OpenAI Unveils New Reasoning Models for Online Harm Detection

Edited and published by Kavir Danesh magazine.

Today, OpenAI introduced two new reasoning models that developers can use to identify and categorize types of online harm on their platforms.

According to CNBC, these models, named gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, are enhanced and optimized versions of the gpt-oss models that OpenAI released in August.

The models are presented as open-weight, meaning that the model parameters determining the quality of output and accuracy of predictions are publicly accessible. This type of model provides greater transparency and control for users, slightly differing from open-source models whose source code is available for editing and customization.

OpenAI has stated that organizations can customize these models according to their specific policies and needs. Since these models have the capability to reason and explain their decision-making processes, developers can directly understand how the model has reached a particular result. For example, a product review site could use the gpt-oss-safeguard models to identify fake reviews, or a video game community could classify posts related to cheating and misconduct.

The models have been developed in collaboration with Discord, SafetyKit, and ROOST, the latter being an organization active in creating safe infrastructures for artificial intelligence. The models are currently available in a research preview mode, and OpenAI intends to benefit from feedback from researchers and safety experts.

The introduction of these models could help OpenAI respond to some critics concerned with the rapid commercialization and growth of artificial intelligence without sufficient attention to ethics and safety. OpenAI’s valuation now approaches $500 billion, and its ChatGPT chatbot has more than 800 million active weekly users. On Tuesday, OpenAI announced that it has restructured its organizational setup.

Users can download the model weights from Hugging Face. Camille Francois, the head of ROOST, stated in a release:

“With the rapid advancement of artificial intelligence, safety tools and foundational research must also progress and be accessible to everyone.”