I won’t be publishing any more stories with storiesby.ai. What was once a quirky lark of a side-side-project hits darker now. Generative AI advanced more quickly than those in the industry thought possible. Many of us who think about the harm such systems can unleash thought we would have more time to prepare.

In December of 2022, Open AI released ChatGPT, and even they were surprised by its reception. Some inflection point was crossed between “wow, that’s kind of neat for something an AI made” to “this is so good I should be using it daily.” Business models that had sustained massive tech companies for decades were declared not long for the world and they flailed around to mount a response, a signal that they weren’t yet to be counted out.

This powerful technology will shape our lives in unimaginable ways over the next decade, offering the potential to solve problems that deeply impact millions. Protein folding has been solved, paving the way for novel treatments for ailments and a boon to drug discovery. We can use ML systems to flag threats to global food security and proactively route aid for famine prevention. People with disabilities could use generative capabilities to communicate more fluently with people around the world and better perceive the environment around them.

But to what end are we harnessing the technology that some of our brightest minds claim is more revolutionary than fire? Cheating on exams. Spamming science fiction literary magazines with so many AI-generated stories that they have to pause submissions. Replacing low-level content creators trying to build skills and experience with cheaper machine-generated content. We’re focusing the power of this new lens on amplifying the rewards flowing to capital holders because, of course we are.

I want technology to be an aid to human creativity, not a necessary crutch. I want to make life better, not shove people aside because they are no longer needed to feed the capitalist greed beast. And by writing stories with AI, I’m strengthening the narrative that human creativity and ingenuity is nothing more than the value of the tokens they generate.

Maybe this was always the case, and the mass hysteria around generative AI systems has made what used to feel fun feel gross instead. Maybe it’s the reaction of the world–that interaction between immense power and the realities of our economic system. Either way, I won’t participate in this project any longer, though don’t begrudge others if they would like to.

I’ll continue writing original fiction, whether or not AI systems continue to progress to the point where human writers can no longer compete. I love creating new worlds, inhabiting other minds, solving plot problems and surprising myself with the places I go. Because creating art is inherently pleasurable and something that enriches my life whether or not it leads to recognition or worldly success, it is something that we should never devalue.

Take care of each other.

Celeste Kallio:

When ChatGPT first blew up the world, I assumed that it was all hype. Then I sat down to use it to help me write stories for this newsletter. My first few minutes with ChatGPT blew my socks off.

I started out asking ChatGPT for some story ideas. After a few misses, ChatGPT gave me the basic concept of the story.

We workshopped the concept:

And then ChatGPT wrote a great opener:

I was both blown away and a bit disappointed. If the process was this easy, what was the fun of it?

I’m working my way through a great craft book, Writing Fiction by Janet Burroway, and each chapter provides several fantastic writing prompts at the end, which I always aspire to do but never quite motivate myself to complete. Why not get AI’s help?

Unfortunately, ChatGPT did not shine in this task. When I asked it to write a scene in which a man asks a woman about her mother. In response, I got ten very bland paragraphs where a woman raves about her wonderful mother and her interlocutor is suitably impressed.

BO-ring.

I asked for a continuation in which the woman’s mother has a dark secret:

“The woman's mother returned to the table and the three of them sat in silence for a moment. It was clear that she had something on her mind. Finally, she took a deep breath and spoke.

"I need to tell you both something," she said. "I was in prison for ten years for embezzlement. I made a mistake and I paid for it. But I've turned my life around and I've been trying to make amends ever since."

The man was shocked but he could see the sincerity in the woman's mother's eyes. "I'm sorry for what you've been through," he said. "But I'm glad you've been able to turn your life around."

The woman's mother smiled, relief etched on her face. "Thank you. It's been a long road, but I'm grateful for every day."”

Wow. How novel. How dramatic. How interesting.

With all the stories I’ve used ChatGPT to write, the first attempts by ChatGPT were banal and trite. It was only through repeated effort that I got anything slightly unexpected, which stands to reason given the design of these systems which, at the end of the day, are only predicting the next word in a text stream. Everything it comes out with should revert to a mean of the most common things others say in the same situation. It’s a recipe for basic.

Andrey Kurenkov:

I’ve written my last two stories, I Think I Am Starting to Fear Death and An Account of The Machine Wars, partially with ChatGPT and largely agree with Celeste that it’s both exciting and kind of underwhelming.

For the first story (I Think I Am Starting to Fear Death), I had a pretty fleshed out idea of what I wanted and had already written more than half of it, and used ChatGPT to try and help me finish it up. Its output was certainly valid as far as ways to conclude the story, but it felt pretty… lame? There was no subversion, no twist, no interesting aspect to it really — it went in a sacharine direction that felt at odds with the generally bittersweet vibe I was aiming for.

But, it did help me out a lot later, when I decided to revise a finished draft of the piece. At first I went with an unusual approach of including almost nothing but dialogue, with no interspersed commentary on character’s expressions or tone. It was a fun thing to try, but after re-reading it I definitely felt something was lacking, and I decided to add more text along the lines of “Despite the heavy subject, Cassy’s tone is light and playful, as if she's looking looking forward to one of our drunken philosophical dialogues.”

ChatGPT was really useful for this! The dialogue already conveyed emotion pretty well, so the added ‘flavor text’ by ChatGPT pretty much seemed right.

It was still not perfect — “trying to keep my voice steady.” is off the mark, as the delivery in the story is meant to be more casual, as if it’s just banter. But I could still mostly use what ChatGPT provided, and it did speed up my editing process.

