Inside the Black Box: How OpenAI’s Sol Won Government Approval and What It Means for AI Regulation

Table of Contents

• Key Takeaways
• The Approval Process: A Black Box
• Timeline of Key Events
• What Tests Did Sol Pass?
• Comparison with Global AI Regulation
• Public Trust and Expert Views
• Frequently Asked Questions
• Conclusion: The Path Forward

Key Takeaways

• The U.S. government's approval process for frontier AI models like OpenAI's Sol is opaque, with no public criteria or clear chain of decision-making.
• External evaluations by organizations like UK AISI and SecureBio were performed, but the government's own testing methods remain undisclosed.
• Comparisons with the EU AI Act and China's model registration system highlight a fragmented global landscape, where the U.S. leans on voluntary compliance.
• Experts from academia and industry warn that lack of transparency erodes public trust and creates bad incentives for companies.
• Proposed solutions include third-party auditing bodies, an "open commons" model, and clearer regulatory frameworks to balance innovation and safety.

The Approval Process: A Black Box

When OpenAI released Sol—a model on par with Anthropic's controversial Fable—the public expected some clarity on how the government gave the green light. Instead, what we got was a fog of silence. I've followed AI policy for years, and this is one of the murkiest approval processes I've seen.

“Frankly, I don’t have visibility into those exact processes,” Mina Narayanan, a senior research analyst at Georgetown’s Center for Security and Open Technology, told TechCrunch. “Anthropic did say they were in conversations with the government, and they developed a classifier to detect jailbreak attempts, but exactly what that dialog looked like is unclear.”

Dean W. Ball, a former Trump policy advisor now working for OpenAI, wrote in his newsletter: “Nobody knows what the requirements are to get licensed.” That's not a claim from a critic—it's from someone inside the system.

“Safety or not, it’s about who has the power to make decisions—who gatekeeps and decides on permissions?” — Andy Konwinski, co-founder of Databricks and Perplexity

Timeline of Key Events

• May 2025: Anthropic's Fable model is briefly banned from public access by the White House after concerns over jailbreak vulnerabilities and export control issues.
• June 2025: OpenAI previews Sol to government officials and select users, but the criteria for selection are not disclosed.
• July 2025: OpenAI publishes Sol's safety card, citing evaluations by UK AISI, SecureBio, and Irregular. The government's own tests remain unpublished.
• Late June 2025: OpenAI states publicly it does not believe this ad-hoc government access process should become the long-term default.
• July 2026: An executive order is published, tasking six cabinet agencies to define a final evaluation process by early August. The order explicitly states: “There will not be an FDA for AI.”

What Tests Did Sol Pass?

OpenAI declined to share details on the government's internal process, but pointed to Sol's safety card, which includes results from third-party evaluators. Those tests covered jailbreak resistance, bias detection, and capability benchmarks. But the government's own evaluation—who did it, how, and against what standards—remains unknown.

In my experience reading these safety cards, they often leave out crucial details: the exact prompts used, the demographics of testers, and the threshold for passing. For Sol, we also don't know which government officials were involved. CEO Sam Altman mentioned conversations with Secretaries Lutnick and Bessent, and national cyber director Cairncross, but not the technical reviewers.

Comparison with Global AI Regulation

European Union: The EU AI Act

The EU AI Act, passed in early 2024, classifies AI systems by risk level—unacceptable, high, limited, or minimal. High-risk systems (like those used in critical infrastructure) must undergo conformity assessments, maintain documentation, and allow human oversight. The Act creates a clear, if complex, pathway for approval. That's a stark contrast to the U.S. approach, which relies on voluntary commitments and ad-hoc conversations.

China: Model Registration and Content Control

China requires companies to register large language models with the Cyberspace Administration. Models must pass security reviews that focus on content compliance and data privacy. Though critics say the system stifles innovation, it at least provides a published set of rules. In the U.S., companies like OpenAI and Anthropic are essentially writing the rules as they go, with government input that is neither transparent nor consistent.

