OpenAI introduces safety models that other sites can use to classify harms

On Wednesday, OpenAI announced the launch of two innovative reasoning models designed to help developers identify various online safety threats on their platforms. Named gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, these models are fine-tuned versions of the existing gpt-oss models released in August. These new models are classified as open-weight, meaning that their parameters are accessible to the public. This transparency allows organizations to tailor the models to their specific policies and needs. Unlike open-source models that provide full source code for users to modify, open-weight models focus on offering clarity in their operations. This feature gives developers insight into how the models reach their conclusions. For example, a website dedicated to product reviews could employ these models to filter out potentially fake reviews, while a forum for video game discussions might use them to categorize posts related to cheating. OpenAI collaborated with Discord, SafetyKit, and the Robust Open Online Safety Tools (ROOST) organization to develop these models, which are currently available in a research preview. The introduction of these models may help OpenAI address concerns raised by critics who argue that the company has prioritized rapid growth over ethical AI practices. Valued at $500 billion, OpenAI's ChatGPT has already attracted over 800 million weekly active users. Additionally, OpenAI recently completed a recapitalization, reinforcing its structure as a nonprofit with a significant stake in its for-profit operations. "As AI technology advances, the tools and research dedicated to safety must keep pace and be accessible to all," stated Camille François, President of ROOST. Interested users can access the model weights through Hugging Face, according to OpenAI.