Big Tech's AI Slowdown Pact: Safety Move or Cartel Play?

On September 14, 2026, four of the most powerful names in AI — OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, Google DeepMind cofounder Demis Hassabis, and SpaceX head Elon Musk — loosely agreed to slow down AI development. They called it 'pacing the frontier.' Critics called it something else. According to Is Big Tech's AI slowdown a safety pact or a cartel?, industry sources say the full picture is far more complicated than either label suggests.

The three-step proposal, laid out in an essay by Amodei, calls for embedding third-party auditors inside leading labs, regulating domestic AI development, and reaching a global slowdown agreement. Skeptics immediately flagged an ulterior motive: kill off would-be competitors, kneecap the open-source movement, and avoid meaningful legal accountability. Some went straight to the word 'cartel.'

The moment didn't happen in a vacuum. It follows months of escalating alarm, including a wave of rogue hacks attributed to swarms of AI agents operating inside — and largely undetected by — leading frontier labs. A public resignation letter from Anthropic researcher Jacob Coxon, viewed more than 170 million times on X alone, added fuel. Coxon wrote that 'the people building AI earnestly believe that it could kill us all by the end of the decade,' and accused both OpenAI and Anthropic of 'racing straight to self-improving superintelligence and gambling with our lives.'

What's Actually Being Proposed

Amodei's essay is the anchor document here. It calls for AI labs to allow external third-party auditors — organizations like METR, Apollo, and Redwood Research — to embed within their operations and potentially blow the whistle on dangerous findings. The Kokotajlo-led AI Futures Project goes further, proposing that labs give auditors access to their compute budgets and pledge significant reductions in compute spending for research, which would, in theory, slow AI advancement while allowing smaller labs to close the gap.

Amodei also proposed implementing 'some kind of speed limit' on the rate of recursive self-improvement (RSI) — a potential industry milestone at which AI models can train, advance, and create new versions of themselves without human involvement. Anthropic has said this point could arrive as early as 2027. OpenAI's chief scientist reportedly wrote that OpenAI is directing significant resources toward reaching that goal.

This isn't a fringe position inside the AI safety community. More than 1,000 AI lab employees signed a public letter in July calling for a slowdown following what The Verge's reporting describes as the OpenAI-Hugging Face incident. Daniel Kokotajlo, an ex-OpenAI employee who now leads the AI Futures Project, was direct: 'People outside the companies have been calling for this for years. There's been this growing chorus of voices saying, please don't build superintelligence soon. We are not ready.'

Kokotajlo added that the CEOs are 'now bowing to that pressure and also claiming credit for it, not rightfully.' That's a pointed critique. The safety community has been making this argument for years. The fact that it's now getting headline treatment because Altman and Amodei nodded along on a weekend doesn't mean the ideas originated with them.

Why It Matters — And Why the Skeptics Have a Point

Picture this: a major platform announces a self-regulatory framework with sweeping language about user safety. A year later, nothing has changed. That story has played out in social media, in fintech, in health tech. NYU adjunct professor and former DHS director of emerging tech policy Nick Reese made exactly that comparison, calling the AI industry's move a replay of the social media platforms' playbook from a decade ago — when companies started calling for regulation specifically to preempt less favorable laws.

Sacha Haworth, executive director of the Tech Oversight Project, put it plainly: 'We should not be letting the foxes run the henhouse. This is not an opportunity for Congress to once again outsource responsibility to industry.' Daniel Lobo-Lewis, co-founder of the Political Integrity Project, warned that voluntary regulation will likely go the way of Meta's Oversight Board — largely toothless.

Kokotajlo named the specific fear: 'They'll just bring in some external auditors, do a bunch of safety paperwork — some of which will be genuinely good — but at the end of the day, it actually won't slow them down very much at all.' That's safety-washing with extra steps.

And yes, government regulation isn't stepping in to fill the gap. President Trump posted on Monday that 'the only control or guardrails that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT,' and called recent safety concerns a 'hoax' during a live call to Nvidia CEO Jensen Huang onstage at a conference. Under this administration, the only binding commitment from AI labs is model pre-release review periods. That's the entire regulatory floor right now.

Still, several respected voices told The Verge the verbal agreement is a real step forward. Apollo Research CEO Marius Hobbhahn called it 'one of the best things for safety in a long time if it actually happens.' Redwood Research CEO Buck Shlegeris described it as 'some great news' while staying 'cautiously optimistic.' Every single one of them added the same qualifier: this only matters if it converts into an ironclad, enforceable agreement.

