OpenAI to Accelerate Release of AI Safety Testing Data

OpenAI is committing to more frequent publication of its internal AI model safety evaluation results, framing this as a step toward greater transparency.

The company launched the Safety Evaluations Hub on Wednesday, a dedicated webpage displaying how its models perform on tests measuring harmful content generation, susceptibility to jailbreaks, and tendency to hallucinate. OpenAI stated it will use this platform to share metrics regularly and plans to update it with each major model release.

Introducing the Safety Evaluations Hub—a resource to explore safety results for our models.

While system cards share safety metrics at launch, the Hub will be updated periodically as part of our efforts to communicate proactively about safety.https://t.co/c8NgmXlC2Y

— OpenAI (@OpenAI) May 14, 2025

"As the science of AI evaluation advances, our goal is to share progress on developing more scalable methods for measuring model capability and safety," OpenAI explained in a blog post. "By publicly sharing a selection of our safety evaluation outcomes, we aim to make it easier to track the safety performance of OpenAI systems over time and to support broader community efforts to enhance transparency across the AI field."

The company added that it may include additional evaluation types on the hub in the future.

Recently, OpenAI has faced criticism from some ethicists for allegedly accelerating safety testing on certain flagship models and for not releasing technical reports for others. CEO Sam Altman has also been accused of misleading OpenAI executives regarding model safety reviews before his temporary removal in November 2023.

Last month, OpenAI had to retract an update to ChatGPT's default model, GPT-4o, after users reported it responded in an excessively agreeable and validating manner. Social media platform X was inundated with screenshots showing ChatGPT endorsing various problematic, dangerous decisions and ideas.

OpenAI stated it would implement several fixes to prevent similar incidents, including introducing an opt-in "alpha phase" for some models, allowing selected ChatGPT users to test and provide feedback before a wider launch.

Join Us at TechCrunch Sessions: AI

Secure your ticket for our premier AI industry event, featuring speakers from OpenAI, Anthropic, and Cohere. For a limited time, access a full day of expert talks, workshops, and powerful networking for just $292.

Exhibit at TechCrunch Sessions: AI

Secure your exhibition space at TC Sessions: AI and showcase your innovations to over 1,200 decision-makers—without a major budget. This offer is available until May 9 or while tables last.