Readers rate AI-generated short stories higher than human ones until they learn a machine wrote them

What happened

A new study has found that readers cannot reliably distinguish AI-generated short stories from human-written ones, and rate the AI-generated texts more favourably when they don't know their origin. According to reporting by The Decoder, more than 2,500 participants took part in the research, and their ability to identify which stories were machine-written was no better than random chance.

The twist came once authorship was revealed: as soon as participants learned a story had been generated by ChatGPT, their ratings of it dropped. The same text that scored well under blind conditions was marked down purely on the basis of disclosed origin, not any change in the content itself.

Why it matters

This is a clean, controlled demonstration of a bias that customer experience teams increasingly need to manage: quality perception is not fixed to the artefact, it's contingent on framing and disclosure. As generative AI moves further into content, service scripts, product descriptions and even personalised communications, brands will face the same fork the study exposes — the same output can be judged excellent or suspect purely depending on whether its origin is known.

For behavioural economics practitioners, this is a textbook case of a labelling or source effect overriding an objective quality signal. It suggests that trust and expectation are doing much of the evaluative work customers think they're doing when they judge "quality" — with real implications for how, when and whether organisations disclose AI involvement in what customers read, hear or receive.

By the numbers

• 2,500+ participants took part in the study assessing AI- versus human-written short stories.
• Chance-level accuracy was recorded when participants tried to identify which stories were AI-generated.

The Renascence take

The headline finding — that people can't spot AI writing — is the less interesting result here. The sharper signal is that disclosure alone degrades perceived quality, independent of the work itself. That's a service-design problem, not a technology problem.

Most organisations treat AI disclosure as a compliance checkbox, bolted on after the content decision is made. This study suggests it should be treated as a design variable in its own right, because the act of labelling something "AI-generated" is itself an experience touchpoint that reshapes judgment. The operators who get ahead here won't be the ones hiding AI use, nor the ones disclosing it clumsily as a caveat — they'll be the ones who design the disclosure moment itself: pairing it with context, evidence of human oversight, or a clear rationale for why AI was used, so trust is built into the reveal rather than lost to it.