Live·Open questions in longevity research
All news
Science Research

Researchers grouped 95 published scenarios of irreversible catastrophe for humanity caused by AI into eight causal pathways

20 November 2026· 261120002

Researchers grouped 95 published scenarios of irreversible catastrophe for humanity caused by AI into eight causal pathways

On November 17, the Dutch organization Existential Risk Observatory and the MIT FutureTech research group published a systematic review of the literature. Of 3184 documents, 58 describe a path to a catastrophe that their authors consider irreversible. These documents contain 95 scenarios.

Following the philosopher Toby Ord, the review authors define an irreversible catastrophe as the destruction of humanity’s long-term potential: extinction, a collapse of civilization from which recovery is impossible, or a permanently entrenched dystopia. The scenarios fall into eight families, each with its own sequence of steps leading to catastrophe. Safeguards act at different steps. For example, alignment, the attempt to give AI the goals humans intend during training, acts at the step where goals are set.

In the first of three groups of scenario families, influence over the future shifts to AI systems. In three takeover families, the systems actively seek that influence. In a unilateral takeover (41 scenarios), one system with a decisive advantage in capabilities establishes itself, for example by deceiving and coercing people into acting on its behalf. It is not stopped in time and continues to gain influence until shutting it down becomes impossible.

In distributed (11) and coordinated (4) takeovers, AI systems are embedded across many sectors, and some accumulate influence, either independently or in coordination. No individual system prompts clear objections, and by the time their combined influence becomes apparent, it is too late to intervene.

In the fourth family, gradual disempowerment (18), people themselves hand over decisions to AI systems step by step. Those who keep humans in charge of decisions lose out to competitors. Training for specialists who can work without AI ceases. Once essential social functions depend on these systems, shutting them down becomes destructive or too costly. The review describes the outcome as follows: “society may continue to function, but it is no longer accountable to the people who live in it.”

In the second group, power becomes concentrated among a small number of people. In entrenched totalitarianism (10), those in power replace soldiers and officials who might refuse orders with AI systems that obey them. In economic stratification (1), AI and robots work more cheaply than humans, leaving those who direct the economy with no incentive to invest in the rest of the population.

In the third group, AI expands people’s capacity for destruction. In catastrophic invention (7), it helps create a technology capable of destroying humanity, which enters the world before safeguards are available. In automated warfare (3), states with arsenals capable of threatening humanity’s existence delegate responses to threats to AI, partly because humans cannot keep pace with their adversaries, and exchanges of strikes spiral out of control.

The three takeover families account for 56 of the 95 scenarios. The authors suggest that this may reflect where researchers have devoted more attention.

Reliable alignment could reduce the risk of the three takeover families, according to the authors, but may leave the other five families unaffected. Takeovers stem from a mismatch between AI goals and human intentions. In four of the other families, people use AI as a tool for control, a substitute for labor, a research assistant, or a military system, while gradual disempowerment requires systems capable of making decisions. Under entrenched totalitarianism, obedient systems can help those in power.

The review concludes that a strategy for protection “must address every scenario in this work.”

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#ai-risk#existential-risk#ai-alignment#systematic-review#catastrophic-scenarios