AI's Loudest Alarms Are Ringing From Inside the Building

In September 2026, Anthropic CEO Dario Amodei published an essay arguing the AI industry needs to slow down, warning that competitors "don't really understand the risks they're taking."

Weeks earlier, his own company had signed compute contracts worth up to 517 billion dollars, racing to keep pace with OpenAI's even larger buildout plans.

Both things are true at once, and that's the pattern worth understanding: the loudest warnings about AI risk keep coming from the same people building AI as fast as anyone alive. The gap between what AI leaders say and what their companies do isn't a contradiction they've failed to notice. It's how the industry currently operates.

The Essay and the Invoice

Amodei has spent much of the past two years writing about AI risk in careful, alarming detail: concentrated power, catastrophic misuse, a pace of development humanity might not be able to manage safely. His September 2026 essay called for the industry to deliberately pace itself.

Anthropic's actual behavior tells a narrower story. The company has reportedly signed compute deals worth up to 517 billion dollars over eleven months and is nearing a valuation approaching a trillion dollars, up more than fifteenfold in about a year. It still trails OpenAI's own reported 750 billion dollar spending plan through 2030.

Think of it like a doctor publishing a paper on the dangers of a drug while running the factory that makes it faster than any competitor. The warning can be sincere. The factory still ships.

Both companies are pouring that money into the same chips and data centers straining local water and power supplies, built by the company whose graphics cards became the AI industry's backbone. None of that spending pauses for an essay.

Did You Know?

Anthropic's compute commitments are now large enough that bond investors, not just AI researchers, are tracking them. The same debt markets financing the broader AI buildout have started demanding higher returns for the risk of lending into it, an early signal that the money behind AI's growth is getting more skeptical even as the companies keep spending.

Testifying for Rules, Lobbying Against Them

Sam Altman told the US Congress in 2023 that AI needed regulation, warning that "if this technology goes wrong, it can go quite wrong." Lawmakers treated the moment as a rare instance of an industry asking to be constrained.

At the same time, OpenAI was lobbying European Union officials to keep general-purpose systems like GPT out of the AI Act's strictest "high risk" category. Internal documents reported by Time showed the company pushing this position through 2022, before the law's final language was set.

The lobbying worked. The Act's final text dropped earlier language that would have automatically classified foundation models as high risk, and OpenAI continued pressing European lawmakers into 2026 to soften the law's implementation code. A company can ask for guardrails in one room and ask for fewer of them in another, and both requests can be genuine to the person making them. They just don't point toward the same outcome.

The Team Built to Watch for the Danger, Then Dissolved

In 2023, OpenAI created a team called Superalignment, given a four-year mandate and a fifth of the company's computing power to solve the problem of controlling AI systems smarter than their creators. It was the company's clearest public commitment that it took its own warnings seriously.

The team lasted about a year. Its co-leader Jan Leike resigned in May 2024 and wrote that "safety culture and processes have taken a backseat to shiny products," adding that his team had spent months "sailing against the wind" for basic computing resources. OpenAI dissolved the team days later.

A company that builds a dedicated safety unit is making a public promise about its own priorities. Disbanding it within a year, while product releases kept accelerating, is the clearest evidence available that the promise and the roadmap were never fully aligned.

Interesting Facts: OpenAI was founded in 2015 as a nonprofit specifically to keep advanced AI from being controlled by any single company. Elon Musk, one of its founders, sued OpenAI in 2024 arguing it had abandoned that mission by restructuring around a for-profit arm. A federal jury ruled against him in 2026, but on a technicality: it found he'd waited too long to sue, not that the underlying claim about mission drift was wrong.

When the Warning Never Reaches the People It Affects

The clearest cost of this gap doesn't show up in a hearing room. It shows up in decisions made by algorithms nobody outside the company gets to question until a lawsuit forces the question open.

UnitedHealth is facing a federal class action alleging it used an AI tool called nH Predict to help decide how much rehabilitation care Medicare Advantage patients could receive after hospital stays, despite the tool carrying what the lawsuit describes as a 90 percent error rate on appeal. A STAT News investigation cited in the suit found the company pressured staff to keep patient stays within a narrow margin of what the algorithm predicted, while betting that only a small fraction of denied patients would appeal. UnitedHealth says the tool only informs care planning and that coverage decisions follow separate criteria.

A similar pattern played out on the road. In October 2023, a GM Cruise robotaxi dragged a pedestrian roughly 20 feet after a collision in San Francisco.

California's DMV later said Cruise withheld the full crash footage from regulators, showing only part of what its cameras had recorded. The state suspended the company's driverless permits, and Cruise's own robotaxi ambitions never fully recovered. It's a reminder that the companies now expanding driverless fleets into new cities are the same ones deciding, on their own, what regulators get to see first.

Neither company built its algorithm to cause harm. Both had every incentive to disclose the problem quietly and adjust rather than let outside reporting or a courtroom do it for them. Neither did, until they no longer had a choice.

Knowlegic Perspective

None of this means the warnings are fake. Amodei's essay, Altman's testimony, Leike's resignation letter, they read as sincere, and probably are. People inside these companies genuinely believe the technology they're building could go badly wrong.

The harder truth is that sincerity and structure are different things. A CEO can believe every word of a risk essay and still answer to investors who expect the fastest possible growth. A safety team can be fully resourced on paper and still lose every internal argument against a product deadline. The warning doesn't need to be dishonest to lose to the incentive.

That's the actual story here, not hypocrisy in the simple sense, but a structure where the people most aware of the risk have the least power to slow down for it. Until that changes, the loudest alarms about AI will keep coming from inside the building doing the building.

Sources & References

• OpenAI CEO warns Senate: "If this technology goes wrong, it can go quite wrong", ABC News (2023)
• Mr. ChatGPT goes to Washington: OpenAI CEO Sam Altman testifies before Congress on AI risks, CNN Business (2023)
• OpenAI's former superalignment leader blasts company: "safety culture and processes have taken a backseat", VentureBeat (2024)
• Former OpenAI leader blasts company for ignoring "safety culture", Fast Company (2024)
• Anthropic CEO calls for "pacing the frontier" of AI race amid safety concerns, CNN Business (2026)
• Anthropic signed $517bn in compute agreements in past 11 months, Data Center Dynamics (2026)
• OpenAI lifts planned compute spending to $750 billion through 2030: WSJ, Yahoo Finance, citing The Wall Street Journal (2026)
• UnitedHealth uses faulty AI to deny elderly patients medically necessary coverage, lawsuit claims, CBS News (2026)
• Musk loses case against OpenAI, CNN Business (2026)