Meta is taking a cautious approach with its AI development. According to a recent policy document, the company might stop working on or releasing AI systems it deems too risky. This framework, called the Frontier AI Framework, highlights two categories of AI systems that could pose significant dangers: “high-risk” and “critical-risk.”

High-risk systems are those that could make cyberattacks or the creation of biological weapons easier, though not necessarily foolproof. Critical-risk systems, on the other hand, could lead to catastrophic outcomes that are hard to control or fix. Meta gave examples like AI that could hack highly secure corporate systems or enable the spread of devastating biological weapons.

Interestingly, Meta isn’t relying on strict tests to judge these risks. Instead, decisions are made based on input from internal and external experts, reviewed by senior-level executives. The company admits that current evaluation methods aren’t precise enough to provide clear-cut answers.

If Meta identifies a system as high-risk, it won’t release it publicly until improvements are made to reduce the risks. For critical-risk systems, Meta says it will add extra security measures and halt development altogether until the system can be made safer.

This new framework seems to be a response to criticism of Meta’s “open” AI strategy. Unlike companies like OpenAI, which keep their AI systems behind closed doors, Meta has been making its AI technology widely available. While this approach has led to widespread use of its AI models, such as Llama, it has also raised concerns about misuse—like when a U.S. adversary reportedly used Llama to create a defense chatbot.

Meta’s move also sets it apart from other open AI providers, like China’s DeepSeek, whose systems reportedly lack safeguards and can easily produce harmful content. Meta argues that balancing the benefits and risks of AI is key to ensuring the technology helps society without causing harm.

By taking this cautious stance, Meta aims to show it’s serious about responsible AI development—even as it continues to push the boundaries of what AI can do.