AI safety concept

The public calls for an AI slowdown were only the tip of the iceberg. OpenAI's global policy chief Chris Lehane confirmed on September 15, 2026 that the company has been holding closed-door safety discussions with rivals Anthropic and Google DeepMind for several weeks — talks that began before any of the executives went public with their warnings.

Lehane made the disclosure to reporters in Washington, where he was meeting with U.S. lawmakers to address catastrophic risks associated with advanced AI models. Bloomberg first reported the story at 08:47 (UTC-7) on September 15.

What the talks are about

These are not vague industry goodwill gestures. According to Lehane, the three companies are discussing concrete mechanisms: a dedicated regulatory body for frontier AI, third-party evaluation frameworks, and quantifiable safety baselines that models would need to clear before deployment.

The conversations have been running in parallel with — and partly behind — the public campaign that kicked off when Anthropic CEO Dario Amodei published an essay on September 12 calling for the industry to collectively slow the pace of frontier model development. OpenAI's Sam Altman quickly endorsed the idea, and Microsoft followed suit on September 14. The White House pushed back, with officials describing the coordinated slowdown talk as "cartel-like behavior."

The timeline

Date (UTC) Event
Sep 12 Amodei publishes essay calling for coordinated AI slowdown
Sep 13 Altman says OpenAI is open to pacing frontier development
Sep 14 Microsoft backs slowdown; White House calls it a cartel
Sep 15 Lehane confirms weeks of secret talks between OpenAI, Anthropic, Google

Sources: Bloomberg, company statements, news reports.

The revelation that talks were already underway for weeks before Amodei's essay suggests the public slowdown campaign was not a spontaneous moment of collective conscience, but a coordinated messaging strategy built on pre-existing private negotiations.

Why it matters

The shift from public posturing to private standard-setting is the real story here. When three companies that collectively control the vast majority of frontier AI capability sit down to write their own safety rules, they are effectively drafting the regulatory framework that governments have so far failed to produce. That is both efficient and dangerous — efficient because these are the only organizations with the technical expertise to set meaningful benchmarks, dangerous because a self-regulatory regime designed by the companies it governs has inherent conflicts of interest.

The White House's "cartel" accusation is not entirely wrong, though it misses the point. The concern is not that these companies will collude to raise prices — it is that they will collude to define what counts as "safe," locking in standards that protect their incumbent position while raising barriers for new entrants. A third-party evaluation framework sounds neutral, but who gets to design the tests? Who pays for the evaluators? The answers to those questions will determine whether this is genuine safety governance or regulatory capture by another name.

Nvidia's Jensen Huang weighed in against new AI regulation, telling reporters that safety should be left to the industry. That position aligns conveniently with what these three labs are now doing — writing their own rules. But it also highlights a split in the industry: the model labs want safety guardrails (which also slow competitors), while the infrastructure provider wants no rules at all (because every model, safe or not, buys Nvidia chips).

The missing player is xAI. Elon Musk has been conspicuously absent from both the public slowdown calls and these private talks. If the Big Three agree on safety standards that Grok cannot meet, xAI faces a choice: comply and slow down, or operate outside the framework and risk being locked out of enterprise customers who demand certified-safe models.

What to watch