For years, the artificial intelligence industry had an almost automatic answer to a simple question: how far can we go?
The answer was, essentially, as far as technology allows us to.
Each new generation of models pushed the frontier a little further. More reasoning capability, more context, better tools, and greater autonomy. Companies competed to build the next system before their rivals, while governments began to treat that capability as a matter of national power.
Now the conversation is beginning to change.
Dario Amodei, CEO of Anthropic, has publicly argued that the development of the most advanced models should proceed more cautiously and that we need mechanisms capable of evaluating their capabilities and risks before allowing the race to continue without restrictions.
The concern is reasonable. A system that answers questions is one thing. One capable of executing code, using tools, interacting with external systems, and maintaining objectives over extended periods is another. The greater a model's ability to act, the harder it becomes to anticipate all the ways it might behave.
But accepting that concern immediately leads us to a more difficult question:
Who decides when we have gone too far?
That is the point where technological safety begins to intersect with competition.
Anthropic is not a regulatory body. Neither are OpenAI, Google, Meta, or the other laboratories developing frontier models. They are the main participants in the race. For that reason, when a company proposes establishing limits on certain capabilities, the proposal should be evaluated on its technical arguments, but also through mechanisms that allow those arguments to be independently verified.
Not because their warnings are necessarily wrong, but because a decision of this kind can directly affect who is allowed to continue developing the technology and under what conditions.
The problem becomes even clearer when we stop looking only at companies.
A Race That Is Already Geopolitical
Artificial intelligence is not being developed within a single market. The United States and China are competing for semiconductors, computing infrastructure, specialized talent, and the ability to develop advanced models.
U.S. restrictions on certain technologies destined for China are part of that competition. From Washington, they can be justified as national security measures intended to prevent sensitive capabilities from being used for strategic or military purposes. From Beijing, they can be interpreted as an attempt to preserve the United States' technological advantage.
The two interpretations are not necessarily incompatible.
A restriction can reduce a risk while simultaneously making it harder for a competitor to reach the same level of capability. That is precisely why it is difficult to completely separate security policy from technology policy.
Recent Chinese criticism of certain U.S. proposals concerning AI reflects this tension. Beijing has challenged the logic of technological containment and called for greater international cooperation, while its own laboratories continue developing increasingly capable systems.
We do not need to decide which position is correct to identify the problem: the same word — security — can describe very different objectives depending on who establishes the rules.
What If Slowing Down Were Also a Strategy?
Here, a possibility emerges that deserves consideration without turning it into an accusation.
Regulation that limits certain capabilities may be necessary to reduce risks. But it can also increase the cost of competing at the technological frontier.
This matters especially because the most advanced models require extraordinary amounts of computing power, infrastructure, and capital. If new regulatory requirements disproportionately affect certain actors, they could ultimately consolidate the position of those who already have the resources necessary to comply with them.
That does not mean regulation is a bad idea.
It means that safety should not be evaluated solely by what it intends to achieve, but also by the structures it creates around it.
A rule can reduce an immediate danger while producing, as a secondary consequence, a more concentrated market.
That is why the question should not simply be whether we need regulation. We probably do.
The question is how to design it so that safety does not accidentally become a barrier to entry or a tool of geopolitical competition.
The Real Problem: Defining the Limit
Suppose that tomorrow a model appears with capabilities far beyond those of current systems.
Who decides whether it can be deployed?
The company itself could do so through its internal evaluations. That would be the fastest and technically simplest option, but hardly sufficient when the consequences affect third parties.
A government could do it. That would provide legal authority, although it would necessarily introduce national security and economic competition considerations.
An international body could be created. In theory, this would offer greater independence, but getting the major powers to accept the same rules would be extraordinarily difficult.
We could also combine all three levels: independent technical evaluation, government oversight, and international mechanisms for certain risks.
The latter option may be less elegant than searching for a single authority, but it may be precisely the reality we need to accept.
Because the problem is not simply knowing whether a model is dangerous.
It is establishing a credible procedure for deciding that it is.
A New Kind of Race
For a long time, we imagined the artificial intelligence race as a competition to build the most powerful model.
Now another, less visible but potentially equally important race is beginning to emerge: the race to define the conditions under which those models will be allowed to exist.
Who establishes the standards.
Who conducts the evaluations.
Which capabilities require restrictions.
What level of evidence is sufficient to stop a deployment.
Which countries are subject to those rules.
And what happens when one of the main competitors decides not to accept them.
In that context, slowing down can be a perfectly legitimate safety measure. It can also have economic and strategic consequences that no one should ignore.
We do not need to assume bad faith to recognize that.
Nor do we need to dismiss the risks of AI in order to question who establishes its limits.
In fact, both discussions should move forward together.
If the risks are important enough to justify restrictions, then the evaluations supporting those restrictions should be open to examination by independent actors. If the measures have strategic consequences, they should be discussed as such rather than presented exclusively as technical decisions. And if regulation can significantly alter competition, its effects on the market should also be part of the conversation.
Artificial intelligence is reaching a point where the question is no longer only who will succeed in building the most capable system.
It also matters who will succeed in defining the rules of that race.
Because controlling the risk of a technology is one thing.
Controlling who is allowed to develop it is something very different.
And it is still unclear which of the two we are beginning to do.
t474-r0b07
T474::AUTH
AI::ASSISTED
HUMAN::DIRECTED
ANTI_HYPE::016
Top comments (0)