AI-written stories fool readers just fine. It is knowing that ruins it, study finds

“People actually prefer AI-generated stories while believing that these stories are worse,” the authors wrote.

Why people devalue AI writing

Gautam Mehra, founder of Winterline AI Research in Pune, which helps manufacturing firms identify and evaluate the right AI technology providers for real-world deployment, says “people’s distrust is embedded in the word ‘artificial’,” and an inherent denial that the machine might be able to outdo a human being when it comes to creativity.

People, he adds, often value human writing because it reflects real experiences, emotions, and personal stakes. They trust a human writer more as they intentionally chose to communicate with the reader.

“Imperfections can signal individuality and sincerity. Effort itself has meaning, much like a handwritten letter compared with a perfectly printed one,” says Mehra, adding that a human author allows readers to attribute an idea to the person who came up with it.

“AI may write more fluently, but fluency is not the same as meaning. Human-created work can feel scarce and original, while AI can produce polished text almost instantly,” Mehra says. “People often value the human passage as evidence of a genuine act: someone thought this, felt this, and decided to say it. In that sense, the history and authorship of a passage are part of its value, not merely the words on the page.”

Priyamvada Dixit, author of the short story collection Banarasi Afsane, offers a writer’s perspective. “An artist creates art through years of toil, tears, and blood and still calls himself a student. AI learns from different artists, and declares itself the master,” says Dixit.

“If a writer uses algorithms in their work,” she adds, “we would see neither the pain of Nirmala nor the stories of Manto; there would be no wit like Mirza Ghalib’s, nor heartbreaking plays like Shakespeare’s. And we would be able to read a thousand stories in only a handful of words.”

Dr Munia Bhattacharya, senior consultant psychologist and head of department at Marengo Asia Hospital in Gurgaon, points to the emotional reasons behind algorithm aversion: a lack of emotional connection with machines, a fundamental trust in humans over machines, fear of AI mistakes or “hallucinations”, privacy concerns, and a sense of losing control. “AI is driven by data, not personal experience,” she says. “That feels alien to us.”

Human or AI? Can we tell

In two other experiments, participants were asked to identify which of two stories was written by a human. In one experiment, 39.4 per cent guessed correctly. In the other, which was run months later when more people had been exposed to AI, performance was effectively random at 52 per cent.

“There is no general evidence that participants were able to distinguish human-written from AI-generated stories,” the researchers concluded.

The researchers also found that those who relied on parameters such as language and personal enjoyment tended to guess incorrectly. This might also explain the increased reliance on AI detection tools such as Pangram.

Recently, allegations were levelled against Daggermouth author H M Wolfe. Their publisher and fans backed it, but debut crime novelist Jerry Falade, who secured a more-than-$2-million two-book deal after a 14-way bid, was not so lucky. He lost the deal after the agent could not authenticate how the manuscript evolved.

Jane Friedman, who reports on the publishing industry, took to Threads to write: “In the current environment, I expect more literary agencies to start using Pangram or some type of AI detection. They should disclose the fact that they use AI detection on their submission guidelines page. It looks something like this (see screenshot). One agent I’ve spoken to about their use says it’s partly a protective action for all involved. The agent needs to know the result before the publisher knows/runs their own detection so they can flag it to the client as a potential problem.”

A screenshot of the post (Threads/@janefriedman)

Regarding why so many respondents preferred AI-generated text, the researchers suggested that AI fiction tends to be more fluent and easier to comprehend. Another reason was that it tended to be more emotionally upbeat than human-written texts. There is also a regression-to-the-mean effect: because language models are trained on large amounts of data, they smooth over individual quirks.

Familiarity with AI models

The study also showed that participants with greater familiarity with AI tools were more likely to correctly identify AI-generated work.

This also became part of the public debate following the Commonwealth Short Story Prize controversy, after three of the five finalists were accused of using AI to write their short stories, prompting questions about the jury’s ability to identify AI-generated work.

Sneha Sharma, a content producer at a private company, says, “My colleagues and I can immediately zero in on AI writing because it’s everywhere. It has some easily identifiable tells. The more AI writing we come across, the easier it is to identify it. It may be fluent, but it lacks the quirkiness that makes each human voice unique.”

The study, however, was limited to stories that took about five minutes to read, not long-form work. Whether AI can sustain quality and character development over the span of a full novel remains an open question.

Nevertheless, the study demonstrates that AI-written texts can fool and even please readers. And yet, when we know it is a machine, we still turn away.