OpenAI Details 10 Safety Practices at Seoul Summit

OpenAI used the AI Seoul Summit as a backdrop to pull the curtain back on the specific safety practices it claims make its models industry-leading, releasing a list of ten measures as it signs onto new Frontier AI Safety Commitments. The timing is no accident. With over a hundred million users now relying on its systems, the company is trying to frame safety not as a bolt-on feature but as a core engineering discipline that spans multiple time horizons, from today’s chatbots to the far more capable systems they expect to build tomorrow.

The commitments, agreed upon by OpenAI and other unnamed companies, push for public safety frameworks—something OpenAI already did last year with its Preparedness Framework. That document sets a hard rule: the company won’t release a new model if it crosses a “Medium” risk threshold until sufficient safety interventions bring the post-mitigation score back down. For GPT‑4o, more than 70 external red-teamers kicked the tires, and their findings were fed back into evaluations designed to probe weaknesses in earlier checkpoints. It’s a rare glimpse into the iterative, unglamorous work that happens before a model ever sees the light of day.

On the abuse detection front, the company is eating its own dog food. It uses GPT‑4 for content policy development and moderation decisions, which it claims creates a faster feedback loop for policy refinement while exposing human moderators to less toxic material. The company also pointed to a joint disclosure with Microsoft about state actor abuse of its technology as proof it’s willing to share uncomfortable findings so others can harden their defenses. Meanwhile, child safety gets specific attention: a partnership with Thorn’s Safer initiative detects and reports Child Sexual Abuse Material to the National Center for Missing and Exploited Children if users try uploading it to image tools.

The election integrity measures are more concrete than what most labs disclose. ChatGPT now points U.S. and European users to official voting information sources. DALL·E 3 images carry C2PA metadata to help people trace media provenance, and the company joined the steering committee of the Content Authenticity Initiative. It’s a recognition that safety in 2024 isn’t just about what a model refuses to say—it’s about the downstream chaos that synthetic media can cause when provenance is a mystery.