The Rivals Who Started Talking: OpenAI, Anthropic & Google on AI Safety

What Happened This Week

On Tuesday, September 15, 2026, a quiet but potentially historic development surfaced in the technology press. Chris Lehane, OpenAI's global policy chief, told reporters that the company has been working with its two fiercest rivals — Anthropic and Google DeepMind — on AI safety for weeks. He said OpenAI had traveled to Washington to help U.S. lawmakers understand and address the catastrophic risks associated with increasingly powerful artificial intelligence systems. Bloomberg broke the story; CNBC, TechCrunch, Fortune, and The Information quickly confirmed and expanded on it.

The timing was not accidental. Just days earlier, on Saturday, Anthropic CEO Dario Amodei published an essay urging the entire industry to slow down how quickly it improves its most advanced models. Amodei argued that the race to build ever-more capable "frontier" models carries risks that no single company can manage alone. The call was endorsed by a strikingly broad coalition: Sam Altman of OpenAI, Demis Hassabis of Google DeepMind, and Elon Musk of SpaceXAI all publicly supported Amodei's message.

But the most revealing part was what happened behind closed doors. According to reporting by The Information, the three companies have been working together to create a standards body for the AI industry — an idea Altman reportedly told staff would need to happen without the support of the U.S. government. That last detail is loaded, and it will matter a great deal to the rest of this article.

Why this matters: For the first time, the world's three most powerful AI labs — companies that spend billions and hire thousands to out-innovate each other — are openly admitting they need to cooperate on safety. If rivals in any other industry started doing this, it would make headlines for entirely different reasons. Here, the stakes are about the future of the technology itself.

Where This Idea Came From

To understand why these talks feel so significant, you need to know a little about the people behind them and the proposal that set things in motion.

Demis Hassabis, the chair of Google DeepMind, published an essay in July on X (formerly Twitter) proposing something bold: a public-private partnership, or what some call a self-regulatory organization, operating under federal oversight. His model was the Financial Industry Regulatory Authority, known by its abbreviation FINRA — the independent body that watches over the U.S. financial markets. Hassabis suggested something similar for AI: a watchdog with real power to screen the world's most advanced models and, if dangers grew, to coordinate slowdowns across the whole industry.

"Dario's essay points towards the right path forward." — Demis Hassabis, on Dario Amodei's call to slow frontier AI.

Think of it this way. When you buy stocks, you don't trust a broker blindly. There are rules, auditors, and regulators who make sure nobody is manipulating the market or taking reckless bets that could crash everything. Hassabis is essentially arguing that AI deserves the same kind of oversight — a referee so that the race to build smarter machines doesn't become a race to the bottom.

Now, three months later, those same rivals are reportedly building the referee together. That is a dramatic shift from the public posturing of recent years, when each lab tried to out-promise the other on speed and capability.

What This Could Look Like in Real Life

It is easy to hear "frontier AI safety standards" and think of abstract policy papers. But the idea touches everyday things. Imagine a few everyday scenarios.

A hospital using AI

A doctor relies on an AI to help diagnose a rare disease. A safety standards body could mean that system was independently tested for dangerous mistakes before it ever reached the clinic.

A self-driving car

The brain behind a new autonomous vehicle is one of these frontier models. Independent verification could ensure it handles unexpected situations — a child running into the street — before it's allowed on the road.

A bank's loan decisions

An AI decides who gets a mortgage. A standards body could check that it isn't unfairly rejecting people based on their postcode or name — protecting you from biased automation.

Your chatbot assistant

The assistant in your phone gets smarter each year. If labs agreed to slow down and verify each upgrade, you might get fewer surprising, harmful, or misleading answers.

None of this means the technology would stall. The idea isn't to stop progress — it's to build guardrails while the car is still moving. But for the first time, the drivers are agreeing on what the guardrails should look like.

How This Could Be Used Across the World

If these talks lead to a real standards body, the applications would reach far beyond the three companies involved.

Healthcare worldwide: Independent safety checks on AI diagnostic tools could let hospitals in low-income countries trust advanced models they otherwise couldn't afford to vet themselves.
Climate and energy: AI now helps design better batteries, predict grid failures, and model climate patterns. A shared safety framework could accelerate trustworthy deployments where mistakes are costly.
Education: Students already use AI tutors by the billions. Standards could ensure those tools are safe, accurate, and appropriate for children before schools adopt them.
Emergency response: During disasters, AI coordinates relief efforts and predicts the paths of storms. Verified, reliable models could mean faster, more accurate help when it's needed most.
Small businesses: A bakery, a clinic, or a logistics firm that uses AI for scheduling or forecasting would benefit from the same baseline of trust that large corporations get.

