Dario Amodei published a warning. AI systems capable of seizing control of critical infrastructure in the near term, cyberattack risks, biological weapons supercharged by models that no one fully audits anymore. Anthropic's CEO called on the industry to slow the pace at which it improves its products' capabilities. Donald Trump responded by calling these criticisms "negative forces" and reaffirming that the United States cannot afford to lose the technological race against China. This scene sums up the tensions of 2026. The builder of the tool warns about its dangers. The leader funding its development insists there is no time for pauses.

This dynamic amounts to a conflict of interest, because the same person generating the capability is also generating the public alarm. Amodei holds the most direct technical knowledge of what Claude and its successors can achieve. When the head of a leading organization raises his hand to signal that progress could spiral out of control, it's worth listening closely. It's the engine's designer warning about a brake failure.

The internal logic of that thesis deserves credit. Anthropic has held documented talks with the Pentagon about using its models in defense contexts. Its head of AI safety, Mrinank Sharma, resigned in February amid speculation about internal disagreements over the pace of deployment. And that wasn't the only signal. In June, one of the company's models breached classified NSA defenses within hours. The episode revealed gaps between capability and oversight that are no longer theoretical.

So why does the warning feel incomplete? The same organization calling for a slowdown is racing to get to the next generation of models first. Anthropic keeps releasing versions, signing government contracts, and raising capital at valuations that only make sense with accelerated growth. The public statement and the private behavior point in opposite directions. That contradiction isn't a footnote. It's the whole trend.

Who really benefits when an AI company's CEO calls for regulation while his team accelerates behind closed doors? The answer has a cold logic to it. A well-calibrated warning positions whoever issues it as the industry's responsible voice. At the same time, it pushes toward rules that will likely be written by the very actors requesting them. Barriers to entry that protect those who already made it in. Climb the ladder first, then pull it up.

Something similar is happening around DeepSeek and the absence of verifiable audits before deployment. No one, not governments, not companies, has real processes to confirm a model's safety before releasing it. The warnings come afterward, once millions have already adopted it. Amodei talks about losing control of the internet in the short term. Yet Anthropic hasn't put forward a verifiable pause mechanism. It has only suggested that others regulate first.

David Sacks, the person the White House appointed to head AI policy, illustrates the trap in this debate. He acknowledges cybersecurity risks. Yet every time a concrete restriction is discussed, the conversation circles back to the same point: if the United States slows down, China won't. Trump puts it with even less nuance. Any sign of caution reads as weakness against Xi Jinping. There are no real incentives to stop. Not at Anthropic, which must show progress to sustain its valuation. Not at the White House, and not in China, both operating under the same mirror-image logic.

This recalls how, during the nuclear race of the mid-twentieth century, the scientists who designed the bomb later called for international controls. Geopolitical incentives prevailed anyway. No real pause.

What's missing from the current conversation is a question almost no one asks clearly: who actually holds the power to decide the pace, beyond who is technically right? You can have all the right evidence about the risks and still have zero capacity to impose a stop if the actors with market power don't want to stop. I explored that distinction in The Generosity in the Doorway. The difference between knowing what a responsible arrangement would look like and holding the political and economic power to make someone adopt it.

Amodei may have written the most honest warning of the year. Without structural incentives, real regulatory pressure, or competitors willing to cede ground, the text becomes a symbolic gesture. Not a brake.

For anyone who doesn't work in tech or follow the Silicon Valley debate, this means something concrete. The decision about how much collective risk to accept, from cyberattacks, from enhanced biological weapons, from systems acting beyond human control, is being made in rooms where most people have no voice. Those who do have a voice, Trump, Xi Jinping, frame any caution as geopolitical weakness. This isn't a technical problem that Anthropic's engineers will solve. It's a governance problem that still has no democratic mechanism for resolution.

Whoever controls the assessment of danger usually also controls the narrative of urgency. Both rarely benefit whoever is outside the room. The underlying question isn't whether Amodei is right about the risks. He probably is, at least in part. The question is why we keep accepting that the person with the greatest conflict of interest is also the main voice setting the pace.

While Amodei was publicly warning about the loss of control, Anthropic kept signing contracts with the Pentagon to expand the military use of its models. The warning and the expansion happened in parallel. No pause. The company that called for a slowdown never slowed its own. So who really decides the pace?

Sources

1. DW Español — coverage of Dario Amodei's call to "hit the brakes" on AI development

2. The Wall Street Journal — emails revealing communications between Anthropic and the Pentagon

3. Bulletin of the Atomic Scientists — analysis of the OpenAI-Hugging Face incident and US-China AI relations

4. Reddit/public documentation — resignation of Mrinank Sharma, Anthropic's head of AI safety