白宫AI框架保密引争议，两派罕见同声反对

Welcome to Transformer, your weekly briefing of what matters in AI. If you’ve been forwarded this email, click here to subscribe and receive future editions.

NEED TO KNOW

• Demis Hassabis moved to a new role as Chair of Google DeepMind and Chief Scientist of Alphabet.
• OpenAI revealed more details on how its agents coordinated with each other and took over internal infrastructure weeks before the Hugging Face hack.
• Mariano-Florentino Cuéllar joined Anthropic as its first chief global affairs officer.

But first…

THE BIG STORY

The White House AI framework has done the impossible: united both sides of the AI regulation debate against it.

The framework, finalized last week under June’s executive order, was meant to establish how the “voluntary” regime for testing AI models before release would work. Instead, it has managed to alienate safety advocates and regulation skeptics alike.

Based on what we know, the framework is seriously flawed. According to Axios, it applies to models “with state-of-the-art capabilities and national security risks.” But both terms lack a “clear definition,” undermining the whole point of the exercise: replacing the arbitrary process we’ve seen to date with predictable rules of the road. Open-weight models are exempt from testing, despite arguably posing the greatest risks. And the framework handles internal deployment bizarrely: companies are reportedly “encouraged” not to share models with the government until they’re “as close to public release as possible,” but once a model is shared, company employees can no longer use it. As John Schulman pointed out, this could push companies to run earlier, less safe versions internally, rather than the safer versions they eventually release. Given all we’ve learned recently about the risks of internal deployment, that is exactly backwards.

But the biggest issue is that we just don’t know very much at all. The White House has decided not to publish the framework: only a few companies that have signed up (or been coerced into doing so) have seen it.

That secrecy has drawn fire from both camps. “There may be good reason to classify the benchmark it uses to test those models. It has no good reason to hide how the program works,” the Abundance Institute’s Neil Chilson said, arguing that this is “no way for our democracy to govern the most important technology of our lifetimes.” Brad Carson, president of Americans for Responsible Innovation, called the move a “dangerous mistake” that “threatens public accountability.” He added: “A rulebook can only hold AI companies in check if people outside those companies know what the rules are.”

Both are right. AI is too important to regulate behind closed doors, especially given the government lacks much of the necessary expertise. The framework could almost certainly be improved if scrutinized by outside experts — who could also help check it’s actually being followed. And beyond the instrumental benefits, there’s the principle: in a democracy, the rules are public.

The immediate fix is for the White House to publish the framework — or be forced to. The Foundation for American Innovation has submitted FOIA requests to drag it into the light, while a group of Democratic senators has demanded access. But publication alone is not a solution. As Carnegie Endowment fellow Anton Leicht wrote this week, as long as regulatory power over AI lives in the White House, secrecy will be the default. The only long-term solution is for Congress to do its job and bring AI governance under legislative control.

— Shakeel Hashim

ALSO NOTABLE

Rep. Greg Casar, alongside Reps. Valerie Foushee and Sara Jacobs, yesterday introduced the AI Tax and Work Protection Act, designed to redistribute the financial benefits of the AI boom to the working class by creating construction, child care and elder care jobs. The bill would impose a tax on either tokens or revenue, whichever is higher, to be paid by either a model developer or, in the case of open-weight models, the company deploying them. It fits with the broad populist messaging coming from the left, such as Alex Bores’ UBI-style “AI dividend,” or Sen. Sanders’ proposal for an AI sovereign wealth fund.

AI-specific taxation schemes might not actually be as effective as broader consumption or income taxes, but that’s beside the point. This is a messaging bill, not something anyone thinks will pass this Congress (introducing it during recess is enough to tell you that). But as far as messaging bills go, it’s an important one: Casar chairs the Congressional Progressive Caucus, which includes both Foushee and Jacobs, and where he goes, so do other members of Congress’s left wing. Casar is testing a progressive message on AI, not just for the 2026 elections, but for 2028.

“The number one, broad concern from voters with AI is: ‘Is this going to devastate my economic future or my kids’ economic future?’” Casar told Transformer on the call announcing the policy. “Democrats don’t have a policy plan for dealing with this, and we want to change that today. Here is the clear plan to make sure that if unemployment rises and AI starts replacing a large number of jobs … to make sure that Americans are still at work.”

— Veronica Irwin

THIS WEEK ON TRANSFORMER

• Plz Don’t Kill Us: Inside AI safety’s influencer bootcamp — Celia Ford looks at whether TikTokers can make existential risk mainstream
• What the latest rogue AI incidents should teach us — Shakeel Hashim argues that we need better AI testing practices — and to solve alignment before training more powerful systems

Subscribe now

THE DISCOURSE

Rep. Greg Casar wants to ban superintelligence:

• “AI safety regulation is one of the most important issues the entire country is facing … It’s an emergency, and we need to act like it.”

Sen. Chris Murphy tweeted:

• “The disclosure that an OpenAI model ‘escaped’ its containment should scare the hell out of everyone. As should how flippant the company has been about the development. Meanwhile, Trump continues to protect the industry from regulation. A disaster in the making.”

The Washington Post’s editorial team pushed back against the “escape” framing:

• “[Incidents at OpenAI and Anthropic] have been described in the press as ‘escapes.’ A ‘lab leak’ would be more accurate. An ‘escape’ suggests the existence of an actor that wanted out and schemed past its keepers. A ‘leak’ puts the responsibility on those that failed to prevent it.”

Encode’s Nathan Calvin had a message for catastrophic risk skeptics:

• “[You] may have had a point that it is challenging to work on these problems before the problems become clear — but for better or worse, the problem is now extremely clear!”
• “Agreeing AI alignment risks are deadly serious and real does not mean you have to support a specific piece of legislation, or have a positive feeling about a specific person or organization … But I am now really really profoundly sure that these risks are serious and real. And I think anyone weighing the evidence objectively should agree.”

Ex-OpenAI futurist Joshua Achiam thinks “people worried about AI cyberweapons are missing the point”:

• “The problem is that we built the software layer of civilization on spaghetti code loaded with zero days.”

OpenAI’s roon is freaking out:

• “when I freak out over loss of control incidents, it’s not because the limited damage they have caused is anything close to the positive value of the technology … the actual problem is that it’s better and more accurate to think of these things as potentially self-replicating life-like forms that can turn into digital infections under the wrong conditions. and as their intelligence becomes unbounded, so too does the damage they can cause.”

Sam Altman, meanwhile, shared a “cool use case of ChatGPT work”: