OpenAI has announced a significant step towards greater transparency in AI safety by committing to publish the results of its internal AI model safety evaluations more regularly. This move aims to enhance accountability and provide insights into the safety performance of OpenAI's systems over time. To achieve this, the company has launched a Safety Evaluations Hub, a webpage that showcases how its models score on various tests for harmful content generation, jailbreaks and hallucinations.
The Safety Evaluations Hub will be updated with major model updates going forward, and OpenAI will share metrics on an ongoing basis. This development comes amid growing scrutiny over AI safety and transparency, with critics accusing OpenAI of rushing safety testing for certain flagship models and failing to release technical reports for others. In response, OpenAI emphasizes its commitment to safety, highlighting its practices such as empirical model red-teaming and testing, alignment and safety research, and monitoring for abuse.
OpenAI's empirical model red-teaming and testing involve evaluating model safety before release, both internally and externally, according to its Preparedness Framework. The company's alignment and safety research focus on building smarter models that make fewer factual errors and output less harmful content, even under adversarial conditions. Additionally, OpenAI leverages tools to detect safety risks and abuse, sharing critical findings to help others safeguard against similar risks.
By publishing its safety evaluation results, OpenAI aims to promote accountability and collaboration in AI development. The Safety Evaluations Hub is a step towards greater transparency, allowing the public to understand the safety performance of OpenAI's systems and supporting community efforts to increase transparency across the field. This move demonstrates OpenAI's commitment to prioritizing safety and transparency in its AI development processes.
