Unpacking the Government's OpenAI Safety Decision

OpenAI is making its cutting-edge large language model, Sol, broadly available to the public. Sol is reportedly comparable in capability to Anthropic's Fable, a model that previously caused sufficient concern within the White House—due to either its advanced features or its ownership—to warrant a temporary restriction from public access.

The question of how these powerful models received authorization for release, however, remains largely unanswered.

Mina Narayanan, a senior research analyst at Georgetown’s Center for Security and Open Technology, expressed this uncertainty to TechCrunch, stating, “Frankly, I don’t have visibility into those exact processes, so yes, I don’t feel like I have enough information to say whether they’re adequate or not.” She noted that Anthropic confirmed discussions with the government, along with developing a classifier to identify jailbreak attempts and implementing defensive strategies against future breaches. Nonetheless, the precise nature of these dialogues between the government, Anthropic, and OpenAI remains obscure.

This sentiment was echoed last month by Dean W. Ball, a former Trump policy advisor now employed by OpenAI, who wrote in his newsletter that “nobody knows what the requirements are to get licensed.”

Andy Konwinski, a computer scientist and co-founder of Databricks, Perplexity, and the Laude Institute, informed TechCrunch that he has yet to encounter anyone, even within leading frontier AI laboratories, who fully grasps this approval process. He emphasized the profound implications, stating, “It’s existentially a problem.” Konwinski further highlighted the core issue: “Safety or not, it’s about who has the power to make decisions—who gatekeeps and decides on permissions?”

Despite being eighteen months into the Trump administration, a clear path forward remains elusive, a situation some critics attribute to, rather than despite, the influence of industry figures in policy-making. Last month saw the release of an executive order, following weeks of internal debate, which outlined a roadmap for evaluating frontier models. However, concrete specifics are largely absent, save for one definitive exclusion. Sriram Krishnan, a former Andreesen Horowitz partner and until recently a senior AI advisor in the White House, told the Financial Times, “There will not be an FDA for AI.”

A significant unresolved issue is the lack of consensus on which types of models necessitate government oversight and which agency or agencies should conduct these evaluations. Presently, the Department of Commerce’s Center for AI Standards and Innovation appears to be spearheading efforts. Yet, the executive order directs six cabinet agencies to establish a definitive process by early August, leaving the current interim approach largely ad hoc.

OpenAI CEO Sam Altman disclosed on CNBC that the approval process entailed discussions with officials such as Secretary of Commerce Howard Lutnick, Secretary of the Treasury Scott Bessent, and US national cyber director Sean Cairncross. However, the identities of the experts who tested the models, and their methodologies, remain undisclosed. OpenAI declined TechCrunch's request for details on the government’s process, instead referring to the outcomes of various external evaluations by entities like UK AISI, SecureBio, and Irregular, as documented in the model’s safety card.

Preceding its broader public rollout, OpenAI provided the government and a select group of users with a preview of Sol, mirroring the approach taken with Anthropic’s Fable. The identities of these users and the criteria for their selection have not been revealed. In a late June blog post, the company acknowledged, “we don’t believe this kind of government access process should become the long-term default,” committing to collaborate with the government on establishing an alternative pathway.

These discussions, however, occurred against a notable backdrop: reports of Altman offering up to 5% of OpenAI’s equity for the administration’s "Trump Accounts," and OpenAI president Greg Brockman being identified as the largest publicly known donor to Trump’s mid-term political operation. For external observers, it becomes challenging to disentangle these activities from the government’s seemingly more lenient regulatory stance on Sol.

In contrast, Anthropic’s Fable was temporarily withdrawn from widespread access after the US government prohibited its use by foreign nationals. This decision stemmed from genuine concerns regarding users exploiting the model for hacking capabilities through jailbreaking, alongside reported personality clashes between Anthropic and the Trump administration. The potential threat of an export ban may also have incentivized OpenAI to be more compliant with the government's (still undisclosed) requests.

While a hands-off regulatory approach might appeal to the industry, one that appears contingent on personal relationships with administration officials inevitably fosters uncertainty and perverse incentives.

Konwinski conveyed his apprehension to TechCrunch, expressing concern that genuine experts in AI technology—including “safety researchers, alignment researchers, interpretability researchers, but also data people, and people from all over the stack”—are not adequately involved in the model release process.

He advocates for an “open commons” approach as the most effective method to balance safety with innovation. Konwinski referenced established models such as the FDA, NIH, or national laboratories, which facilitate collaboration among researchers, government officials, and private companies to build consensus on critical safety matters.

This dynamic is partly rooted in the capitalist incentives that have driven AI research for over a decade, a factor highlighted in Elon Musk’s lawsuit challenging OpenAI’s corporate structure. Ball underscored that the AI business model necessitates companies rapidly recouping substantial training costs soon after model release to maintain a competitive edge. Konwinski added, “Even if their intentions are good, there’s very clear legal obligations and fiduciary responsibility that are built right into the operating procedures.”

In his post, Ball proposed that the path forward should involve government-licensed, third-party auditing organizations responsible for assessing the safety protocols of frontier laboratories. Konwinski also expressed optimism for novel institutional structures, such as focused research organizations, which could empower impartial experts from academia and the non-profit sector to access and evaluate advanced AI models.

Currently, the pervasive secrecy surrounding AI development shows no signs of abating. This opacity is likely to generate political challenges for an industry already facing growing skepticism from the American public. As Remzi Arpaci-Dusseau, a computer science professor at the University of Wisconsin-Madison, remarked last week at the Open Frontier conference, “There’s not a sense that responsible people are driving forward these changes.”

At the same conference, David Siegel, the computer scientist and founder of Two Sigma—one of the most successful quantitative hedge funds—invited attendees to contemplate a scenario he deemed “very bad”: one where “a small number of firms control the technology; the government, in their secretive laboratories, is evaluating whether or not the technology is suitable for use; and the general public and scientific community doesn’t really have any access to any of that stuff.”

This hypothetical scenario, it appears, may already be unfolding.