<div> When Social AI Agents Help or Hurt:&nbsp; </div> <div> Dual Agency and Goal Incongruence in Influencer Platforms </div>

Social AI agents are increasingly deployed on social media platforms to stimulate user engagement. Although prior research highlights the engagement benefits, users may interpret AI involvement as a manipulative persuasion attempt when their interests and influencers' interests are not fully aligned. We conduct a randomized field experiment on a leading online video-sharing platform to understand how comments posted by social AI agents affect viewer engagement with influencers' videos. The experiment compares two content genres that differ in goal incongruence between influencers and viewers: (i) Cooking Tutorials videos, where influencers' instructional goals largely align with viewers' learning objectives, and (ii) Restaurant Discovery videos, where influencers' commercial incentives may conflict with viewers' desire for authentic, unbiased recommendations. Within each genre, influencers are randomly assigned to either a treatment condition in which AI comments are posted on newly released videos, or a control condition without AI comments. The results reveal asymmetric effects. In low-goal-incongruence settings (Cooking Tutorials), AI comments increase viewer engagement, including likes, shares, and comments. In contrast, in high-goal-incongruence settings (Restaurant Discovery), AI comments reduce engagement, and the negative effect is stronger for influencers with greater advertising content in their prior videos. A complementary online experiment further shows that in high-goal-incongruence settings, AI comments reduce engagement by increasing perceived manipulative intent. Our findings identify goal incongruence as a boundary condition for the effectiveness of AI-generated social interactions and underscore the importance of aligning AI engagement strategies with users' trust expectations.