Does partnering with an AI chatbot genuinely enhance creativity, or does it simply create the illusion of creativity through more complex language? Prior work suggests that human--AI collaboration can raise creativity scores, but that this advantage often disappears once verbosity is controlled for, whereas human--human collaboration tends to yield more diverse and original ideas. However, existing co-creativity research remains fragmented in how creativity is measured, and largely focuses on final products rather than the collaborative process. 

To address these gaps, we examined co-creative short story writing with expert ratings and automated metrics of diversity, novelty, surprise, and linguistic complexity, and analyzed the conversations for idea generation, elaboration, and synthesis. In a within-subjects design, 40 participants (N = 20 pairs) each wrote stories alone, with a human partner, and with a generative AI chatbot (GPT-5.1). Our results replicate the previous studies. We found that human--human collaboration produced more semantically diverse and surprising stories than human--AI collaboration, whereas human--AI stories were more lexically complex. 

These results point to a fundamental problem of an "AI hivemind'': large language models tend to pull collaborative output toward semantically similar, average responses (Jiang et al., 2025). Process analyses clarified this pattern. In human--AI pairs, AI chatbot contribute much more to the final story, while human-human pairs have a more balanced contribution, providing a potential explanation for the lowered creativity in human-AI pairs. Indeed, higher creativity was best predicted by a more even contribution balance, or a better synthesis of ideas from both sides. This result illustrates that it is the human partner's active, sustained contribution that can counteract the AI's tendency toward homogenized output and injects the variability needed to keep the final story original.