[隐私政策] How will OpenAI mitigate harmful biases and other negative impacts caused by the models served through the API?

Mitigating negative impacts such as harmful bias is an extremely important challenge facing the entire industry. Ultimately, API models do have biases (as shown in the GPT-3 paper), and these biases sometimes appear in generated text. API models may also cause harm in ways not yet imagined. The following are steps taken to address these issues:

Work with users to develop usage guidelines to help them learn from one another and mitigate these issues in practice.
Work closely with users to gain an in-depth understanding of their usage and develop tools to flag and intervene in manifestations of harmful bias.
Conduct our own research on harmful bias and broader issues of fairness and representation, which will help advance work for users.