Safety is at the Core of Our Work

Responsibly developing and deploying advanced AI systems

We believe that powerful artificial intelligence must be safe. We invest significant resources in safety research, risk assessment, and the development of safeguards to ensure AI benefits all humanity.

Safety Principles

Prevent Harm

We actively identify and mitigate potential risks, including misuse, unintended consequences, and systemic risks. Before releasing new models, we conduct rigorous safety assessments.

Human Oversight

We introduce human feedback (RLHF) during training to ensure model behavior aligns with human intentions and values, reducing bias and harmful outputs.

Continuous Monitoring

We continuously monitor system usage after deployment, quickly respond to emerging threats, and continuously update models to improve security.

Our Safety Measures

We embed safety practices at every stage of the AI development cycle.

Red Teaming

We invite external experts ("red teams") to attack our models, trying to induce them to produce harmful content, bias, or misinformation. This helps us discover and fix vulnerabilities before release. For example, before GPT-4 release, we invited over 50 experts from AI safety, cybersecurity, biological risk and other fields to conduct testing.

Reinforcement Learning (RLHF)

We use reinforcement learning technology based on human feedback to fine-tune models. Human trainers rank and score model responses, guiding the model to generate more useful, truthful, and harmless replies. This is key to ChatGPT's ability to have fluent conversations while maintaining safety.

Content Moderation API

We developed a free content moderation API to help developers identify and filter inappropriate content such as hate speech, self-harm, violence, etc. This is also integrated into ChatGPT to prevent generating content that violates our usage policies.

Collaboration & Governance

We collaborate with governments, civil society, and other AI labs to develop AI safety standards and best practices. We support regulation of high-risk AI systems and are committed to improving transparency in AI development.

Your Data Privacy

We understand the importance of data privacy. When you use ChatGPT, you have control over your data.

Your data is not used for training (optional): Enterprise users' data is not used for model training by default. Individual users can also opt out of data training in settings.
Data retention control: We only retain deleted conversation records for review for security purposes (up to 30 days).
SOC 2 Compliance: Our security measures have passed SOC 2 Type 2 compliance audit, ensuring enterprise-level security standards.