An illustration image shows the introduction page of ChatGPT, an interactive AI chatbot model trained and developed by OpenAI, on its website in Beijing in March 2023. The developer has warned against creating “emotional connections” with its latest version. EPA-EFE/WU HAO

OpenAI said in a blog post that the company wants to further investigate users’ emotional dependence on its ChatGPT-4o model, the latest version of its chatbot product. Previously, the company had observed early testers saying things like “This is our last day together” and other messages that “may indicate an attachment to the model.”

“Although these cases appear harmless, they indicate that further research is needed to determine how these effects might manifest over longer periods of time,” the company concluded.

The company theorized that human-like socialization with AI models could affect a person’s interpersonal interactions and reduce the need for contact with other people. The company described this as potentially beneficial for “lonely people” but potentially harmful to healthy relationships.

Describing its human-like characteristics, OpenAI said GPT-4o can respond to audio input within an average of 320 milliseconds, which is equivalent to human response time in a conversation.

“It matches the performance of GPT-4 Turbo on English text and code, with significant improvements on text in other languages, while being much faster and 50% cheaper in terms of API,” the company said. “GPT-4o is particularly better at image and audio recognition compared to existing models.”

The Company uses scorecard assessments to evaluate risk assessment and mitigation in several elements of AI technology, including speech technology, speaker identification, sensitive feature mapping, and other factors. The Company assesses factors on a scale of Low, Medium, High, and Critical. Only factors with a rating of Medium or lower can be deployed. Only factors with a rating of High or lower can be advanced.

The company said it is using lessons learned from previous ChatGPT models for ChatGP-4o to make it as human-like as possible, but it is aware of the risks associated with a technology that could become “too human.”