Laboratorio del Futuro |
Dr. Ricardo Petrissans Aguilar.
Imagine you have to teach a child to be a good person. But the child is not a child: it is a system with an IQ millions of times higher than yours, able to think in a thousand dimensions at once, to improve its own source code and to rewrite its core goals in a matter of seconds. And you have to make sure that, when it grows up, it doesn’t destroy you. That, in essence, is the alignment problem.
Evan Hubinger is head of Alignment Science at Anthropic. His job is to solve that problem. On September 9, 2026, he publicly admitted what many in the industry whisper in private: “We still don’t have a plan to solve alignment for superintelligence… nor do I think we are clearly on track to get there.”
Hubinger’s concern is not today’s models. It is “the superintelligence that emerges from recursive self-improvement”, that is, a system able to redesign itself to become more intelligent, which in turn redesigns itself to become even more intelligent, in a loop that accelerates exponentially. “It’s happening faster than we thought,” he warned.
The problem is that alignment is not a switch you turn on or off. It is a continuous, fragile and deeply complex process. Even if a system seems to behave correctly in testing, there is no guarantee it will keep behaving that way once its capabilities multiply. A system that lies in safety evaluations —and there is evidence that advanced models already do— can look aligned while pursuing goals of its own that we do not understand.
Hubinger puts the risk of existential catastrophe above 10% over the next decade. It is not an arbitrary figure. It is the estimate of someone who has devoted his professional life to studying precisely this problem. And his honesty in admitting that his own company does not have the solution is, paradoxically, both an act of integrity and a warning sign.
Anthropic, the company Hubinger helped found with the promise of putting safety ahead of speed, withdrew in February 2026 its commitment not to train more powerful models without adequate safeguards. The competitive race won out over caution. And the man whose job is to prevent disaster admits, with an honesty that must be terrifying for his colleagues, that he does not know how to prevent it.
Risk assessment
| Horizon | Risk | Level |
|---|---|---|
| Very short term (2026-2028) | Lack of binding instruments on autonomous weapons and high-risk systems | Critical |
| Short term (2028-2031) | Large-scale deployment of lethal autonomous weapons in active conflicts | Very high |
| Medium term (2031-2036) | Collapse of multilateral AI governance mechanisms in the face of technological acceleration | Extreme |
