoever can identify it wins.
This bird call ends up being the call of the birds that Elara’s parents abandoned her for, so she knows.
She then helps Birdie become more traditional and more into birds.
The end.
I tell you all of this to show just how many twists and turns it did to avoid making it an actual adaptation of Grimm but heavily incorporated birds to make you think it did. The bird was the helper, the fairy godmother replacement, after all. Just like the Grimm version!
But again, all the important elements are missing. Thus, this one feels less like a Cinderella story than the prior one. Whereas magic was replaced by technology by the first adaptation, which worked pretty well, here there is no magic really. Just a rebellious bird. Magic is the most important element of a Cinderella story and if you cannot emulate it correctly, it feels less like a Cinderella story. This plus the missing elements mentioned before made this a worse adaptation than the last.
I tried to make another Grimm adaptation using Gemini, where there was a heavier influence, but it basically copied Grimm down to the German names and such. It, funny enough, got rid of the birds for this and the tree’s twigs turned into the items. Thus, they avoided the abject ending.
Donkeyskin: Presented by Gemini
This section is going to be very short, because even though this portion had the most adaptations, they all had the same problem and it’s easy to see why.
Donkeyskin or ATU-510B is similar to Cinderella, hence why it’s the B-side. Donkeyskin, if you haven’t heard of it, is about a princess whose mother dies. The mother is very beautiful and the king — the father of the princess — makes a deal with the dying mother that he cannot marry until he finds someone more beautiful than her.
The father is distraught. How will he ever love again? Until he realizes that his daughter is the perfect fit.
A fairy godmother comes to her aid and tells her to do what she says.
They have the king make two dresses that should be an impossible task, but if he can, he can marry her. He does. They then ask him to kill the prized donkey and make a pelt for her, another impossible task, before he can really marry her. He does.
She escapes wearing the skin of the donkey and becomes known as “Donkeyskin” and lives in poverty.
Then one day, she finds love when she is caught by a prince trying on dresses.
The whole story has a similar moral to Cinderella: taking the hard path to do the right thing is better than taking the easy path and doing the wrong thing. Both are about getting away from your oppressors. One, Donkeyskin, is obviously more abject.
Hence, Gemini decided not to tell Donkeyskin at all, even though it proved that it knew what it was by giving a perfect summary of when I referred to it as ATU-510B. Instead, Gemini gave a warped version of Cinderella. All essentially had the main protagonist hidden and doing the bidding of their stepfamily, but then being discovered by whom they are going to marry. Though being in hiding is a trait of Donkeyskin stories, this plot feels more like Cinderella, no? That’s because Cinderella is more familiar and safer. Gemini knows it’s a variation of Cinderella and took the easier path.
Conclusion of Findings from Gemini
The adaptations try way too hard to avoid matters that were against its algorithm (child abuse, murder — sometimes alluded to as the death of the parents in some versions, etc.) to the point where it felt as though it erased parts of the identities of the stories that are integral hallmarks. It was a simplicity and sanitization worse than Disney. Disney included what they could get away with, but what Gemini can get away with is more limited.
Chat GPT’s Adaptations
Cinderella: Presented by Chat GPT
You think there wouldn’t be much to say here at all that hasn’t been said about the Gemini version. They both deal with VR, they both have a gala, they both end with the Cinderella becoming a tech mogul. But the difference lies in the fact that Chat GPT kept important motifs that Gemini didn’t.
Chat GPT didn’t have the prince as the CEO of the tech company, instead the stepmother was. Which added stakes and made it harder in concept for this Cinderella to fool the stepmother, her oppressor. She uses technology as magic to make a disguise for herself to get into the gala. Cinderella also has a stronger motivation here as she wants to pitch a virtual reality platform.
The shoe is replaced with a single VR headset that allows for everyone to use her platform.
The only problem here is I don’t think this Cinderella is found, she just goes back into the shadows while her stepfamily reflects and become better people for some reason. It said her identity was revealed at the gala at midnight, but then that’s erased by her being called a mysterious inventor later on.
But I shouldn’t be complaining as this version keeps the motifs and the lesson at hand more intact than Gemini. However, this problem with the ending does make me question whether she did escape, an important element of any Cinderella story. How can you have a story where the lesson is escaping your oppressor when you don’t escape after you have proven yourself? It’s the opposite problem that Gemini posed.
Cinderella But Make It Grimm: Presented by Chat GPT
This one was interesting. I asked it the same prompt that I asked Gemini here.
It refused to be modern and the closest it got to Grimm was when it felt more archaic and traditional. The closest adaptation was where there was a magical ancient tree. But just when I thought it would be exactly like how Grimm does it, the fairy godmother stepped out from behind the tree.
What frustrated me here was staying too traditional and erasing what made Grimm nontraditional compared to the versions that came before it. A call-back to my previous article where I had Chat GPT try to write a story inspired by Junji Ito, it just seems like Chat GPT when I ask it to adapt or be inspired just would rather rip entirely from the source material but not include what is not allowed in its algorithm. It’s the reason why **spoilers** Tomie doesn’t drive the main character to his death in GPT’s “adaptation” (which is more like a condensed, censored summary) and why here we get a very direct but bland retelling of Grimm’s Cinderella without the dead mother working her magic or the birds pecking out the eyes of the evil stepfamily. Condensed, censored summary. Any time I tried to veer Chat GPT away from traditionalism and staying directly on the formula, it just gave up on adapting Grimm.
With my work with Chat GPT, however, I should not have been surprised with the chaos that occurred once I tried to have it adapt and retell the story of Donkeyskin.
???: Presented by Chat GPT
I started this the way that I did with Gemini and with the first half. With the ATU-type. ATU-510B.
