AI is Hot - AI Regulation is a Hot Mess

**Ray Rike** (0:08)
Welcome to today's AI to ROI Big Story. I'm Ray Rike, CEO of Benchmarkit.

**Peter Buchanan** (0:16)
And I am Peter Buchanan, the managing partner at NewPlan.

**Ray Rike** (0:20)
You know, I feel I could say this every week, Peter, but we had some really big news in the last week. And this news was from the federal government on AI because on June 12th, the US Department of Commerce issued a directive ordering Anthropic to suspend, just totally suspend access to their latest models, Fable 5 and Mythos 5
And they're the two most powerful AI models in the world. And they wanted every non-US citizen not to have access. Now, Anthropic could not verify users' nationalities and quickly, so it just shut down access to both models, to every user globally, both in the US and beyond the US. So as a result, API calls that were built upon these models returned errors, even active sessions ended, and Cloud Code and Cloud.AI automatically defaulted to older models. The directive was pretty strongly worded, Peter. In fact, it carried criminal penalties for non-compliance, and it really didn't issue any justification for this immediate kind of season.

**Peter Buchanan** (1:36)
Right, that was the only specific thing.
You're going to do this or, you know, there's a criminal compliance thing, nothing else was very specific in it at all. So let's put this in context. It was the first time in history that the US government used its export controls to take a widely deployed frontier model offline. So there's no law that authorizes the US government to do this specifically. There's no standard for doing it. There's no existing process for doing it. So today's episode is about what this means, who reacted and what a real regulatory framework would need to look like to prevent this from happening again. But before we do that, Ray, because this is a technically deep and charged episode, we'd like to have a word from our sponsor.

**Ray Rike** (2:29)
Okay. Now, we are not policy wonks, right?
But I want to walk through this very methodically, step-by-step. So let's look at the mechanics of the shutdown. The directive only targeted non-US citizens from accessing these two models. But once again, Anthropic had no real-time ability to identify and shut off the nationality of every user. So that was a big issue. Now, what's amazing is as a result of the US just having to shut down the model, it even affected Anthropic's own non-US citizen employees in San Francisco. And of course, it impacted all the technologists out there around the world working for allies, other governments, and over 150 companies across 15 countries.
And this was sent, people were using this model not for nefarious reasons, right? These were organizations using Fable 5 and Mythos 5 to identify and patch critical software vulnerabilities and power grids, hospital systems, banking infrastructure, and enterprise systems of Ratchagrid. Mid-task shut off.

**Peter Buchanan** (3:54)
Yeah, it's terrible. So the business exposure here is what you get every enterprise executive's attention because there was no warning, there's no fallback, there's no contractual protection. Every contract has forced majeure clauses in it, but they don't cover government mandated AI shutdowns. At least they didn't before June 12th.
So Gary Tan of Allspring Global Investments was very blunt about this. He's a pretty famous investor and he said, US AI frontier models are now strategic assets that are tightly controlled and that level of control is likely to increase as Chinese models close the performance gap.

**Ray Rike** (4:37)
Yeah. Now, we know AI is evolving so quickly and we know how long it takes governments to create statutory-based regulation. But let's talk about the US government state rationale here. They said that the ban was issued reluctantly, and they're saying they did it because an Anthropic partner and someone who has the White House's ear, Amazon, flagged the security vulnerability that Anthropic had actually declined to remediate voluntarily. So it appears that Amazon flagged the security concern to Anthropic and they declined to remediate it. So former White House AI's R David Sacks actually wrote about this publicly and said that the administration truly values Anthropic's technical capabilities and considers the issue serious but resolvable. So they're saying it's resolvable. But Anthropic took kind of a different perspective. They said, hey, the jailbreak was narrow.
It provided no meaningful uplift over other publicly available models and the directive against them and their models was disproportionate.

**Peter Buchanan** (5:55)
Yep. So the company's explicit statement said it supports the government's authority to block unsafe AI deployments.
In fact, they want mandatory testing of frontier models, but they want it through a statutory process that's transparent, fair, clear, and grounded in technical facts. And they don't think that happened this time at all.