For the second story (An Account of The Machine Wars), I had already written a full draft, but wanted to add more details to it and refine the writing a bit, so I basically asked ChatGPT to do a full revision of the story:

At first I asked it to revise the whole story, which was already somewhat long. Instead of revising the story, it for some reason created a sort of summary without adding anything to the original story:

I was kind of surprised and tried to get ChatGPT to course correct, only to be met with this response:

It was pretty obvious at this point I used up the context window of the language model, and this was pretty easy to confirm:

I did have some success getting it to work by just having it revise the first and second half of the text separately, and it added some decent details without really altering the writing much. But the general reaction I had to what it did was “meh”. I later decided to instead use SudoWrite’s guided writing feature to instead directly add in paragraphs with the details I wanted in the middle of the story, and this worked much better!

Now I am working on my third story with ChatGPT, and this time I started with just a general outline from which I wanted to get ChatGPT to write a draft.

It wrote a pretty reasonable draft, which I then worked to revise with follow up requests like “make this story even more detailed” , “Describe what the man went through in the jail cell in more detail.”, and “Describe the physical appearance and layout of the jail cell in more detail”. Like before, the results were valid, but also kind of… boring.

Perhaps unsurprisingly, ChatGPT tends to output predictable and bland writing, and it’s not easy to steer it in the direction of something that better resembles your “voice”. My take-away is that it can be good for edits that don’t require it to do anything interesting, but otherwise tools that are specifically aimed to help authors with AI (such as SudoWrite and Laika) are far more useful.

Share

Leave a comment

What if we can play between these two extremes, and explore the goofy and creative side of AI’s emerging capabilities for writing fiction and generating art? As an AI engineer/researcher and creative writer, I’m fascinated by the ways that incorporating AI into the creative process can change how we tell stories. And in writing for Stories by AI, I have gotten to explore that fascination and discover how AI tools can be leveraged by artists such as myself today. Let me tell you about it!

It starts with an idea, and in this aspect, writing with AI is much easier than writing a story solo. When I have control over the writing process, I usually combine two ideas in an interesting way, then develop characters that resonate with the idea and change as they navigate the plot. I’ll often have glimmers of ideas bouncing around in my brain for years before I figure out what to do with them! With AI however, I don’t have total control over the writing process, so planning things out defeats the purpose of exploring the AI’s suggestions.

The seed of a story can come from anywhere. Fairytale retellings are evergreen and can always benefit from a strange, machine-generated twist (eg Rapunzel and the Two Towers). Tropes that you love in books and movies can be good inspiration, too. I’ve always been fascinated by intergenerational space travel and the social dynamics therein, which spawned The Mealsludge Revolts. I also keep a folder of interesting articles from Science magazine around, which I often mine for my original fiction.

Starting with an idea alone rarely leads to a compelling story, though. It helps to have a few bullet points. Sometimes, I use conversational AI prototypes to brainstorm plot ideas, though they can often lead to anti-plots. That’s not a total waste of time though - treacting against a bad idea can breed good ones!) This feels very similar to workshopping plot ideas with a human friend. The surreal surprises are part of the appeal, so I don’t get too detailed with it–a couple of bullet points of a plot skeleton are my right balance between meandering freedom and rigidity. And if something weird happens, I go with it and see what happens.

My current tool of choice is SudoWrite, which is based on OpenAI’s GPT-3 with some special sauce added to support the goal of creative writing in particular. Another option is to start with a story outline as a seed, but I’ve found the way the AI tries to pull the plot toward the outline to be a bit awkward so prefer to steer manually. I write the first paragraph or so to introduce the tone and concept, then use the “Write” feature to generate two possible continuations. Usually, one of them is much closer to what I’m looking for than the other, but often they both contain interesting elements, so I will combine bits of them to include in the story.

Without steering, the AI suggestions tend to become circuitous and repetitive, so I often prod it with one of my bullet points to advance the narrative. The best moments, though, are those when the AI takes a wild turn I hadn’t expected. My plan was never for Nicole to fall in love with her sourdough starter.

My goal is to get about 1000 words in a story and edit it down to about 800 so it can be read in a little over five minutes. The AI writing isn’t as concise and clear as ideal, so this is fairly easy to do by removing extraneous phrases and repetitive sections. Of course, my own human-generated text also benefits from polishing in the process.

At that point, I share the story draft with the other folks working on Stories by AI. One of the most interesting things about the project from my POV is that different people find different things interesting. Some are  more interested by the AI aspect, and want the stories to be “less human” and not “too normal.” When we shared our initial batch of stories with our writer friends, they were much more interested in the human side – how the co-writing process worked, and what it said about human creativity and ingenuity. As an insider in the tech world, I am fairly skeptical about AI but appreciate these SOTA models for what they are capable of. I think that creating alongside AI is a fascinating way to explore my own creativity and storytelling along with these new capabilities.

Other writers are exploring this world, too. Technology journalist and novelist Vauhini Vara co-wrote Ghosts, an essay about grief, with GPT-3. (Her debut novel, The Immortal King Rao, was not co-written by AI, but is the best book I’ve read this year.) K. Allado McDowell did use AI as her writing partner for Amor Cringe. You can find more explorations of this mashup in this article by The Verge.

This project has given me a new lens through which to view writing. Creativity can be a lonely and rejection-filled endeavor. With this project, I rediscovered approaching writing with an explorer’s mindset, where I’m excited about what might happen without expectations. AI can be a great creative partner: filling in the gaps when you get stuck, or suggesting an offbeat path that you never would have noticed. The human is still at the center though–the human writing, the human steering, and most importantly, the human reading.