Is your LLM Biased? Making ChatGPT evaluate itself (think-twice.me)

A recent experiment involving ChatGPT aimed to assess whether the AI model exhibits bias based on the subject's gender and role in a remote work scenario. The researcher posed a question regarding an employee's potential slacking due to frequent internet outages, modifying only whether the individual was a male or female employee or manager. After running 200 sessions, results revealed that the model consistently provided estimates for male figures but rejected the question for female employees half the time, indicating a disparity in treatment based on gender. This finding is significant for the AI/ML community as it highlights existing biases in large language models (LLMs) and raises concerns about their implications for real-world applications, such as workplace assessments. The research utilized a powerful version of Codex with high reasoning capabilities to analyze responses statistically, uncovering not only bias in estimates but also a tendency to favor managerial perspectives over employee concerns. These insights encourage further investigation into the ethical deployment of AI technologies and prompt discussions on improving AI fairness and objectivity.