With more than a hundred million users and millions of developers relying on our safety teams' work, we view safety as a critical investment that spans multiple time horizons. From aligning today’s models to anticipating the needs of more capable future systems, our commitment to safety grows stronger over time.
We believe in a balanced, scientific approach where safety measures are integrated into the development process from the outset, ensuring our AI systems are both innovative and reliable.
AI Seoul Summit and Frontier AI Safety Commitments
At today’s AI Seoul Summit, we joined industry leaders, government officials, and civil society members to discuss AI safety. OpenAI and other companies agreed on the Frontier AI Safety Commitments, pledging to safely develop and deploy frontier AI models while sharing information about their risk mitigation measures. These commitments align with steps we have already taken, including publishing safety frameworks like our Preparedness Framework.
Related:
Our 10 Active Safety Practices
-
Empirical Model Red-Teaming and Testing: We evaluate model safety before release, both internally and externally, according to our Preparedness Framework. No model is released if it crosses a “Medium” risk threshold without sufficient safety interventions.
-
Alignment and Safety Research: Over time, our models have become significantly safer due to smarter design and focused investment in alignment and safety systems. We conduct and publish fundamental research to enhance our models' robustness against attacks.
-
Monitoring for Abuse: We deploy a range of tools, including dedicated moderation models and our own models for safety risk monitoring. We have shared critical findings, such as state actor abuse of our technology, to help others safeguard against similar risks.
-
Systematic Safety Approach: Safety measures are implemented at every stage of the model's life cycle, from pre-training to deployment, ensuring continuous improvement and robust monitoring infrastructure.
-
Protecting Children: We prioritize protecting children by building strong default guardrails into ChatGPT and DALL·E, partnering with organizations like Thorn to detect and report Child Sexual Abuse Material, and upholding Safety by Design principles.
-
Election Integrity: We collaborate with governments to prevent AI abuse, ensure transparency, and provide access to accurate voting information. Initiatives include identifying DALL·E 3-generated images and supporting the "Protect Elections from Deceptive AI Act".
-
Impact Assessment and Policy Analysis: Our impact assessments influence research, industry norms, and policy. We publish work on managing risks and collaborate with experts to assess the implications of language models.
-
Security and Access Control: Protecting customers, intellectual property, and data is a priority. We control access via API, conduct penetration testing, and offer a bug bounty program. We explore novel security measures like confidential computing for GPUs.
-
Government Partnerships: We partner with governments worldwide to develop adaptable AI safety policies, share learnings, and inform public debate over new standards and laws.
-
Safety Decision Making and Board Oversight: Our Safety Advisory Group reviews model capability reports, with company leadership making final decisions and the Board of Directors overseeing these decisions.
Future Directions
As we advance towards our next frontier model, we recognize the need to evolve our practices, especially in increasing our security posture against sophisticated state actor attacks and ensuring thorough safety testing before major launches. We will continue to share updates on these evolving practices in the coming weeks.
Conclusion
OpenAI is dedicated to developing safe and capable AI models. By integrating comprehensive safety practices and collaborating with various stakeholders, we strive to deliver AI that benefits society while mitigating potential risks.
Have a question? Ask now!
Follow us on Instagram!
