How did the government decide OpenAI’s frontier model was safe to release?

OpenAI has begun the public rollout of its latest language model, named Sol, which is said to rival Anthropic’s Fable. The approval process for these advanced models has raised eyebrows, as many experts are left questioning how they were deemed safe for release. Mina Narayanan, a senior research analyst at Georgetown’s Center for Security and Emerging Technology, expressed uncertainty about the review processes, stating, “I don’t feel like I have enough information to say whether they’re adequate or not.” While Anthropic has mentioned engaging in discussions with government officials and developing a classifier to detect unauthorized access attempts, the specifics of those conversations remain elusive. Dean W. Ball, a former advisor to the Trump administration and now part of OpenAI, highlighted the lack of clarity surrounding the licensing requirements for AI models. Andy Konwinski, a computer scientist with a background in multiple startups, echoed this sentiment, noting that even individuals within frontier labs are unclear about the decision-making process. This ambiguity fuels concerns about who ultimately holds the power to approve AI technologies. Despite the confusion, the Biden administration recently issued an executive order outlining a plan for evaluating frontier models. However, the details of this approach are still vague, with many experts questioning which models will require government oversight and which agencies will be responsible for these evaluations. Currently, the Department of Commerce’s Center for AI Standards and Innovation appears to be taking a leading role, but the timeline remains uncertain. OpenAI CEO Sam Altman mentioned that discussions with government officials were part of the approval process, but details about the experts involved in testing the models are scarce. OpenAI has refrained from disclosing specific information about the government’s evaluation processes, although they referenced various external assessments that contributed to the model's safety designation. In a blog post from late June, OpenAI indicated their intention to collaborate with the government to establish a more transparent access process, distancing themselves from what they described as a temporary measure. Complicating the situation is the reported offer of up to 5% equity in OpenAI as part of negotiations with the administration, raising questions about the influence of personal connections on regulatory decisions. In contrast, Anthropic’s Fable faced a temporary suspension from public access due to concerns over its potential misuse and conflicts with the Trump administration, which highlights the complexities of regulatory dynamics in the AI sector. The potential for export bans may also have influenced OpenAI's willingness to cooperate with government requests. Industry experts warn that a hands-off regulatory approach, while appealing, could lead to uncertainty and misaligned incentives. Konwinski argues that a more inclusive approach, involving safety and alignment researchers, is crucial for responsibly managing AI advancements. He advocates for the establishment of independent auditing organizations to provide oversight of frontier labs. The ongoing secrecy surrounding AI development is likely to persist, presenting ongoing challenges for an industry facing growing public skepticism. As concerns about the concentration of power among a few firms rise, experts are urging for greater transparency and public engagement in the evaluation of AI technologies, fearing a future where technology is controlled away from public scrutiny.