Xiaoke Zhang

Video-sharing platforms like TikTok have introduced AI voice to assist creators in video creation. This creates a new form of multimodal human-AI collaboration, where AI automates audio narration and creators control textual and visual modalities. While the AI voice may reduce creators’ cognitive load in video production by offloading the effort of speech recording, it may also diminish their social benefit by substituting their unique vocal identities. Drawing on Social Exchange Theory, we theorize and empirically investigate the impact of AI voice adoption on creators’ video production and consumption leveraging a unique dataset of 554,252 TikTok videos. Using a stacked difference-in-differences model with propensity score matching, we find that AI voice adoption increases creators’ video production by 27%. Moreover, while AI voice reduces audio novelty, it enhances textual and visual novelty by freeing creators’ cognitive resources. However, AI voice in videos lowers viewer engagement. Mechanism analyses reveal that less-experienced creators and those with higher vocal identity congruence with the AI voice benefit more from AI voice adoption, as they produce more videos and experience smaller declines in video engagement. This study contributes to the user-generated content and human-AI collaboration literature while providing practical insights for video creators and multimodal UGC platforms.

Large language models (LLMs) are transforming journalism by directly entering journalistic workflows, introducing new opportunities and challenges for the institutional press. This study investigates how LLM assistance affects journalists’ information sourcing in news production using a mixed-method approach. We begin with a qualitative study of 43 journalists to identify and theorize how LLM assistance affects three core journalistic values: publication promptness, information source quantity, and information source originality. We then compile a large-scale dataset of 1,073,742 news articles from 111 South Korean news outlets and collaborate with industry experts to detect undisclosed LLM-assisted articles. Our event-level analysis shows that LLM assistance accelerates publication but reduces the number of information sources used in news articles, with a larger decline in primary sources than in secondary sources. Heterogeneity analyses and a randomized experiment suggest that this reduction is driven by two mechanisms: an LLM generation mechanism that narrows the set of retrieved and represented sources, and a metacognitive regulation mechanism that reduces journalists’ active search and evaluation. We further show that these effects extend beyond individual articles. A journalist-level difference-in-differences analysis indicates that LLM adoption leads to persistent reductions in source usage over time. Our findings offer practical implications for LLM system design, newsroom practices, and institutional disclosure policy.