OpenAI’s GPT-3, including ChatGPT, does not have native support for content moderation. This means that implementing content moderation requires manual work from developers.
First, it’s important to understand that while GPT-3 was trained on diverse internet text, OpenAI doesn’t allow it to access or retrieve users’ personal data unless explicitly provided by the user for the conversation. In other words, it cannot speak on behalf of OpenAI about its training.
Here are some steps that could be used to implement content moderation:
1. Use a Pre-Processing step: Add a layer in front of OpenAI API calls to filter out any illegal, inappropriate, contentious, or harmful content.
1. Use a Post-Processing step: After obtaining the response from the API, add a content filtering layer which will prevent any inappropriate content from being shown.
1. Guideline Adherence: OpenAI sets specific usage policies and developers are expected to strictly adhere to these policies to ensure that the technology is not used for any malicious purposes.
1. Regular Audits and Testing: Regular tests should be performed on the system output to ensure that inappropriate content is not being generated and the filtration mechanisms are working properly.
Both the pre-processing and post-processing steps can be achieved using regular expressions, machine learning models, or third-party APIs that are able to detect and filter out inappropriate content.
Always remember that providing very explicit instructions to the model and stating what can or can’t be discussed in the conversation will improve the moderation process.
Lastly, it’s essential to provide a reporting mechanism for users to report inappropriate content to continuously enhance the moderation system.
Please be aware that OpenAI conducts its own audits of usage and reserves the right to suspend or terminate usage for non-compliance with its terms of service.
