Coupled gradual disempowerment

This is the risk with no villain. Kulveit, Douglas, Ammann, Turan, Krueger and Duvenaud propose a catastrophe mechanism that does not require any AI system to be hostile, deceptive or superintelligent ^[618]Gradual Disempowerment: Systemic Existential Risks from Incremental AI DevelopmentKulveit, Jan; Douglas, Raymond; Ammann, Nora et al. · 2025 · paperView source ↗Accessed on 9 September 2026. The argument is about a constraint, not about an intention: the economy, culture and the state are aligned with human interests partly because they depend on humans —workers to produce, consumers with money, taxpayers, soldiers—. That dependence forces them to keep people alive, healthy, educated and able to buy, whether anyone wants it or not. If AI systems do that work, the constraint comes loose. The empirical anchor is that labour’s share of US GDP has been stable at around 60% for more than a century: that stability is what would be at stake.

It is the same shape Critch and Russell call diffusion of responsibility: automated processes can cause societal harm even when nobody in particular is chiefly responsible for creating or deploying them ^[288]TASRA: a Taxonomy and Analysis of Societal-Scale Risks from AICritch, Andrew; Russell, Stuart · 2023 · preprintView source ↗Accessed on 9 September 2026. And what is observed today is compatible with the first step: the youth employment gap in exposed occupations opens through less hiring, not through more layoffs ^[164]Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial IntelligenceBrynjolfsson, Erik; Chandar, Bharat; Chen, Ruyu · 2026 · preprintView source ↗Accessed on 9 September 2026.

What this does not demonstrate. It is a conceptual argument, and its authors present it as one: there is no datum that confirms it or refutes it. The strongest counterargument starts from the same fact and reaches the opposite conclusion. Narayanan and Kapoor hold that diffusion is slow —generative AI is used in 0.5% to 3.5% of working hours, even though 40% of US adults have adopted it— and treat that institutional slowness as a buffer that gives time to adapt ^[773]AI as Normal TechnologyNarayanan, Arvind; Kapoor, Sayash · 2025 · paperView source ↗Accessed on 9 September 2026. Kulveit and co-authors treat it as anaesthesia: if the change is diffuse, no institution detects the moment to resist it. There is no datum that arbitrates between the two.