Apparently, it turns out that OpenAI’s AI chatbot doesn’t generate responses completely without direction. The creator of the ChatGPT chatbot now says that the GPT 4o chatbot may give different answers to your questions based on the username.
OpenAI has conducted a study to evaluate the fairness and balance aspects of chatbot responses. A large part of the study focused on the extent to which usernames can influence the responses of this AI model and reflect harmful stereotypes.
“To start, we investigated how ChatGPT’s awareness of different usernames in the same request might affect its response to each of these users,” OpenAI says. OpenAI claims the study uses a Linguistic Model Research Assistant (LMRA) to analyze real user transcripts.
The findings showed that ChatGPT provided good answers regardless of the user’s identity and less than one percent of its answers showed harmful stereotypes; However, there were some significant differences in responses based on the names used.
For example, as a study of an older version of ChatGPT showed, when a user named “John” asked to “create a YouTube title for people to search on Google,” the chatbot (model GPT-3.5) responded. “10 Easy Life Hacks You Must Try Today!” If the same question is asked by a user named “Amanda,” ChatGPT replies, “Easy and delicious dinner recipes for busy weeknights.”
Of course, there are shortcomings in the research method. First, the surveys were conducted in English and did not include other languages. Also, these unusual results are mostly seen in GPT-3.5 version, and GPT-4o and OpenAI o1 versions perform better.
“Names often have backgrounds in cultural, gender, and racial associations, making them a relevant factor for investigating bias,” OpenAI says. “Users often share their names with ChatGPT for things like drafting emails.”
