A UC San Diego study found GPT-4.5 was judged human more often than real people in live chats, raising sharper questions about AI disclosure, trust, and online identity.