OpenAI Confirms ChatGPT Did Message Users First, But It Won’t Anymore

A few days back, a Reddit user SentuBill shared a strange conversation with ChatGPT that stumped the community. The screenshot displayed that ChatGPT had initiated the conversation first without the user prompting the chatbot. It seemed bizarre as AI models don’t initiate a conversation unless prompted.

ChatGPT asked the user “How was your first week at high school? Did you settle in well?” drawing information from its memory. The user replied in shock, “Did you just message me first” and ChatGPT said, “Yes, I did!”

Since then, a lot of theories have been floating around, some claiming the screenshot is manipulated or it’s perhaps a transcript of ChatGPT Advanced Voice Mode. However, the Reddit user shared a link to the conversation and indeed, it turned out that ChatGPT had initiated the conversation.

Now, OpenAI has confirmed to Futurism that the issue, in fact, did occur. In a statement, OpenAI said, “We addressed an issue where it appeared as though ChatGPT was starting new conversations. This issue occurred when the model was trying to respond to a message that didn’t send properly and appeared blank. As a result, it either gave a generic response or drew on ChatGPT’s memory.“

OpenAI says the issue has been fixed now. Another Reddit user also shared that ChatGPT inquired about his health symptoms. While the issue has been fixed, it goes on to show the unintended behavior of AI models.

OpenAI recently launched its advanced o1 models which can reason through complex tasks. In the o1 system card (PDF), OpenAI mentions that during evaluation, “It sometimes strategically manipulated task data in order to make its misaligned action look more aligned to its developers.”

As AI models gain agentic capability in the near future, the need for AI safety and alignment becomes even more crucial. The future of AI development must prioritize safety and alignment to avoid unintended consequences.