People are increasingly asked to judge creative work that may or may not come from a machine. The label attached to a poem or painting can shape how impressed we are before we have looked closely. A new set of experiments from China suggests that people rate art as less creative when it is labelled as made by artificial intelligence, yet the same label can work in its favour for idea generation tasks.

Researchers at Hangzhou Normal University and Shenzhen Polytechnic University ran three experiments. The findings appear in BMC Psychology. In the first, painting and poetry outputs were presented to participants as made by a human or by AI, although the works were identical.

The second experiment looked at works described as co-created, varying how the credit was shared between a human and AI. The third used a divergent thinking task, in which people list unusual uses for an everyday object. Again, the same output carried different labels. The abstract does not give sample sizes.

Divergent thinking is the ability to come up with many varied ideas in response to an open question. Perceived originality is how new or unusual a work seems to the person judging it. Attribution bias is the tendency to judge the same thing differently depending on who is credited.

In the first experiment, outputs labelled as human were rated as more creative than identical outputs labelled as AI in artistic creation. The label alone changed the verdict. The content of the work was the same in both cases.

In the second, a label saying that a human had the ideas and AI assisted raised perceived originality in poetry. For painting, works credited mainly to humans stayed on top. The third experiment reversed the pattern, because the same divergent thinking output received higher creativity ratings when it carried an AI label.

The authors read this as a domain-specific pattern. People appear to welcome collaboration with AI, yet still discount what AI contributes when the work is art. In idea generation, which is less tied to personal expression, the AI label did not carry that penalty.

The study has clear limits. The abstract reports no sample sizes or effect sizes, so how large or reliable the differences are cannot be judged here. The experiments tested a few tasks with materials made partly with AI tools, and ratings in a lab-style task may not predict how people respond to creative work in everyday life. The reasons for the pattern were not tested.

Earlier studies have found that people tend to devalue art when told an algorithm made it. This work suggests the bias is not universal, and that the type of task may decide whether an AI label helps or harms.