The authenticity penalty

A comprehensive research project comprising 16 experiments has revealed a persistent reality for publishers: readers consistently undervalue creative writing when they believe AI was involved in its creation, even when the quality of the content remains the same. This “AI disclosure penalty”, influenced by perceived authenticity, proved resistant to multiple evidence-based interventions and persisted throughout a 15-month study period (March 2023–June 2024).

The publishing industry faces an unprecedented dilemma. Generative AI tools offer significant productivity gains, yet transparency regarding AI usage may fundamentally undermine audience appreciation. This tension is intensifying as the US Congress considers the AI Disclosure Act of 2023, which could potentially mandate the disclosure of AI involvement in creative works. For publishers, this creates a critical trade-off: whilst AI enhances production efficiency, disclosure requirements could reinforce negative biases that affect how content is received and its commercial viability. This research provides the first large-scale, systematic evidence of how AI disclosure affects the evaluation of creative writing.

Researchers from Wharton, the University of Michigan and NYU Stern conducted 16 pre-registered experiments involving 27,491 participants. The studies used creative writing generated by ChatGPT and award-winning short stories written by humans. Participants assessed samples they believed had been created by ‘an AI tool’, ‘a human’ or through ‘human-AI collaboration’.

The meta-analysis revealed a robust negative relationship between AI disclosure and evaluations (p < .001). In 14 out of 16 studies, the effect was negative and significant. The penalty reduced evaluations by 6.2 per cent on average, with a Cohen’s d of 0.24—a small but meaningful effect. Crucially, the effect remained consistent across different samples, participant groups and the entire study period, which spanned a period of rapid AI evolution.

Researchers systematically tested interventions that have mitigated algorithmic bias in other contexts—all failed to reduce the penalty.

Content characteristics: Studies examining narrative perspective, format (poetry versus prose), emotional tone, and the ‘humanity’ of characters (human, animal, alien or robot protagonists) found no reliable moderation patterns.

Evaluation context: Reframing evaluation as artistic rather than objective or utilitarian made no difference. The penalty persisted equally in both contexts.

Information on AI capabilities: Exposing participants to articles about AI’s emotional or cognitive sophistication successfully altered their perceptions but did not mitigate the penalty. Believing that AI possessed greater capabilities did nothing to reduce bias.

Attempts at humanisation: anthropomorphising AI through names, gender and backstories produced inconsistent results across replications, offering no reliable mitigation strategy.

Human-in-the-loop framing: Most crucially, emphasising human involvement provided no relief. Studies comparing disclosures of AI-only, human-only and human-AI collaboration found that both AI-only and collaboration disclosures produced almost identical negative effects. Using award-winning stories written by humans rather than ChatGPT content confirmed this pattern, ruling out differences in quality as the cause.

Readers were just as dismissive of the writing when told it involved collaboration between humans and AI as when told it was generated purely by AI.

Across eight studies, perceived authenticity emerged as the consistent mediator. Disclosure of AI use was negatively associated with perceived authenticity; authenticity was strongly associated with evaluations; and controlling for authenticity rendered the direct effect insignificant.

The meta-analysis revealed that the total effect of AI disclosure was -0.327 (p < .001). When broken down, the indirect effect via perceived authenticity was -0.292 (p < .001), whilst the direct effect, controlling for authenticity, was only -0.034 (p = .240)—statistically insignificant. The penalty operates almost entirely through reduced perceptions of authenticity, with authenticity’s relationship to evaluations (β = 0.611, p < .001) remaining robust even when controlling for other factors.

Lead researcher Manav Raj and his colleagues articulated the dilemma: "Creators cannot be transparent about using AI to produce creative content without undermining appreciation for that very content—unless there are ways to reduce the penalties associated with disclosing AI use." They emphasised the persistence of this effect: “The AI disclosure penalty is remarkably persistent, holding true throughout the duration of our study; across different evaluation metrics, contexts and types of written content; and across interventions derived from previous research.”

Publishers face complex challenges. The research shows that the negative impact of disclosure cannot be mitigated through framing, manipulation of context or emphasising human collaboration. If disclosure legislation is passed, publishers required to disclose AI involvement may see a decline in readership, regardless of quality.

The fact that the effect remained stable over 15 months suggests that these biases are “stubbornly difficult to mitigate, at least for the time being”. However, the researchers note that this may change “as we become more accustomed to AI-generated creative works”.

Understanding how the penalty arises from perceived authenticity provides a conceptual framework. Rather than focusing on the method of creation, publishers might emphasise what makes a piece of work valuable beyond its production—unique perspectives, emotional truth and editorial curation. For content where AI provides clear value (data-driven journalism, real-time updates), publishers might develop new frameworks for discussing authenticity, centred on editorial integrity and serving readers.

The industry requires longitudinal research to track developments as AI becomes ubiquitous, as well as studies into whether editorial framing can mitigate these effects in real-world publishing contexts involving established brands and reader relationships.