AI Content Generation Is Already a Fiction Engine
AI content generation is not mostly a marketing feature; it is already a fiction engine.
That matters more than the latest publishing scandal. According to WIRED’s reporting on a recent arXiv study, a large share of chatbot activity involves fiction generation, including prose, poetry, fan fiction, erotica, and role-play. The exact number is still contested, but the user behavior is not. If you build products, workflows, or policies around generative AI and assume the main job is ad copy, blog drafts, or product descriptions, you are reading the market through the wrong end of the telescope.
I have seen this mistake in practice. In one client workshop, a team came in asking how to standardise AI writing tools for campaign copy. Within 30 minutes, the most active users were not trying to write landing pages. They were stress-testing memory, character consistency, branching dialogue, and prompt iteration because that was the fastest way to expose where the model and interface broke down. Creative use showed us the product requirements faster than the business use case did.
AI content generation is already a fiction habit
The strongest signal in the source material is not cultural. It is operational. Melanie Walsh, a coauthor from the University of Washington Information School, told WIRED that users know the stories are produced by AI and do not care; in fact, that seems to be part of the attraction. That is a blunt correction to the common executive assumption that machine-written narrative has no durable audience.
The study drew on WildChat, a public prompt dataset, and found that more than a third of chatbot interactions involved some form of fiction generation. Even if you discount that number because the sample is public and self-selected, the pattern is still hard to dismiss: people are spending serious time on AI storytelling, not just sampling it once for novelty.
The Petra Ferraz de Novaes example in the WIRED piece makes the point better than any dashboard can. She used SillyTavern to run open-source models locally and kept returning because the model took the story somewhere unexpected. In business writing, unpredictability is usually a defect. In role-play AI or fan fiction AI, it is part of the product.
The numbers point to power users, and that is still a big deal
Here is the part I think many teams misread: a power-user pattern does not make the category irrelevant. It usually means the opposite.
The researchers looked at more than 500,000 prompts. They found that heavy users generated a disproportionate amount of the fiction activity, including cases where one person iterated thousands of times on a single scenario. After adjusting for that skew, they still estimated that 7 percent of chatbot users engage in fiction generation. That is not everybody, but it is far beyond a rounding error.
There is also a valid challenge to the headline. A separate 2025 OpenAI and NBER working paper suggested that 1.4 percent of ChatGPT prompts were fiction-related. OpenAI also noted, via WIRED, that WildChat is not representative of all ChatGPT users. Fair enough. But for product teams, the split between 1.4 percent and 7 percent does not change the design lesson very much.
If even 1.4 percent of prompts on a mass platform are fiction-related, that is still an enormous usage class in absolute volume. If it is 7 percent of users after controlling for prompt obsessives, then the case is stronger still. Either way, AI content generation is not confined to marketing copy.
Why hallucinations become features in creative writing AI
This is where most enterprise discussions fall apart. Teams inherit evaluation habits from summarisation, customer support, or knowledge retrieval, then apply them to creative writing AI as if every use case should be graded on factual stability.
That is the wrong frame.
For fiction, the model’s loose edges can create the feeling of collaboration. Novaes told WIRED, “What fascinates me is the unpredictability.” I believe that, because I have watched users keep going precisely when the model stops behaving like a deterministic template machine. They branch. They retry. They pin a character trait, then deliberately loosen the scene description to see what comes back. That is not failure recovery. That is the workflow.
The practical consequence is that AI writing tools aimed at creative use need a different interface from enterprise content systems. They need:
- fast iteration on prompts
- visible memory and context controls
- branching or version comparison
- lightweight safety controls tuned to role-play, not only business prose
- enough continuity to sustain sessions over hours, not minutes
This is also why private role-play and public publishing should not be lumped together. A person generating AI fiction prompts on a local setup or in a chat interface is looking for co-creation, surprise, and persistence. A publisher evaluating a manuscript is looking for authorship, consistency, and trust. Same model family, different product job.
The steel-man case against this thesis
The strongest counter-argument is easy to state: consumer entertainment behavior does not tell enterprise buyers much. Publishing controversies show that readers still care about authorship. Public datasets like WildChat are noisy. And a lot of fiction prompting may simply reflect a small group of enthusiasts pushing the system hard rather than a broad market.
I think that argument is serious, not straw.
If your only question is whether professional publishing will accept AI-generated novels as equivalent to human-authored work, the answer is clearly no, at least not cleanly and not soon. Readers, editors, and booksellers are treating authorship as part of the product. The BBC’s reporting on a withdrawn novel and The Atlantic’s discussion of AI-related literary backlash both point in that direction.
But that is also why publishing scandals are a bad benchmark for judging AI storytelling demand.
Why the rebuttal still stands
Users do not need the publishing industry to validate a behavior for it to matter. They only need enough utility or enjoyment to repeat it.
That is the core signal here. Repeated prompting, scenario variation, long sessions, and private role-play all indicate that AI content generation is functioning as an interactive medium. Not a fake author. Not a bad junior copywriter. An interactive medium.
Once you see it that way, a lot of product decisions look different. Memory stops being a nice-to-have. Prompt UX becomes retention infrastructure. Session controls matter more than one-shot output quality. And team training becomes important because people inside companies will discover these uses whether leadership planned for them or not.
For media, publishing, and software teams, the near-term question is not whether AI will replace fiction writers. It is whether your product or workflow recognises creative intent when it appears. If users keep bending a copy tool into a story engine, that is evidence. In one recent review I ran, the biggest drop-off point was not model quality. It was that users could not easily revisit earlier branches of a conversation. They wanted a dungeon master. We had given them a form field.
That is also where a service like AI Content Generation Solutions fits best: not as a promise that every team should mass-produce prose, but as a way to map which content tasks need repeatability and which need open-ended generation.
The hot take, then, is simple: the market has already voted with its prompts. AI content generation is not just a copy workflow with extra press coverage attached. It is a creative behavior category, and teams that ignore that will mis-spec the product, the training, and the guardrails.
Written by the Encorp team. Talk with us: book a 30-min call or follow us on LinkedIn.
Martin Kuvandzhiev
CEO and Founder of Encorp.io with expertise in AI and business transformation