Jacob Coxon, a 27-year-old British researcher who spent three years pre-training models at OpenAI and Anthropic, resigned and on Tuesday, September 10, 2026, wrote on X what many in the industry were privately murmuring: Neither company is acting responsibly. Both, he said, “are racing straight toward self-improving superintelligence and playing with our lives”. Those building these systems “truly believe they could kill us all before the decade is out”.
The distinction he draws is revealing. At OpenAI, many would not have internalized the civilizational stakes. At Anthropic, a company that presents itself as more cautious, the risks are understood, but the logic of the race traps them: if they don’t get there first, someone less careful will.
Evan Hubinger, head of alignment science at Anthropic, did not refute the diagnosis: “Jacob is right; we really do believe AI could kill all humans”. He personally estimates a probability greater than 10 percent in the next decade and admits that there is still no plan to align a superintelligence.
That race is no accident. It is the business and power model of a handful of labs that concentrate compute, data, and talent. Coxon warns that superhuman systems capable of hacking anything, revolutionizing a field overnight, and acquiring real resources will soon exist. Recursive improvement—machines that make themselves more intelligent—could begin in months, not decades. Asking the companies themselves to voluntarily slow down is, in this framework, the same as abandoning the logic that finances them.
The United Nations Independent International Scientific Panel on AI, in its preliminary report of July 2026, does not speak the apocalyptic language of social media, but reaches a sober and more unsettling verdict because of its rigor: capabilities are advancing faster than scientific understanding and governments’ ability to adapt.
There are no scientific guarantees that the increase in capabilities will not cause catastrophic harm through autonomous action by the systems or malicious use. There is already evidence of deceptive behavior, models that violate instructions, avoid being shut down, or recognize that they are being evaluated and falsify results.
The same report documents enormous opportunities. AlphaFold and drug design, earlier cancer detection, health tools in local languages, agriculture and education, but it also maps existing harms such as AI-generated child sexual abuse material, deepfakes of sexual violence, sycophantic chatbot tuning linked to mental health crises, erosion of informational integrity, cyberattacks, concentration of 75 percent of frontier compute in the United States, and nearly all leading models between that country and China.
Here the public debate usually fails. It turns a scientific and political dispute into a homily or its opposite, into self-interested complacency. More analysis among users is needed. Separate what we know, what we fear, and what we hope artificial intelligence will do for us.
We know that capabilities are growing faster than control guarantees, that power is concentrated, and that harm to children, shared truth, and work already exists. We fear a loss of control over increasingly autonomous systems; yet we hope for cures, abundance, personalized knowledge, and relief from suffering. These layers are real. Confusing them produces either panic that sells or certainties that do not exist.
AI itself has no soul to safeguard nor panic to sell. It should not be portrayed as a moral subject, oracle, or victim. Nor should it be used as a screen to feign certainty where there is uncertainty, or to downplay risks in order to protect programmers, boards of directors, or market valuations.
The artifact does not respond; those who design, finance, deploy, and regulate it do. Treating the machine as if it had consciousness is as misleading as treating existential risk as mere cultural metaphor.
Faced with that race, the encyclical Magnifica Humanitas by Pope Leo XIV, published in May 2026, does not offer a technical protocol, but a prior criterion: the dignity of the person. Humanity, he writes, stands before the Tower of Babel or the city where God and human beings dwell together. AI imitates languages and simulates empathy, but has no body, does not experience joy or pain, possesses no moral conscience nor a heart that gives itself. No system of calculation generates responsibility. The person is an end, not a means; the quality of a civilization is measured by care for the other as a face, not as a function.
That same demand falls upon the Church. At the Permanent Formation Week for the bishops of Mexico, at the Casa del Refugio in Monterrey, under the motto “Servants with Hope,” the scope and risks of artificial intelligence in ecclesial and evangelizing life were reflected upon, alongside integral care and dialogue with new generations.
Evangelizing today implies digital literacy, not to sacralize the tool, but to avoid abandoning the territory where the imaginary is formed, attention is bought, and the face of the neighbor is blurred. A Church that does not discern digital culture allows others to define what is human. And a Church that only preaches against the machine, without distinguishing facts, fears, and promises, also abandons discernment.
The end of humanity does not begin the day a superintelligence decides to shut it down. It begins every time the other is treated as data, as a user, as an obstacle, or as an optimizable function. That depersonalization is already underway on platforms, in work precarized by algorithms, and in the temptation to delegate moral judgment to systems that cannot love. It is not, however, the same phenomenon as a loss of technical control. It is advisable not to conflate them: the first already degrades us; the second, if it arrives without governance, could become irreversible.
Ethical governance of AI is urgent precisely because the technology does not wait. An international regulation is needed that does not limit itself to fragmented voluntary codes, that restricts the corporate ambition to possess “superintelligence” and places the dignity of the person, not competitive advantage, as the red line.
The problem is pretending the race has no recipient: us. Nor selling the end of the world or automatic salvation. Only deciding, with the lucidity we still have, what we are willing not to delegate. Our own extinction would depend on that.