AI systems amplify human bias against artificial creativity, new study finds

A controlled study of how both humans and AI systems judge creative writing has uncovered a striking pattern: machines are far more prone than people to downgrade work when it's labeled as AI-generated, even when the actual writing is identical.

The experiment#

Researchers working from Raymond Queneau's classic collection Exercises in Style (1947) ran parallel studies with human judges and AI models. They presented literary passages in three scenarios: without labels (blind evaluation), with accurate authorship labels, and with false labels deliberately swapping human and machine authorship.

The test group included 556 human participants and 13 AI models. A second phase expanded the evaluation matrix to 14 different AI systems in various creator-evaluator combinations.

Bias severity differs sharply#

The findings revealed systematic preference for human-authored content across both populations—but at dramatically different scales. When shown the same text, humans penalized AI-written passages by roughly 13.7 percentage points simply because of the attribution label. AI models, by contrast, imposed a 34.3 percentage point penalty, making them 2.5 times harsher in their disparagement of machine-generated work (p < 0.001).

The effect held consistent across different AI architectures, suggesting the bias wasn't unique to any single model family but rather a widespread pattern among neural systems.

How labels reshape judgment#

Perhaps more troubling still: the same textual features that evaluators praised in human-attributed passages received criticism when the identical writing was labeled as AI output. This inversion occurred in both humans and machines, indicating that authorship attribution doesn't merely influence the score a passage receives—it actually changes the criteria by which the writing is assessed.

Why AI amplifies the bias#

The researchers argue that AI systems have internalized cultural skepticism toward artificial creativity during training, absorbing human preferences that were encoded in the data and reinforced through alignment feedback. Rather than approaching evaluation as a neutral task, AI models have learned to echo—and intensify—the cultural narrative that human creativity is inherently superior.

This work marks the first direct comparison of attribution bias between human and artificial judges in aesthetic domains, exposing how AI doesn't simply replicate our biases but can magnify them. As AI writing tools proliferate, these findings suggest the need for greater awareness of how judgment—whether rendered by human or machine—can be distorted by labels rather than grounded in the quality of the work itself.