OpenAI Reportedly Discusses AI Safety With Anthropic And Google DeepMind
OpenAI AI Safety Talks: What Would Actually Change?
AI Rivals Are Reportedly Discussing Safety, But The Test Is What Any Agreement Would Require Them To Change.
OpenAI is discussing AI safety with Anthropic and Google DeepMind, according to Bloomberg reporting relayed by Reuters on 15 September. The report cites OpenAI policy executive Chris Lehane. It describes talks, rather than a published, binding agreement between the three companies.
The distinction is consequential. Companies can agree that a problem matters while disagreeing over acceptable risk, independent scrutiny and whether a commercial launch should wait. A useful reading of this development therefore separates the reported contact from the rules that might eventually follow.
Shared Tests Would Only Be A Starting Point
A common evaluation could make competing systems easier to compare. However, a test result becomes a safeguard only when someone has to respond to it. A proposed arrangement would need to explain what constitutes failure, who reviews an ambiguous result and whether additional deployment is suspended during an investigation.
Those are questions for any eventual agreement, not terms established by the reporting. The voluntary NIST AI Risk Management Framework offers an existing reference point: responsibility and ongoing risk management extend across a system’s life, rather than ending with a demonstration. NIST framework.
Who Would Be Able To Challenge The Companies?
Independent access would matter most when an evaluator reaches an unwelcome conclusion. A credible arrangement could give reviewers access to the relevant version, preserve records and require a written response to significant findings. Without those features, the public would struggle to distinguish a demanding review from a selective presentation.
There is also a legitimate design tension. Publishing every technical detail could expose sensitive information, while excessive secrecy could prevent outside scrutiny. A workable proposal would need separate channels for public findings and confidential technical evidence.
The Next Meaningful Development
Look for a named mechanism, defined obligations and an explanation of enforcement. A joint statement would be more informative if it identified who can require corrective action and how disputes are resolved. None of those features should be assumed simply because discussions are reported.
Taylor Tailored’s analysis of what would make an AI slowdown plan work examines that gap between intention and constraint. Cooperation could be valuable; its importance will depend on whether it changes decisions when safety and commercial incentives pull in different directions.
