Meta has unveiled a new artificial intelligence (AI) risk assessment framework designed to prevent the development and release of AI systems deemed too dangerous. The framework, detailed in a recent policy report titled The Frontier AI Framework, categorizes AI risks into three levels: critical, high, and moderate.

According to Meta, AI models posing the highest level of threat will not be developed or released. The risk evaluation process begins by identifying catastrophic threat scenarios, such as cyberattacks or the creation of biological weapons. AI models are then assessed to determine whether they could enable such threats, followed by a classification based on potential impact.

Under the framework:

• Critical risk: AI systems that could uniquely enable catastrophic scenarios. These will not be developed or deployed.

• High risk: Models that increase the likelihood of such scenarios occurring. A small group of experts may have limited, tightly controlled access to these under strict oversight.

• Moderate risk: Systems that do not significantly escalate harmful outcomes but will still be subject to safeguards before release.

Regulatory Pressures and Open-Source Scrutiny

The introduction of this framework comes amid heightened scrutiny of Meta’s AI strategy, particularly its open-source approach. Unlike OpenAI and Google, which limit access to their AI models, Meta has made its Llama models widely available. This has sparked concerns over potential misuse, with reports last year indicating that Chinese military researchers used Llama to develop a defense chatbot.

Regulatory challenges have also influenced Meta’s AI plans. In 2024, the company paused AI model training using user data in the European Union and the United Kingdom following legal pushback. It also temporarily halted the rollout of some AI systems in Europe due to ongoing regulatory uncertainties.

As AI governance continues to be a global concern, Meta’s risk assessment framework is expected to shape discussions on responsible AI development and deployment in the tech industry.