There is something profoundly absurd about the contemporary race for artificial intelligence: some of the human beings who best understand the potential danger are sitting inside the laboratories accelerating toward it.
We are not talking about technological prophets writing from a cabin or amateurs terrified by ChatGPT. We are talking about frontier AI researchers, people who train the models, study their behavior, try to solve their alignment, and know from the inside the capabilities that emerge as scale increases. And some of them are saying something extraordinary: we may be building systems whose control we do not know how to guarantee.
Jacob Coxon has just left Anthropic after also having worked at OpenAI. His accusation deserves to be read without the apocalyptic pyrotechnics of certain headlines: he argues that both companies are advancing toward a self-improving superintelligence and that they are “gambling with our lives.”
What happened afterward is what is truly uncomfortable.
Evan Hubinger, a researcher who remains at Anthropic, publicly backed an essential part of the warning. According to his statement, he personally assigns a probability of more than 10% to AI causing human extinction within the next decade. And he acknowledges something even more important: there is currently no demonstrated plan that solves the alignment problem of a future superintelligence.
Let us stop there.
More than ten percent.
If an aeronautical engineer said that an airplane had more than a ten percent chance of killing all its passengers, no one would argue about whether the correct percentage was ten, nine, or twelve. The plane would not take off.
But if the experimental object is a potentially planetary technology, it seems we have invented a new ethics: first we take off and then we try to build the wings.
The moral obscenity does not lie in researching artificial intelligence. Nor does it lie in automatically assuming the extinction hypothesis to be true, which remains precisely that: an extreme and disputed hypothesis, not a scientifically demonstrated outcome.
The obscenity begins elsewhere.
It begins when someone seriously believes that there is a meaningful probability of causing an irreversible catastrophe and, even so, considers it acceptable to continue increasing the system’s capabilities while waiting for another department to solve the control problem.
Here a question arises that is far more uncomfortable than any fantasy about killer robots:
what responsibility do researchers bear?
Because “I only do research” is a phrase that the history of science has heard too many times already.
Moral responsibility does not disappear because an organization has divided the work into teams, benchmarks (standardized tests), safety departments, layers of authorization, and confidentiality agreements. Bureaucratic fragmentation can distribute a task. It does not magically distribute its consequences.
If those who possess privileged technical knowledge genuinely believe that we are approaching a potentially uncontrollable threshold, remaining silent ceases to be neutrality.
It becomes a decision.
And continuing to work does too.
Naturally, there are researchers inside these companies precisely because they want to prevent disaster. This is an essential distinction. Anthropic was born, to a large extent, around the problem of AI safety and maintains teams dedicated to alignment, interpretability, and risk assessment. It would not be intellectually serious to lump together the researcher trying to build safeguards and the executive who wants to win a commercial race.
But neither can we indefinitely accept the perfect paradox:
“We must keep building an increasingly powerful machine because we need an increasingly powerful machine to research how to prevent the increasingly powerful machine from becoming dangerous.”
At some point, the reasoning begins to eat its own tail.
And then the other question appears. The most unpleasant one.
Who authorized this gamble?
I did not.
You probably did not either.
No world assembly voted on what percentage of extinction risk it considers acceptable. There was no planetary referendum. There is no parliament of humanity that has decided that 1%, 5%, or 10% constitutes a reasonable price for achieving artificial general intelligence.
Yet a handful of private corporations possess the chips, the data centers, the models, much of the scientific talent, and the billions of dollars needed to push the frontier forward.
Humanity stakes the planet.
They place the bet.
And so we arrive at a politically legitimate suspicion, but one that we must formulate precisely as a suspicion and not disguise as a proven fact.
The economic and technological elites have Plan Bs.
We know that some billionaires have purchased isolated properties, built extraordinarily sophisticated shelters, or spoken openly about private survival strategies in the face of catastrophes. That phenomenon has been documented since long before the current explosion of AI.
What we do not know —and should not claim without evidence— is whether the leaders of the major artificial intelligence companies possess a coordinated plan to survive a catastrophe caused by their own systems.
But the mere existence of radical inequality in survival raises a formidable political problem.
Because even without any conspiracy, risks are not distributed democratically.
Someone who possesses hundreds of millions or billions can acquire land, independent energy, water, private security, satellite communications, underground facilities, international mobility, and logistical redundancies.
Most of humanity has a door with a lock.
That is why the question needs no conspiracies.
It is worse without them.
What happens when those making decisions capable of producing planetary risks have an incomparably greater capacity to protect themselves from the consequences of those very decisions?
Here we encounter a classic problem of moral hazard: whoever decides how much risk to assume does not necessarily bear the same exposure as those upon whom the harm will fall.
The worker in Dhaka, the teacher from Valparaíso, the farmer in Senegal, or a family in Gaza do not own shares in Anthropic, OpenAI, or Google DeepMind. They did not participate in their boards of directors. They did not choose the pace at which their models are trained. They do not know their internal safety assessments.
But they would share the consequences if the worst-case scenario ever came true.
And there is something even more perverse.
Competition between companies can transform even reasonable people into agents of an irrational dynamic.
OpenAI cannot stop because Anthropic might get ahead.
Anthropic cannot stop because Google DeepMind might get ahead.
Google cannot stop because China might get ahead.
The United States cannot stop because China might get ahead.
China cannot stop because the United States might get ahead.
And thus we arrive at the perfect mechanism of collective irresponsibility:
everyone knows it would be advisable to slow down, but no one can slow down because no one trusts that the others will do the same.
We do not need a rebellious artificial intelligence to understand this phenomenon.
Human beings invented it a long time ago.
It is called an arms race.
The difference is unsettling: during the nuclear arms race, governments understood approximately what a bomb was, what it destroyed, and how deterrence worked.
Here, the specialists themselves debate what capabilities will emerge, when they will appear, and whether we will be able to control them afterward.
We are racing toward something whose final form we do not know, while those holding the steering wheel acknowledge that they are still designing the brakes.
And that is precisely why Coxon’s resignation matters.
Not because he has demonstrated that we will all die in 2030.
He has not demonstrated that.
It matters because it shatters a moral comfort.
It forces those who remain inside to answer a question that can no longer be resolved through papers, corporate statements, or new “safety” departments:
If you truly believe there is a significant possibility that what you are building could escape human control, what level of danger would have to be reached before you refused to keep building it?
Twenty percent?
Fifty?
Ninety?
Where is the line?
Because if there is no line, then there is not really a precautionary principle either.
There is only a race.
A runaway race toward superintelligence that still has no reins, driven by companies competing to get there first and watched by researchers who, every so often, step out of the building to warn us that the vehicle may have no brakes.
The problem is that we are on board too.
Except no one asked us whether we wanted to get on.