OpenAI moves third-party safety assessments to earlier stages of model development
BIBIBI
AT A GLANCE
OpenAI will involve more third-party organizations in early safety evaluations of AI models.
Article
OpenAI expands third-party involvement in early-stage AI model safety assessments
PANews, September 23: According to Bloomberg, OpenAI announced that it will allow more third-party organizations to participate in technical safety assessments during the early stages of AI model training, evaluation, and deployment. Previously, the company generally brought in external teams to assess capabilities and risks only before a model's release. OpenAI stated that such assessments must demonstrate independence, scientific rigor, robust safety mechanisms, and clear allocation of responsibilities. OpenAI is in discussions with both new and existing partners, including METR and Redwood Research. Some evaluators may in the future be assigned to work at the company's internal offices and participate in testing of the most sensitive components.
OpenAI announced that it will allow more third-party organizations to participate in technical safety assessments during the early stages of AI model training, evaluation, and deployment.
02
Previously, OpenAI generally brought in external teams to assess capabilities and risks only before a model's release.
03
OpenAI stated that evaluations must be independent, scientifically rigorous, supported by strong safety mechanisms, and have clearly defined responsibilities.
04
OpenAI is in discussions with multiple new and existing partners, including METR and Redwood Research.
05
Some evaluators may in the future be placed in OpenAI’s internal offices to participate in testing of the most sensitive components.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Put simply, the point at which external organizations participate in safety checks may expand from before a model's release to the early stages of training, evaluation, and deployment.
Why it matters to readers
This shows that the scope and timing of external safety evaluations of AI models are expanding.
Readers can focus on whether evaluations are independent and scientifically rigorous, and whether responsibilities are clearly assigned.
Risks and unknowns
The report does not specify the final list of additional third-party organizations.
The report does not state whether the discussions have resulted in a formal partnership.
Arrangements for potentially entering internal office premises in the future remain undecided.
Related Developments
Loading event timeline…
Related concepts
AI models
This term is not in the glossary yet. Browse related concepts in the glossary.
