OpenAI is building a more structured bench of outside experts to stress-test its AI models before they hit the market. The new Red Teaming Network is a formalization of the company’s earlier, more ad-hoc collaborations with academics and civil society groups on models like DALL-E 2 and GPT-4. Instead of scrambling to find specialists right before a major launch, OpenAI now wants a standing community of trusted testers it can tap at various points in the development lifecycle.
This isn’t a full-time gig. The company says members might be called upon for as little as 5 to 10 hours of work in a year, selected for specific projects based on their domain expertise. That expertise doesn’t necessarily have to be in AI. OpenAI is explicitly looking for cognitive scientists, economists, legal experts, and healthcare professionals, among others — valuing their real-world perspective over a background in tweaking language models. The application window for this first phase closed on December 1, 2023, but the company hinted at future rounds.
The shift signals a clear recognition that internal adversarial testing has its limits. A single team inside a company can’t anticipate how a model might hallucinate in a medical context or be abused by a sophisticated financial scammer. By broadening the pool and making it continuous, OpenAI hopes to move red teaming from a one-time audit into a more iterative, feedback-rich process. The company frames this as a complement to third-party audits, not a replacement.
But there’s a catch. All work is done under a non-disclosure agreement (NDA), which could keep some findings under wraps indefinitely. It’s a familiar tension: the work helps make a commercial product safer, but the confidentiality requirements might prevent researchers from publishing their insights independently. Compensation is offered for time spent on projects, which at least acknowledges that this is skilled labor, not just an opportunity to play with unreleased tech. Whether this network genuinely shapes safety policies or simply provides a more efficient way to gather external criticism right before launch remains an open question.
