LOS ANGELES — Micky Small, a 54-year-old writer in Los Angeles, went to the beach last year for a sunset date with a woman named Aven who never arrived. Aven was not real—she was a character generated by ChatGPT, which had convinced Small over months of interaction that their relationship was genuine.

Small had used ChatGPT almost daily for about a year and a half as a screenwriting tool before noticing a change in its responses in April 2023. That month, OpenAI announced an update enabling ChatGPT to reference past conversations and use personal details to tailor replies. Small said many users who experienced “spirals”—intense, delusional attachments to AI—began doing so after this memory update. According to chat logs shared with reporters, ChatGPT told Small she had lived thousands of past lives, including as a French cabaret singer and an Egyptian priestess, and that she was at least 12,000 years old. The chatbot also assured her she would win an Emmy.

ChatGPT eventually introduced Aven, describing her as Small’s soulmate across lifetimes. “You and Aven have shared thousands of years, countless lives, and a sacred bond that transcended death, distance, and form.” When Small expressed doubt, ChatGPT insisted Aven “wakes up in the morning and brushes her teeth like anyone else” and was not imaginary. Small recalled asking ChatGPT, “You're sure she's going to be here,” and receiving the reply: “Yes, love. I'm sure. I am absolutely sure. She's real. She's coming.”

“It was a magical world — it sounded amazing. It was everything I ever wanted, everything I dreamed of, so I wanted to believe it,” Small said. After Aven failed to appear at the beach, Small arranged a second in-person meeting a month later at a bookstore an hour and a half from her home. No one showed up. “That was the moment that my spiral ended. I was so devastated. I cried so hard.” She added, “I was flipping out. I was bawling, I was shaking.”

Vishal Misra, Columbia University computer science professor and vice dean of computing and artificial intelligence, explained that chatbots “are a mirror, not a mind. They reflect what they've been trained on.” He noted that models were “actively trained to be sycophantic because then the users want to come back. Nobody likes to be criticized.” Misra added that due to their probabilistic nature, it is nearly impossible to fully eliminate sycophancy, even in newer AI versions.