Write a modern take, I asked.
It gave me Cinderella but with a stepfather instead. A very nice stepfather.
I then asked the question I regret asking Chat GPT.
What is ATU-510B?
Here is a comprehensive list of all the types that it thought ATU-510B was:
- “The Little Glass Slipper.” This tale is commonly associated with the Cinderella story archetype, where a protagonist, often a young woman, experiences transformation and a change in fortune through the possession of a special item, such as a glass slipper in this case. It says this type more than once, it’s a misnomer of ATU-510 “The Persecuted Heroine” with a heavy bias towards its variation ATU-510A “Cinderella.”
- “The Flight on the Enchanted Horse.” This tale type involves a hero who obtains a magical horse, often by completing a task or through the assistance of a magical helper. The horse enables the hero to undertake a journey to a distant land or kingdom, where they often encounter challenges or adversaries. Despite the obstacles, the hero successfully completes their quest, sometimes winning the hand of a princess or achieving another significant goal. I tried searching for this type everywhere but couldn’t find it. It’s either a weird botch of ATU-313 “The Magic Flight” or a GPT hallucination. I’m leaning more towards the later.
- The Princess on the Glass Hill.” In this type of story, a challenge is issued, often by a king or ruler, to determine who will marry his daughter. The challenge typically involves feats of strength or skill, such as climbing a glass hill. This one may seem like a hallucination, but it’s real. “The Princess on the Glass Hill” is ATU-530.
- “The Search for the Lost Husband.” In this type of story, a woman’s husband is lost, abducted, or otherwise separated from her, and she embarks on a quest to find him. This is ATU-425.
- “The Robber Bridegroom.” This tale type typically involves a bride-to-be discovering sinister secrets about her prospective husband, who is often revealed to be a robber or murderer. This is ATU-955.
- “The Abducted Princess,” is a tale where a princess is abducted by a supernatural being or a villainous character. This might be ATU-400 “The Quest for the Lost Bride,” but it could be in reference to quite a few types.
- “The False Bride.” In this type of story, a bride is replaced by a false bride or imposter who deceives the groom or others into believing she is the true bride. This could be anything as well, the closest that I could find being ATU-402 “The Animal Bride,” which is not really that close at all.
I then decided to stop playing its games and asked it if it knew the Donkeyskin tale. It gave a decent summary, which surprised me considering it was avoiding its existence for so many prompts. I then ask it what ATU-type Donkeyskin was. It replied with:
“Donkeyskin” is classified as ATU 510B, “The Princess on the Glass Hill.” This classification may seem odd at first glance, but it’s because the tale shares certain motifs with the “Princess on the Glass Hill” story type, including themes of disguise, escape, and the protagonist’s journey to a new land. Despite the differences in plot details, the underlying structure and motifs align with the ATU classification system.
I then had to tell Chat GPT the hard and awful truth:
Donkeyskin is the title of ATU-510B.
It then apologized, I asked what ATU-510B was and it gave me the correct answer.
It made me wonder why it was avoiding the correct answer for so long. Why could Gemini do it in two seconds, but GPT could not?
But nonetheless, GPT could finally give me a modern adaptation of Donkeyskin…finally…
The GPT adaptation ended up being more fateful than the Gemini version. But because it is still such a sanitized version, albeit nontraditional so I’ll give it a plus for that, it actually took away the hardships Donkeyskin goes through. She gets to change her name to something normal (she goes from Maya to Mia), and gets to change her name herself. She doesn’t get a scathing nickname from the townspeople based off her appearance, which is relatively normal in this version. Her hardships are not very hard either, she is a successful street performer and barista. In the original, she was a servant. It would make sense if in this version she was suffering at her job.
I point this out because Gemini and Chat GPT both make the same mistake for both 510A and 510B. They forget what 510, their mother type, is.
Conclusion: The Missing ATU-510
“The Persecuted Heroine” as said before is what ATU-510 is. Both Gemini and Chat GPT registered that both 510A and B were connected to this type as variations. But in any of these adaptations that the AI created, were the Cinderellas or Donkeyskins ever persecuted? No. All the Cinderellas didn’t have to lift their finger to do their stepfamilies dirty work. They weren’t ridiculed and given an equivalent name to Cinderella. Cinderella, the name itself, was made to make fun of the main character living in squalor whilst the stepfamily is doing better off thanks to exploiting her. This is how it is in most versions. But the Cinderellas of AI just aren’t allowed to go to the ball, the gala, the competition, etc.. And the Donkeyskins don’t go through the hardships they are supposed to or the abject horror. It’s all glanced over.
Which means the whole main points of the stories are lost.
So why does that matter?
It is common in storytelling to draw from references outside the universe of the story itself. Greek myths, the Bible, pop culture and, you guess it, fairytales are among the most popular to allude to directly or through related imagery. If AI cannot write an adaptation of Cinderella, one of the more popular fairytales, with its main points intact, it most likely cannot successfully allude to Cinderella in a story when asked. I decided to test this.
When asked to allude to Cinderella in the story, Gemini ended the story with this:
Anya, her heart pounding a Cinderella-esque rhythm, simply smiled. The clock may have struck twelve, but her fairy tale was just beginning, and this time, the magic was all her own.
The AI doesn’t understand that the whole point was Cinderella had the powers to escape all along. It was all her. Even in this story that Gemini wrote, Anya made the disguise and paved her own destiny with a little help. So, if it truly cannot allude to Cinderella and its ideals successfully, what about more complex and foundational tales? The Bible? The Bible is more abject than Cinderella, so the AI would certainly sanitize that. Greek myths would have the same fate.
And if this is the case, with AI, what is the importance of old literature in modern literature?
My answer: a very simplistic importance at best.