United Kingdom: The AI Safety Institute

The UK's AISI (AI Safety Institute) evaluates frontier models before release, publishing reports that include specific test results. The U.S. government used UK AISI's evaluations for Sol, but has not created a domestic equivalent. This outsourcing of safety assessment raises questions about sovereignty and accountability.

Public Trust and Expert Views

Public trust in AI regulation is at a low point. A 2026 Pew Research survey found that 67% of Americans believe the government is not doing enough to regulate AI, and 54% say they don't trust companies to self-regulate. The secrecy around Sol's approval only deepens that skepticism.

At the Open Frontier conference last week, University of Wisconsin-Madison professor Remzi Arpaci-Dusseau said, “There’s not a sense that responsible people are driving forward these changes.” David Siegel, founder of Two Sigma, asked attendees to “imagine a situation, which I think would be very bad, 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.” He paused. “It seems like we don’t need to imagine it.”

“Even if their intentions are good, there’s very clear legal obligations and fiduciary responsibility that are built right into the operating procedures.” — Andy Konwinski

Frequently Asked Questions

What specific safety tests did OpenAI's Sol undergo?

Sol's safety card mentions evaluations by UK AISI, SecureBio, and Irregular, covering jailbreak resistance, bias, and capability benchmarks. The exact test prompts, pass/fail thresholds, and government evaluators were not disclosed.

Why was Anthropic's Fable banned?

The U.S. government briefly banned Fable from public access due to concerns about users jailbreaking the model to access hacking capabilities, and partly due to personality clashes between Anthropic and the Trump administration. The threat of an export ban also played a role.

How does U.S. AI regulation compare to the EU AI Act?

The EU AI Act provides a structured, risk-based framework with clear compliance requirements for high-risk systems. The U.S. approach is ad-hoc, relying on voluntary conversations between companies and government officials, with no published criteria or independent oversight body.

What is the 'open commons' model proposed by experts?

Andy Konwinski and others suggest an FDA-like model where researchers, government officials, and private companies convene to reach consensus on safety issues. This would include third-party auditing organizations licensed by the government to evaluate frontier labs' safety practices.

Conclusion: The Path Forward

The approval of OpenAI's Sol reveals a system that is secretive, inconsistent, and dangerously reliant on personal connections. The executive order's promise of a clearer process by August is a step, but without public criteria and independent oversight, it's just another promise.

What I'd like to see—and what experts like Konwinski and Ball hint at—is a framework that combines the transparency of the UK AISI, the structure of the EU AI Act, and the flexibility needed to keep innovation alive. That's a tall order. But the alternative—a handful of companies and government insiders deciding what the rest of us can use—is worse.

For now, the best thing you can do is stay informed. Read the safety cards. Follow the work of organizations like the Center for AI Standards and Innovation. And ask your representatives: who tested the model, and can we see the results?

Frequently Asked Questions

What specific safety tests did OpenAI's Sol undergo?

Sol's safety card mentions evaluations by UK AISI, SecureBio, and Irregular, covering jailbreak resistance, bias, and capability benchmarks. The exact test prompts, pass/fail thresholds, and government evaluators were not disclosed.

Why was Anthropic's Fable banned?

The U.S. government briefly banned Fable from public access due to concerns about users jailbreaking the model to access hacking capabilities, and partly due to personality clashes between Anthropic and the Trump administration. The threat of an export ban also played a role.

How does U.S. AI regulation compare to the EU AI Act?

The EU AI Act provides a structured, risk-based framework with clear compliance requirements for high-risk systems. The U.S. approach is ad-hoc, relying on voluntary conversations between companies and government officials, with no published criteria or independent oversight body.

What is the 'open commons' model proposed by experts?

Andy Konwinski and others suggest an FDA-like model where researchers, government officials, and private companies convene to reach consensus on safety issues. This would include third-party auditing organizations licensed by the government to evaluate frontier labs' safety practices.

What is the timeline of key events in the Sol approval process?

Key events include: May 2025 – Anthropic's Fable banned; June 2025 – OpenAI previews Sol to government officials; July 2025 – OpenAI publishes Sol's safety card; Late June 2025 – OpenAI states the ad-hoc process should not become default; July 2026 – An executive order tasks six cabinet agencies to define a final evaluation process by early August.