OpenAI, the creator of Chat-GPT, has recently introduced a new framework called the ‘Preparedness Framework‘ aiming to address the growing concerns about the safety of advanced AI models. This framework includes a range of tools and processes that are intended to enhance AI safety in the development of artificial intelligence models.
An Innovative Approach to AI Safety
At the heart of OpenAI’s new initiative is the Preparedness team, dedicated to ensuring the safety of frontier AI models. This team collaborates closely with other divisions, including the Safety Systems team, which focuses on current model misuse, and Superalignment, which is building safety foundations for future superintelligent AI. The Preparedness Framework is a science-driven, fact-based approach that aims to forecast and mitigate emerging risks effectively.
OpenAI prioritizes safety in AI development, utilizing real-world deployments to enhance safety measures. The Preparedness Framework (Beta) introduces a new methodology for secure AI model development and deployment. Some key elements of the framework include:
- Evaluations and Scorecards: OpenAI will conduct thorough evaluations of its frontier models, including during significant increases in computational power. These assessments will help identify and mitigate risks, with ongoing updates to model risk “scorecards.”
- Risk Thresholds and Safety Measures: The company has established risk thresholds across categories like cybersecurity and model autonomy. Models must meet specific safety criteria before deployment or further development, ensuring a rigorous safety standard.
- Dedicated Oversight and Advisory Groups: A dedicated team will oversee technical assessments, while a cross-functional Safety Advisory Group will review findings and provide input to both the company’s leadership and its Board of Directors.
- Protocols for External Accountability: Regular safety drills and independent audits will be part of the framework, highlighting OpenAI’s commitment to external feedback and accountability.
OpenAI is taking a comprehensive approach to ensure the safety and responsible use of its AI technology. To achieve this, the company is collaborating with both internal and external teams, including Safety Systems and Superalignment, to track and mitigate any instances of real-world misuse and emerging risks. OpenAI is also leading research into risk evolution as AI models scale, with the aim of proactively forecasting potential risks.