The common thread is credibility. When an independent body says a model is safe enough to use, ordinary people and ordinary companies can trust it — something the industry has struggled to earn on its own.

Where This Technology Is Heading

Let's picture three possible futures, from near to far.

Short term (1–2 years)

The three labs likely roll out "third-party evaluators" embedded inside their own systems — independent teams that check models for safety before each major release. You might not notice this directly, but the apps you use would carry a quieter stamp of verification.

Mid term (3–5 years)

If the standards body takes root, it could become a recognized authority — like a safety agency for AI. New models across the industry might need certification before they can be sold or deployed, much like cars pass safety tests or foods pass health inspections.

Long term (10+ years)

If AI keeps growing in capability, a robust standards body could become essential infrastructure — a global referee ensuring that increasingly powerful systems stay aligned with human values. Whether that future is reassuring or unsettling depends entirely on who controls the referee.

How This Technology Evolved to Reach This Moment

The road here was long. From 2023 to 2025, AI moved from chatbots that wrote emails to systems that could reason, plan, and act — an era many call the "conversational" phase giving way to "agentic" AI that can take real-world actions. Each leap in capability has outpaced the rules meant to govern it.

The current talks are a reaction to that momentum. As models got better at coding, science, and even cybersecurity, concerns grew that speed alone was becoming the only metric that mattered. The emergence of AI-focused cyber tools, for instance, prompted over 100 companies this year to demand new safeguards. The safety talks are the industry's way of pushing back against a pure speed race — an admission that capability without caution may be a dead end.

The Good, The Risky, and The Uncertain

Any development this consequential cuts both ways. Let's look honestly at both sides.

The potential benefits

Slower, safer progress: Coordinating on safety could prevent reckless races that prioritize power over prudence, reducing the chance of a harmful mistake.
Public trust: Independent verification gives ordinary people a reason to trust AI — which the technology needs to deliver on its promises in healthcare, transport, and beyond.
A shared safety net: No single company can manage existential risks alone. A coordinated body spreads responsibility and pools expertise.
Global leadership: If the U.S. and its allies build a credible framework, it could shape how the whole world regulates AI — rather than leaving it to chaos.

The risks and concerns

Antitrust red flags: Three competitors agreeing to slow down could look like a cartel. Altman himself has warned such talks might violate competition law, and regulators worldwide will be watching closely.
Barriers for newcomers: If safety certification is expensive and complex, only the biggest labs could afford it — locking out startups and reducing competition, the opposite of what competition regulators want.
Self-policing skepticism: A body run by the very companies it oversees may lack the independence needed to enforce real consequences. "Referees funded by the players" is a hard sell.
Geopolitical gaps: While the West debates slowdowns, other nations may pour resources into unchecked AI development, complicating any coordinated pace-setting.

There is also a political wrinkle that makes this especially complicated. President Trump has publicly dismissed AI safety concerns as a hoax and pushed back on regulation, arguing that any slowdown would give China a leg up in the technological race. His AI advisor, David Sacks, has said similar, calling fears of existential risk overblown. Meanwhile, the very companies pushing for standards say they want to do it without government support — a delicate balancing act that could satisfy neither regulators nor competitors.

The Bottom Line

What we're witnessing is a possible turning point. The labs that spent years shouting about who would build the smartest machine are now quietly admitting that none of them can guarantee it will be safe on their own. That is a rare and perhaps necessary moment of humility in an industry built on bold claims.

But humility alone doesn't build guardrails. The success of these talks will depend on whether the resulting body is truly independent, genuinely open to smaller players, and willing to enforce real standards — not just issue friendly recommendations. The world will be watching to see whether this becomes a new era of responsible AI, or just another PR move in a race that never really stopped.

One thing is clear: the next five years of AI will be shaped less by raw intelligence and more by who we trust to steer it. The rivals have finally started talking. The real question is whether they'll build something worth trusting — or just talk themselves into the same old race, with better brakes.

The Rivals Who Started Talking