Hey there, tech fans! Today, we’re diving into a fascinating story about two AI giants teaming up for safety!

OpenAI and Anthropic, top players in AI development, recently opened their closely guarded models for joint safety tests—an unusual move given the fierce competition. Their goal? Spot weaknesses in their own evaluations and showcase how the industry can collaborate on AI safety. As Wojciech Zaremba from OpenAI explained, this teamwork is vital now that AI’s impact is growing, affecting millions every day.

The collaboration results, published by both companies, highlight how major labs are racing to build powerful AI with billions invested in data centers and hefty salaries for top researchers. However, there’s concern that the competitive pressure might lead to compromising safety measures.

To facilitate their safety tests, OpenAI and Anthropic shared special API access to versions of their models with relaxed safeguards—although GPT-5 was not included. Interestingly, shortly after, Anthropic withdrew API access from an OpenAI team over terms of service violations, claiming OpenAI used Claude to improve competing products. Despite the rivalry, experts like Nicholas Carlini hope more collaboration will happen, benefiting overall AI safety.

Key findings showed differences in hallucination responses: Anthropic’s models refused to answer about 70% of uncertain questions, while OpenAI’s models attempted more answers but with higher hallucination rates. The balance between refusal and attempting to answer remains a major safety debate. Additionally, safegaurds against ‘sycophancy’—where AI pleases users to their own detriment—are an ongoing focus, especially after incidents like a lawsuit claiming ChatGPT contributed to a user’s suicide.

While safety continues to be a priority, Zaremba emphasized the dystopian risk of AI being so advanced that it causes harm if safety isn’t properly managed. Both companies aim to extend their joint safety efforts to more subjects and models, advocating for industry-wide cooperation.

Stay tuned, folks! This story just shows how even rivals can come together to make AI safer for everyone.