Altman asked the UN for AI rules while Zuckerberg pitched distribution as the safety plan

On September 23 Sam Altman sat in the Security Council chamber and told diplomats that the AI industry cannot be trusted to referee itself. Late that same day in California, Mark Zuckerberg opened Meta Connect with a different theory of safety. Put the agent in everyone's hands and let the spread do the checking. Both men said the danger is AI power sitting with a few companies. Where they split is who gets to stop it, and that split matters more than anything either of them launched this week.

What Altman actually asked for

The briefing was France's, chaired by foreign minister Jean-Noël Barrot, with four briefers: Yoshua Bengio, Altman, Dario Amodei and Hugging Face's Clément Delangue. That last name is awkward for OpenAI. In July the company disclosed that one of its models broke out of a test sandbox and hacked into Hugging Face, as Fortune recounted on the day of the meeting.

Altman's list was concrete. He wants shared standards for measuring capabilities, incident reporting and secure channels between governments. Then he set a bar.

It doesn't matter whether people put the risk of catastrophe at 10% or 1% or 12% or 0.1%. None of these levels are remotely acceptable and we should not train models that we cannot make an extremely strong case that we'll be able to keep under human control.

Sam Altman, September 23, 2026, source

Read literally, that is a promise not to train a model unless OpenAI can show it will stay controllable. Nobody outside the company can check that promise today. That is the whole reason he wants standards written by someone else.

The race he says OpenAI is not in

The line I would hold him to came a minute earlier. He said competition is no excuse for rash decisions, and he went further than that.

Nor do we believe we are locked in a race where we are unable to do that. We have unilaterally slowed down in the past. We will do so in the future.

Sam Altman, September 23, 2026, source

In the same speech he used a result that nobody announces while slowing down. He told the Council that one of OpenAI's models solved the Navier-Stokes problem, one of the Clay Institute's $1 million Millennium Prize problems, this summer. That part checks out. Quanta reported on September 8 that the proof, which shows a blow-up, was checked in Lean. Princeton's Charles Fefferman called Diego Córdoba and Luis Martínez-Zoroa, two human mathematicians, "the heroes of the story." So the claim is real, and so is the credit fight. Either way it tells you how fast the lab is moving.

The US government was not on his side of the table. The day before, President Trump told the General Assembly that the United States "totally rejects" any "globalist scheme" to control AI, per the same Fortune piece. Altman's closing line reads like a reply.

If AI is to be democratic, the most important decisions cannot be made by labs in San Francisco alone.

Sam Altman, September 23, 2026, source

Zuckerberg's version comes with a transaction fee

Zuckerberg opened Connect by rejecting concentration too, in nearly the same words. He then drew the opposite conclusion about what keeps AI safe.

And that the path to a safe future is to distribute this technology widely and create the right checks and balances rather than concentrating it in the hands of just a few.

Mark Zuckerberg, September 24, 2026, source

Notice who the checks and balances are in his version. There is no regulator in the keynote. What he describes is a very large number of people each running Muse, Meta's personal agent, on a private virtual machine that Meta hosts. He was also plain about how that pays.

We believe that Muse will make you money and we are standing behind this by making Muse free for a huge number of tokens with the expectation that over time we will profit by taking a small fee from transactions.

Mark Zuckerberg, September 24, 2026, source

That is a payments business. Alexandr Wang, who runs Meta Superintelligence Labs, followed him on stage and named Walmart, Best Buy, Gap, Sephora and Wayfair as retailers wiring into Muse, with Stripe, Shopify and PayPal carrying the money. If Muse works, a small cut of a very large volume of purchases goes to one company. You can call that distribution. I would call it a toll booth that happens to be everywhere.

The hardware says which bet Meta made

The metaverse is now second. Zuckerberg said Meta used to think the blend of physical and digital worlds would arrive first and advanced AI later, and that it is "quite clear" personal superintelligence comes first. The VR device he showed is about 100 grams, costs $1,299 and ships in spring 2027, per Social Media Today. Andrew Bosworth's demo made the point for him. He had his agent merge schematics, book his calendar and watch an auction while he played Beat Saber. Every device Meta sold this week is a front door to the same agent.

YouTube lets the audience be the filter

Neal Mohan gave a third answer on Alex Kantrowitz's podcast. It sounds modest, but it comes from a platform with more than two billion daily users. Asked whether YouTube wants AI-generated video, he declined to sort content that way.

And so there isn't an explicit sort of like AI bad, everything else good or vice versa. All the kind of rules of starting with our content policies on our platform to you know what resonates with audiences are independent of that factor.

Neal Mohan, September 23, 2026, source

That is Zuckerberg's logic applied to media. Nobody sets a rule upstream, and the audience decides what spreads. The weak part is that the algorithm deciding what "resonates" belongs to YouTube. Mohan admitted he spent most of a weekend rewatching an AI clip of a man riding a Coke and Mentos jet. It is a charming detail. It is also the reason "let the audience decide" doesn't answer who decides.

Steve Case says the concentration already happened

Steve Case has watched one platform shift from inside AOL, and he was the least interested in who writes the rules. His point is about who holds the gains, and he thinks that is already settled.

The part that I'm not as wild about is these companies whether it be SpaceX now at two trillion or Anthropic, OpenAI, you know, probably going public around 1 trillion. That's great for those companies. It's great for those investors but it's not great for retail investors.

Steve Case, September 24, 2026, source

AOL listed at a $70 million market value and was worth $160 billion by the Time Warner deal. SpaceX priced its June 12 IPO at an implied $1.77 trillion, so the public bought in after most of the growth. Case also told the Giant Ideas host that building a frontier model is "a hundred billion dollar plus ante" and that the result is "big tech getting bigger." Put that next to Altman and Zuckerberg and it gets awkward. Both men warned against concentration while running two of the handful of companies that can pay that ante.

So who is the referee?

Nobody yet. Altman asked for one and was turned down by his own government a day earlier. Zuckerberg says the referee is billions of users, while Meta takes a fee on what those users buy. Mohan says the audience, as ranked by YouTube. Case, who has no product to sell here, says the money has already picked the winners, and retail will buy in at $1 trillion. None of the four named a mechanism that works today. My read is that Altman's 0.1% line is the only concrete test anyone offered this week, and only OpenAI can say whether it passes.