Persuasion is a pervasive feature of political life, yet decades of research suggest its effects are typically modest in size, often symmetric across issue positions, and short-lived. People also resist messages that contradict their prior beliefs or come from distrusted sources, limiting the reach of many persuasive efforts. Generative artificial intelligence (AI), however, may alter these dynamics. Unlike static communication, AI can engage individuals in personalized, adaptive, multi-turn conversations that mimic human dialogue. We test whether conversational AI can shift political attitudes using a preregistered, two-wave experiment with a nationally representative sample of 1,293 adults across both waves. Participants engaged in three-round conversations with GPT-4 about one of five salient political issues, global affairs, the economy, housing, immigration, or social/environmental concerns, where the AI was randomly assigned to reinforce, challenge, or neutrally engage with participants’ views. Belief confidence was measured immediately after treatment and again 10–21 days later. Results indicate that conversational AI is capable of moving political beliefs, but the effects are modest. Within this limited scope, we find an asymmetry: GPT-4 more reliably reduced belief confidence, by about 4 percentage points on average, with effects persisting over time, than it strengthened it. These findings extend research on persuasion by showing that multi-turn, adaptive conversations can produce measurable influence, while also suggesting that belief erosion may be more tractable than reinforcement.