How did the government decide OpenAI’s frontier model was safe to release?
Key Takeaways
- OpenAI has released its newest large language model, Sol, which it positions as at least comparable to Anthropic’s Fable, a model that was temporarily restricted from public access after drawing White House concern, according to TechCrunch.
- TechCrunch reports that researchers, policy advisors, and even employees at frontier AI labs say they lack visibility into how the U.S. government actually evaluates and approves these models before release.
- An executive order published last month, following weeks of internal disagreement, sets a roadmap for evaluating frontier models but leaves key details unresolved, with six cabinet agencies tasked to determine a final process by early August, per TechCrunch.
A Release Process Nobody Can Fully Explain
When OpenAI made Sol widely available, it did so as a model the company itself frames as on par with Anthropic’s Fable. That comparison matters because Fable had a rockier path to the public: it was briefly pulled from wider access after the U.S. government barred its use by foreign nationals, according to TechCrunch, a decision tied partly to genuine worries about users jailbreaking the model for hacking purposes and partly to friction between Anthropic and the Trump administration.
Yet when TechCrunch asked how either model actually cleared government review, the answers were strikingly thin. Mina Narayanan, a senior research analyst at Georgetown’s Center for Security and Emerging Technology, told TechCrunch she does not have visibility into the exact processes involved and cannot say whether they are adequate. She noted that Anthropic has said it was in conversations with the government, built a classifier to catch jailbreak attempts, and adopted defense-in-depth strategies against future exploits — but the substance of the government’s dialogue with Anthropic and OpenAI remains unclear, per TechCrunch.
That uncertainty is not confined to outside observers. Dean W. Ball, a former Trump policy advisor now at OpenAI, wrote in his newsletter last month that nobody knows what the requirements are to get licensed, according to TechCrunch. Andy Konwinski, a computer scientist who co-founded Databricks, Perplexity, and the Laude Institute, told TechCrunch he has never spoken to anyone who understands the process — not even employees inside frontier labs. He framed this as more than a technical gap, telling TechCrunch it is existentially a problem centered on who holds the power to gatekeep and grant permissions.
Why the Ambiguity Matters for the AI Economy
Eighteen months into the Trump administration, TechCrunch reports there is still no consensus on which models require government scrutiny or which agency should conduct evaluations. The Department of Commerce’s Center for AI Standards and Innovation appears to be leading for now, but the executive order gives six cabinet agencies until early August to settle on a final process. Sriram Krishnan, a former Andreessen Horowitz partner who served as a senior White House AI advisor until last month, told the Financial Times, as cited by TechCrunch, that there will not be an FDA-style body for AI.
OpenAI CEO Sam Altman told CNBC, per TechCrunch, that the run-up to Sol’s release involved conversations with officials including Secretary of Commerce Howard Lutnick, Secretary of the Treasury Scott Bessent, and U.S. national cyber director Sean Cairncross. But TechCrunch notes it remains unclear who actually tested the model or how those evaluations were conducted. OpenAI declined to detail the government’s process to TechCrunch, instead pointing to external evaluations referenced in Sol’s safety card from groups including the U.K. AI Security Institute, SecureBio, and Irregular. OpenAI also previewed the model for government officials and select users before wider release, echoing Anthropic’s approach with Fable, though TechCrunch reports the identities and selection criteria for those users were not disclosed. In a blog post at the end of June, OpenAI stated it does not want this kind of government access arrangement to become permanent and said it intends to work with the government on an alternative path, according to TechCrunch.
The context around these conversations includes financial and political ties that TechCrunch says are hard to separate from the government’s apparently lighter regulatory touch: Altman reportedly offered as much as 5% of OpenAI’s equity toward the administration’s so-called Trump Accounts, while OpenAI president Greg Brockman is described by TechCrunch as the largest publicly known donor to Trump’s mid-term political operation. For anyone tracking capital flows into AI — including crypto-adjacent investors who treat frontier-model access and compute deals as market signals — this opacity around licensing and approval raises the stakes of policy risk baked into valuations across the sector.
What Comes Next for Oversight
Konwinski told TechCrunch he worries that genuine experts — safety researchers, alignment researchers, interpretability specialists, and data practitioners — are not sufficiently involved in release decisions. He argued for an “open commons” model resembling institutions like the FDA, the NIH, or national laboratories, which bring together researchers, government officials, and private firms to reach consensus on safety. Ball separately proposed that progress depends on third-party auditing organizations, licensed by government, to assess frontier labs’ safety practices, while Konwinski pointed to focused research organizations as a way to give outside academic and nonprofit experts real access to evaluate these models, per TechCrunch.
Underlying all of this, according to TechCrunch, are commercial pressures that shape how labs behave: Konwinski noted that legal obligations and fiduciary duties are built into how these companies operate, even when intentions are good, given that firms need to recoup massive training costs quickly while they hold a competitive edge. At the Open Frontier conference, University of Wisconsin-Madison professor Remzi Arpaci-Dusseau said there isn’t a sense that responsible people are steering these changes, and Two Sigma founder David Siegel warned attendees to imagine a scenario where a handful of firms control the technology, government evaluates it in secret, and the public and scientific community are shut out entirely — remarks TechCrunch reported directly from the event.
Hype Check
Claim: The U.S. government has a rigorous, transparent process for vetting frontier AI models like OpenAI’s Sol before public release. Reality: According to TechCrunch, researchers, policy insiders, and even lab employees say they cannot describe how these evaluations actually work, no agency or standard has been finalized, and the executive order meant to formalize the process still leaves core questions — including who tests the models and against what criteria — unanswered until at least early August. Verdict: Mostly Hype. This is not financial advice.
Source
Researched with AI assistance, fact-checked and edited by a human. Not financial advice.