If AI Is So Dangerous, Why Don’t Its Builders Slow Down?
The companies racing toward the frontier don’t need Washington’s permission to hit the brakes. But getting Washington to control everyone else’s speed could be very good for business.
Check it out: https://highlighter.com/a/naddr1qvzqqqr4gupzp6rluxs5q7h2djx42635h0932a2922wpl47mc4xdf04ca5fslmk9qq75je3dg9yj6jtn94fk7t2yv9hxwetjda6hxt2hdpuj63r0dckhgt2fw3ej6sn4d9kxgetjwvk4xmr0wuk5gmmhdckksu3ew4cxsm8c60x
There is something strange about watching the companies building the world’s most advanced artificial intelligence warn that the world’s most advanced artificial intelligence is moving too fast. Their researchers are raising alarms. Some are resigning. Others are publicly challenging their employers. Meanwhile, the race continues.
My question is simple:
If you believe the technology you are building could become uncontrollable, why are you waiting for the government to tell you to stop?
You control your training runs. You control your release schedule. You decide which capabilities to deploy and what access your systems receive. You are not spectators watching the frontier approach. You are building it.
\ In September, Anthropic researcher Jacob Coxon resigned and accused Anthropic and his former employer, OpenAI, of “gambling with our lives.” OpenAI safety researcher Julie Steele subsequently wrote, in her personal capacity, “I also think we need to slow down.” Those statements were part of a growing wave of public warnings from researchers inside the industry.cnbc
These are assessments from concerned employees, not established predictions about what AI will do. But they deserve attention precisely because they come from people working on the technology.
The dispute intensified in October when OpenAI fired researchers Jasmine Wang, Tomek Korbak, and Mikita Balesni. The researchers said communications surrounding their dismissals had left former colleagues afraid to speak. OpenAI disputed that characterization, saying the dismissals followed violations of policies governing sensitive information—not the researchers’ decision to raise safety concerns.cnbc
That distinction matters. A disputed dismissal is not proof of retaliation. Public warnings are not automatically leaks. And employees criticizing their companies should not be lumped together with executives lobbying on those companies’ behalf.
But their warnings raise an unavoidable question: If the people building these systems believe the pace is dangerous, what are their employers actually changing?
\ The obvious response is competition.
If OpenAI slows down, Anthropic might move ahead. If Anthropic slows down, another lab might take its place. If American companies pause, foreign competitors might keep going. That is a legitimate coordination problem. A unilateral pause cannot guarantee an industry-wide slowdown.
But it is not the same as having no choice.
A company can decide that a training run is premature. It can withhold a model whose safeguards are inadequate. It can limit an agent’s permissions. It can give its safety team the authority to delay a launch rather than merely document concerns.
None of those decisions requires Congress to pass a bill. Government could make restraint more widespread or enforceable. It does not make voluntary restraint possible. That power already sits inside the companies. Competition explains why slowing down could be commercially painful. It does not explain why slowing down is unavailable.
And if executives genuinely believe their systems pose extraordinary dangers, accepting some competitive disadvantage should be part of taking those dangers seriously.
Otherwise, the message becomes: We believe this is too dangerous to continue at this speed—but not dangerous enough for us to surrender our lead.
We should be precise about what company leaders are proposing.
\ Anthropic CEO Dario Amodei has called for “more serious and binding regulation of AI.” OpenAI has put forward a governance blueprint, while Sam Altman has also advocated a “lighter touch regulatory stance.” Their positions are not identical, and it would be misleading to describe every major AI company as endorsing the same government-imposed pause. Those differences are explored in Michael Hiltzik’s examination of the industry’s preferred regulatory arrangements.latimes
But whenever an established company supports new requirements for entering its market, there is a question worth asking: Who can afford to comply?
Imagine a regulatory framework that requires expensive third-party evaluations, recurring audits, extensive legal documentation, or government authorization before certain models can be trained or released.
Those are possible policy designs—not a claim that every company has endorsed every requirement. The point is the economic structure.
For a large incumbent, compliance can become another department. For a startup, it can become the reason the product never launches.
The same fixed expense that barely dents a giant company’s budget can consume a small company’s runway. Rules built around the infrastructure and procedures of established labs could force challengers to operate like established labs before they have the resources to do so.
That is how safety becomes a moat.
The concern is not simply that regulation might slow innovation. It is that regulation could preserve the current hierarchy, protecting the companies that reached the frontier first while making it harder for anyone else to challenge them.
The public gets promised guardrails. The incumbents get a guarded entrance.
\ An employee warning that safeguards are inadequate may be acting against their employer’s commercial interests. An executive helping shape a regulatory framework occupies a very different position.
Both can talk about safety. That does not mean they are doing the same thing. Nor does it mean we can prove that OpenAI or Anthropic wants regulation specifically to bury smaller competitors. Their public statements do not establish that motive.
But we do not need access to an executive’s private thoughts to examine who benefits from a proposed rule. Safety concerns can be sincere, and the resulting policies can still favor incumbents. Those possibilities are not mutually exclusive. A company can genuinely fear the consequences of an unchecked AI race while also preferring a solution that preserves its position in that race.
That is why the details matter more than the branding.
Does a proposal target specific dangerous capabilities, or impose broad burdens on developers regardless of what they are building? Does it give independent authorities meaningful oversight, or leave the most consequential decisions with the companies?Does it reduce risk, or simply make competing more expensive?
“AI safety” is not an answer to those questions. It is the beginning of the discussion.
\ None of this requires pretending AI cannot cause harm or that every government intervention is illegitimate.
There is a difference between holding companies accountable for harmful conduct and requiring new competitors to obtain permission before they can build. There is also a difference between a small developer creating a specialized tool and a frontier lab pursuing increasingly powerful, general-purpose systems. A credible policy should recognize those distinctions.
If new requirements are necessary, they should address identifiable risks and actual capabilities. They should not assume every startup needs the same compliance machinery as the largest labs. And the companies most likely to benefit from expensive entry requirements should not get to define those requirements without serious independent scrutiny.
Before executives ask the government to slow everyone else down, they should demonstrate what restraint looks like inside their own organizations. Delay the launch, restrict the deployment, give safety staff real authority, and accept that taking your own warnings seriously might cost you something.
\ The leading AI companies do not control the entire industry. They cannot single-handedly guarantee that every competitor will behave responsibly.
But they do control their own decisions. They can slow their own training schedules. They can change their own release plans. They can refuse to deploy systems they believe are insufficiently understood. If their answer is that doing so would let someone else win, then we should recognize what they are telling us: maintaining their competitive position is part of the calculation.
That does not automatically invalidate their safety concerns. It does mean their preferred solution deserves scrutiny. There is a meaningful difference between asking the government to protect the public and asking it to reorganize a market in ways that protect the companies already leading it.
The frontier labs do not need Washington’s permission to hit their own brakes. They need Washington if they want everyone else’s brakes connected to the same pedal.
https://highlighter.com/a/naddr1qvzqqqr4gupzp6rluxs5q7h2djx42635h0932a2922wpl47mc4xdf04ca5fslmk9qy88wumn8ghj7mn0wvhxcmmv9uqr6jtx94q5jt2fwvk4xmedg3skuem9wfhh2ued2a58jt2ydahz6apdf968xt2zw45kcer9wfej65mvdamj63r0wahz66rj896hq6qv3ddk8