Content moderation demands meticulous effort, sensitivity, a profound understanding of context, as well as quick adaptation to new use cases, making it both time consuming and challenging. Traditionally, the burden of this task has fallen on human moderators sifting through large amounts of content to filter out toxic and harmful material, supported by smaller vertical-specific machine learning models. The process is inherently slow and can lead to mental stress on human moderators.

We're exploring the use of LLMs to address these challenges. Our large language models like GPT‑4 can understand and generate natural language, making them applicable to content moderation. The models can make moderation judgments based on policy guidelines provided to them.

With this system, the process of developing and customizing content policies is trimmed down from months to hours. 

1. Once a policy guideline is written, policy experts can create a golden set of data by identifying a small number of examples and assigning them labels according to the policy.  
2. Then, GPT‑4 reads the policy and assigns labels to the same dataset, without seeing the answers. 
3. By examining the discrepancies between GPT‑4’s judgments and those of a human, the policy experts can ask GPT‑4 to come up with reasoning behind its labels, analyze the ambiguity in policy definitions, resolve confusion and provide further clarification in the policy accordingly. We can repeat steps 2 and 3 until we are satisfied with the policy quality.

This iterative process yields refined content policies that are translated into classifiers, enabling the deployment of the policy and content moderation at scale.

Optionally, to handle large amounts of data at scale, we can use GPT‑4's predictions to fine-tune a much smaller model.

We are actively exploring further enhancement of GPT‑4’s prediction quality, for example, by incorporating chain-of-thought reasoning or self-critique. We are also experimenting with ways to detect unknown risks and, inspired by Constitutional AI, aim to leverage models to identify potentially harmful content given high-level descriptions of what is considered harmful. These findings would then inform updates to existing content policies, or the development of policies on entirely new risk areas.

Judgments by language models are vulnerable to undesired biases that might have been introduced into the model during training. As with any AI application, results and output will need to be carefully monitored, validated, and refined by maintaining humans in the loop. By reducing human involvement in some parts of the moderation process that can be handled by language models, human resources can be more focused on addressing the complex edge cases most needed for policy refinement. As we continue to refine and develop this method, we remain committed to transparency and will continue to share our learnings and progress with the community.