Do authorship labels bias how we judge literary quality?

Two controlled studies built on Queneau's Exercises in Style (1947) find what the paper calls "systematic pro-human attribution bias." Study 1 gave 556 human participants and 13 AI models literary passages from Queneau and GPT-4-generated versions, under blind, accurately labeled and counterfactually labeled conditions. Humans showed +13.7 percentage points of bias (Cohen's h = 0.28, 95% CI: 0.19–0.37). AI models showed +34.3 points (h = 0.70, 95% CI: 0.62–0.78), which the paper calls a 2.5-fold stronger effect (P<0.001). Study 2 used a 14×14 matrix of evaluators and creators and found the bias across AI architectures at +25.8 points (95% CI: 24.1–27.6%). The paper's summary is that AI systems "systematically devalue creative content when labeled as 'AI-generated' regardless of which AI created it."

The design holds the story content fixed and varies only the attribution label, so the difference is attributed to the label. In the Discussion, the same model judging the same two passages under correct and reversed labels gave opposing assessments: a dialect feature praised as "authentic" under a human label became "exaggerated" under an AI label. The paper reads this as evaluators who "bring learned biases about creative agency to their assessments." The abstract adds that the bias may come from training, "including through the preference signals on which they are aligned." That step is offered as a suggestion; the excerpt does not test the alignment data. The Limitations section also notes that AI explanations are generated after each choice, so they show what a model says about its selection, not the process that produced it.

This sits against the nearest notes in a specific way. How much does rhetorical style shift AI review scores? finds that LLM reviewer scores move when only the rhetoric of a manuscript changes. This excerpt finds a similar movement when only the label changes, which suggests the judges respond to cues about a text more than to what the text reports. Does polished AI output trick audiences into trusting it? argues that polished output borrows authority from its presentation; the excerpt shows that the same text loses credit when labeled machine-written, so the authority in these studies runs on provenance as well as polish. Can humans detect AI text if machines can measure it? is about detection. The excerpt does not test detection; its claim concerns judgments that move once provenance is known.

Several things are not established here. The counterfactual condition is named, but its results are not in the excerpt. Human recruitment and sampling are not described beyond N=556. Study 1 uses one generator (GPT-4), one source narrative and minimal prompting, and the human reference set is one 1947 text in one English translation, across thirty selected exercises. The effect sizes therefore describe this paradigm, a style task on one retold story, and are narrower than a general law of how AI judges AI work. What the excerpt supports is that label-sensitive judgment is a measurable risk for creative evaluation, and that the AI shift is larger than the human one in this setup.