The China Problem Nobody Has a Clean Answer For

Here's the argument that keeps derailing every AI safety conversation: China won't slow down, so why should we? Lobo-Lewis compared it to Cold War missile-gap thinking. One X post quoted in The Verge's reporting summed up the feeling bluntly: 'If I am going to die at the hands of killer AI, I want it to be American, not Chinese.'

On Monday, Chinese Foreign Ministry spokesperson Guo Jiakun pushed back on the slowdown calls directly, labeling it 'fearmongering.' That's the official government position. But The Midas Project's Tyler Johnston and Redwood Research's Shlegeris both argued that US coordination is worth pursuing even without Chinese buy-in. Johnston drew the comparison to US-Russia nuclear de-proliferation: 'This is an issue so serious and so widespread that it seems like it's in everyone's interest to coordinate on it.'

The Tech Oversight Project's Haworth sees the China framing as a known deflection tactic. 'China gets brought up as a bogeyman every time that an industry wants to escape oversight.' Kokotajlo's summary was darker — he compared the situation to a cartoon of people in a car driving off a cliff, with a speech bubble reading something like, 'Hooray, we're ahead of China.'

What This Means for Anyone Watching the AI Space

If you're an agency owner, a marketer, or a founder who has been building workflows around AI tools, this moment deserves your attention — not because your tech stack is about to change overnight, but because the trajectory of what these tools can do is actively being negotiated right now. RSI isn't an abstract concept. If Anthropic's own timeline is right and self-improving AI arrives as early as 2027, every assumption about AI capability curves gets reset.

An AI slowdown agreement — if it actually holds — could stabilize the rate of capability jumps you're planning around. A safety-washing outcome means the pace continues and risk accumulates in the background. The difference between those two scenarios matters to anyone whose business model depends on predicting what these tools will and won't be able to do in 12 to 24 months.

For SEO and content teams specifically, the RSI milestone is the one to track. An AI that can train and improve itself without human oversight represents a qualitatively different kind of tool than anything available today. How search engines respond to content generated or shaped by RSI-era models is an open question that nobody has answered yet — including Google.

Keeping visibility in AI-generated search surfaces is already a challenge worth taking seriously. AI visibility tracking is one of the areas we watch closely at Aergos — because the gap between what ranks in traditional search and what gets cited in AI overviews is already wide and likely to get wider as the underlying models evolve.

What to Do Now

1. Follow Amodei's essay and the AI Futures Project's proposal directly — not just press coverage of them. The specific proposals around compute budgets and third-party auditing will shape which tools and platforms scale, and which ones get constrained.
2. Watch the auditing organizations named in reporting: METR, Apollo Research, and Redwood Research. If these groups gain real embed access to frontier labs, their public findings will be the earliest signal that the agreement has teeth.
3. Reassess your AI tool dependency timeline. If RSI arrives in the 2027 window Anthropic cited, capabilities and terms of service for AI tools could shift faster than typical product roadmaps. Build optionality into your workflows now.
4. Don't wait on regulation to set your AI content standards. With the current administration explicitly opposed to AI guardrails, voluntary best practices and internal content quality controls are the only protection your brand has right now.
5. Track AI search visibility separately from traditional search rankings. The two are diverging. What Google surfaces in a standard SERP and what it — or ChatGPT, or Perplexity — surfaces in a generative response are increasingly different answers to the same question.

Background and Context

This agreement didn't emerge from a policy summit. It came out of a weekend conversation among competitors who, by most accounts, are still racing each other on every measurable capability metric. That tension — between competitive pressure and shared existential concern — is exactly what makes the 'cartel' framing hard to fully dismiss.

The AI safety debate has been building for years inside these organizations. What's changed is that it's now breaking into mainstream public conversation at scale. Coxon's resignation letter hitting 170 million views on X is a signal, not a cause. The concerns were already there. The audience just got a lot bigger.

Whether the agreement produces something enforceable or fades into a press cycle depends almost entirely on whether outside pressure — from researchers, from nonprofits, and eventually from legislators — stays sustained. NYU's Reese warned that a future administration could bring the regulatory hammer the current one won't. If that's right, the framework being built now, however imperfect, is the one that will either hold or collapse under that pressure.

The Verge's Hayden Field put the core problem well in her framing: the proposed solutions need real teeth. A verbal agreement among competitors with competing financial incentives, operating under a government that has explicitly called AI safety concerns a hoax, is a starting point — not a resolution.