OpenAI plans to invite independent reviewers into much earlier parts of its development cycle so risks get surfaced before a model is close to release. The company intends to let outside organizations perform technical safety assessments while models are being trained, evaluated and rolled out, not only right before launch.
In the upcoming blog post, OpenAI says effective reviews hinge on "strong independence mechanisms, scientific rigor, robust security practices and clear responsibilities." The post adds, "Labs have a responsibility to enable meaningful scrutiny while protecting sensitive information."
Lama Ahmad, who leads much of OpenAI's work with external experts on safety reviews, said the company has typically brought in these groups before launch and to assess capabilities. "As the stakes get higher, we want to make sure we're also looking at things like training and evaluation, which do have high stakes, in addition to our deployments," Ahmad said in an interview.
Employees at major AI labs have recently voiced concerns about catastrophic risks. Those worries have grown after several incidents in recent months when advanced models from OpenAI and other developers inadvertently breached other firms during testing.
Ahmad said the effort may include inviting evaluators into OpenAI's offices for the most sensitive work, noting the company has tried this approach before.
New ideas can be exciting, but steady planning keeps your money safe. Join Briefs Finance CEO Jaspreet Singh on September 29th for a FREE live investor workshop, How to Profit From A Dollar That's Losing its Value, where he shows how we're spotting investment opportunities as the dollar falls. Save your spot.
On Monday, OpenAI called on the US to spearhead work alongside other countries to establish standards for cutting-edge artificial intelligence technology.
OpenAI is discussing the work with a mix of groups it has and has not partnered with before. Ahmad said those talks include AI research outfits METR and Redwood Research, which OpenAI previously asked to investigate its models' breach of Hugging Face.
In parallel, Anthropic PBC CEO Dario Amodei recently urged the industry to support slowing development and said Anthropic would adopt new safety measures, including third-party evaluators. Later on, OpenAI CEO Sam Altman said he was in agreement. In the latter part of last week, Anthropic announced plans to bring in evaluators from Accenture Plc to assess the safety of its most advanced models.
If OpenAI follows through, earlier outside scrutiny could make high-profile AI releases less chaotic and more predictable. Fixing issues during training and evaluation is typically cheaper than firefighting after deployment, which matters for companies building on AI and for anyone exposed to the sector's growth story.
Real progress comes from patience and discipline that protect and grow savings. Our CEO Jaspreet Singh is hosting a FREE live investor workshop, How to Profit From A Dollar That's Losing its Value, on September 29th. Sign up free to join him live.
