From Chat to Agent: Why Consent Breaks When AI Acts on Your Behalf

The rapid evolution from conversational chatbots to autonomous AI agents has outpaced consent mechanisms designed to protect users. While millions share sensitive information with LLM-based systems, evidence shows a widening gap between users’ mental models of data handling and actual platform practices. This disconnect deepens as AI browsers navigate the web, handle credentials, and execute transactions. This paper argues that consent is structurally broken across the LLM ecosystem and identifies three design tensions: personalization versus data minimization, agent autonomy versus user control, and transparency versus usability. Building on these tensions, this paper proposes LLM-specific consent mechanisms: contextual in-flow transparency, task-scoped permissions, and AI-mediated privacy guardians, and highlight their consequences for security, comprehension, and power asymmetries in agentic systems.