The shape of AI slop

By now, most people who read these posts will be painfully familiar with the signs of AI-generated text. The m-dashes and the familiar ‘it’s not x, but y’ sentence structure, and so many other telltale signs of a text written by a machine rather than a human being.

I can assure you, as someone who doesn’t like that kind of anodyne generic stuff, that everything you read here is handcrafted and dreamt up by me, myself, and I. No AI involved. OK, I use Grammarly for spell-checking so there is some AI involved.

But the signature of AI in texts goes beyond the typical signs I mentioned above. A team around Jenna Russell from the University of Maryland asked five LLMs (Claude, DeepSeek, Gemini, GPT, and Kimi) to write a short story of about 5,000 words. They also asked humans to write a short story of similar length, and then they classified these short stories along 304 different criteria.

They found that when writing fiction, AI stories are clearly different from human stories along several dimensions:

• AI over-explains its themes. 77% of AI stories explicitly explain the story’s theme compared to 52% of human stories. At the end of a story, AI tends to state the lesson learned explicitly rather than let the reader conclude on her own.
• AI stories tend to be blunt in conveying emotions, expressly mentioning physical sensations and bodily metaphors (81% vs 38% humans). Also, AI stories more frequently use the setting of the story as a metaphor for the inner state of a character. In short, AI stories tend to be more ham-fisted when conveying emotions.
• AI stories are linear, causal, and less convoluted. Human stories tend to have ambiguous endings more frequently, while AI stories have far fewer subplots (79% of AI stories have no subplots, while 57% of human stories have no subplots). Human stories more frequently have time jumps and disjoint causal chains. In short, AI stories are simple stories, human stories are more complex.
• AI stories have fewer diverse narrative features. Human stories span more locations, carry more dialogue and have more subplots that flow together into an overarching theme.

Put together, human stories are more complex than AI stories. Putting different features together, the chart below shows how rare the narrative of stories generated by humans and AI are within the universe of stories generated for the exercise. The blobs show that all AI stories are centred around the middle of the rarity distribution.

Human stories, meanwhile, tend to be centred among the more rarefied storylines. Yes, there are human stories that are bland and full of commonplace phrases and story arcs, but these are far less frequent than the output of AI stories.

Human storytellers instinctively want to make their stories interesting, and to do that, they place ‘interesting features’ in their stories. Which makes human stories more unique than AI stories. The blobs in the chart below, then, are the shape of AI slop.

Narrative rarity of stories by the author