OpenAI promises greater transparency on model hallucinations and harmful content

Open started a new website called Security Assessments Center to share information about things related to things like hallucination rates of models. If a model produces malicious content, it will emphasize how the hub instructs and how they treat jailbreaks.

Tech Company claims that this new page will show additional transparency in Openai in a company facing context Multiple trials Used illegally copyrighted material to bring up AI models. Yes, yes and you need to remind it New York Times Claiming Tech company Accidentally evidenced In a plagiarism of the newspaper against him.

The Security Assessment Center is designed to expand on Openai system cards. They just launched security measures, while the HUB must provide ongoing updates.

“We aim to share our progress on the development of more expanding roads to measure the model abilities and safety of the AI ​​assessment,” he said. “We hope that the security of the security of the assessment here will make it easier to understand the security of Openai systems in time, but also will not support community efforts to increase transparency throughout the field.” Openai adds that the company is working to have more active communications in this area.

Interested parties can view HUB sections and see information about appropriate models such as GPT-4.1 4.5. Openai notes that the information provided in this center should only look at the “instant” and stakeholders on system cards. Evaluations and other releases for further information.

One of the adults Amps Openai to the entire security center is the choice of these tests and which information is open to share. As a result, there is no way to guarantee all the problems or concerns of the company to share the public.