Historical GPTOnline editorial cover for the September 2023 Red Teaming Network announcement; it is not a current application notice or evidence of a completed safety evaluation.
On September 19, 2023, OpenAI announced the OpenAI Red Teaming Network to inform its AI risk-assessment and mitigation work. The initiative sought a community of outside experts rather than only one-off engagements before model releases.
Red teaming is emerging as a crucial step in the process of AI model development, especially as generative technologies gain widespread acceptance. It serves as a mechanism to identify, although not necessarily rectify, biases in models such as DALL·E 2, known to magnify stereotypes related to race and gender. It also identifies prompts causing text-generating models, including ChatGPT and GPT-4, to overlook safety filters.
OpenAI had previously worked with external experts to assess model risks. The Red Teaming Network formalized that collaboration and complemented its Researcher Access Program. Its separate security bug bounty focused on vulnerabilities in systems; it should not be treated as the same programme as behavioural model-risk assessment. The objective, as stated in a company blog post, is to deepen and broaden OpenAI’s collaboration with scientists, research institutions, and civil society organizations.
OpenAI envisions this initiative as a supplement to externally-specified governance practices, like third-party audits. Members of the network, based on their expertise, will be invited to participate in red teaming at various stages of the model and product development lifecycle.
In addition to commissioned red teaming campaigns by OpenAI, members of the Red Teaming Network will have the opportunity to interact with each other about general red teaming practices and insights. OpenAI clarifies that not every member will be involved with every new model or product, and individual time contributions, which could be as few as 5–10 hours in one year, will be determined individually.
The 2023 call invited a diverse range of domain experts to join, including those with backgrounds in linguistics, biometrics, finance, and healthcare. Prior experience with AI systems or language models is not a prerequisite for eligibility. OpenAI said membership itself would not restrict publishing research or pursuing other opportunities, but work on particular projects could be covered by NDAs or remain confidential indefinitely.
The company values a willingness to engage and contribute unique perspectives on the impact assessment of AI systems. The original call welcomed applicants globally and prioritized geographic and domain diversity. A later editor’s note on the same announcement states that applications for that phase closed on December 1, 2023; this is not a current open invitation.
However, doubts remain if red teaming alone is sufficient. Critics argue it might not be.
In a March 2023 opinion essay, Aviv Ovadya advocated “violet teaming.” This involves identifying potential harm a system like GPT-4 might cause to an institution or public good, and subsequently supporting the development of tools using that same system to protect the institution and public good. He argued that incentives and time for such work were limited. This was his proposed extension to risk assessment, not an established guarantee of safer outcomes.
Red teaming can reveal problems, but identifying them does not automatically fix them or prove that a system is safe. The network was presented as one part of a broader evaluation and governance process, not a proven best solution.
