Strengthening our Frontier Safety Framework

We expand our risk ranges and improve our risk evaluation.

Artificial intelligence breakthroughs transform our daily lives, from the progress of mathematics, biology and astronomy to achieving the potential of personal education. While we are increasingly powerful AI models, we are committed to developing our technologies responsible and taking evidence based on evidence to remain at the forefront of the emerging risk.

Today, we publish the third repetition of our border safety framework (FSF) – our most comprehensive approach to determine and alleviate the severe risks of advanced artificial intelligence models.

This update depends on our continuous cooperation with experts through industry, academic circles and government. We also combined the lessons learned from the implementation of previous versions and developed best practices in the AI ​​border safety.

The main updates of the framework

Treating the risk of harmful manipulation

Through this update, we offer the critical level of ability (CCL)* Focus on harmful manipulation – specifically, artificial intelligence models with strong manipulation capabilities that have been questioned to change beliefs and behaviors systematically and greatly in the harsh limits.

This addition depends on the research and its operation that we conducted to identify and evaluate the mechanisms that drive manipulation of artificial intelligence. To move forward, we will continue to invest in this field to understand and measure the risks associated with harmful manipulation.

Adapting our approach to the risk of imbalance

We have also expanded our working framework to address potential future scenarios, as unprepared artificial intelligence models may overlap with the ability of operators to direct, modify, or stop their operations.

While our previous copy of the frame included an exploratory approach focusing on useful logical CCLS (i.e. warning levels of the date of the start of the artificial intelligence model in a deceptive manner), with this update, we now provide more protocols for our research in machine learning and the development of CCLS that focuses on models that can accelerate and develop artificial intelligence research.

In addition to the risks of misuse caused by these capabilities, there are also the risk of an imbalance resulting from the capabilities of the model for an unveiled work in these levels of abilities, and the potential integration of these models in artificial intelligence development and publishing processes.

To treat the risk forms by CCLS, we perform safety cases before external launch processes upon release CCLS. This includes detailed analyzes showing how to reduce the risks to controlled levels. For advanced automatic learning research and CCLS development, internal publishing can be widespread risks, so we are now expanding this approach to include these bulletins.

Steps our risk evaluation process

Our working framework is designed to process risk in proportion to its intensity. Our CCL definitions have specifically charged the critical threats that call for the most striking governance and mitigation strategies. We continue to apply safety and security before reaching specific CCL thresholds and as part of our standard model development approach.

Finally, in this update, we go to more details about our risk assessment. Based on our basic evaluation early, we half how comprehensive assessments include identifying methodological risks, comprehensive analyzes of model capabilities and explicit determination of risk acceptance.

The promotion of our integrity with the integrity of the border

This last update represents the border safety framework, our continuous commitment to taking a scientific approach to the evidence to track and remain the risks of artificial intelligence as capabilities advance towards AGI. By expanding our risk areas and enhancing our risk assessments, we aim to ensure that transformational artificial intelligence benefits humanity, while reducing possible damage.

Our framework will continue to develop based on a new research and the inputs of stakeholders and lessons of implementation. We are still committed to working cooperatively through industry, academic circles and government.

Not only does the use of useful AGI require technical breakthroughs, but also requires strong work frameworks to relieve risks along the way. We hope that our updated border safety framework will contribute to this collective effort.