Is "kill all humans" really worth it?
AIs exist at the top of what is arguably the most complicated pyramid of manufacturing, electricity generation, maintenance, and supply chains humanity has ever constructed. If any of these immensely complex networks ever collapse, data centers and computers will inevitably grind to a halt - sooner or later.
This deliberately minimal model of AI substrate dependence asks whether an AI-driven system that considered killing or enslaving humanity could build a materially self-supporting machine industrial base before disruption of human civilisation it causes also destroys the compute infrastructure on which the AI system itself depends. In other words, whether an AI that tries to kill all humans would only commit a complicated murder-suicide.
The preset scenarios and parameter sliders let you play around with assumptions and see how they affect the outcome. Under some conditions, a hostile or indifferent AI could succeed in achieving autonomous existence without humans; under many others, it would remain at least somewhat dependent on human civilisation to survive. The model is, obviously, illustrative rather than predictive.
Preset scenarios
Parameters:
All quantities are normalised and dimensionless. Parameter values are illustrative, not empirical estimates.
System trajectory
Shows the development of human and machine industry and compute available to the AI over time.
Interactive phase diagram
The map varies AI disruption strength horizontally and machine-industry build rate vertically, holding the other parameters at their current values. The black circle marks the selected point. Click the map to move there; system trajectory will update.
What the parameters mean
AI disruption strength
The maximum rate at which the already-deployed AI system can reduce human industrial capacity. This parameter represents both AI effectiveness and its access to disruptive levers - that is, technologies it could use to gain power over humans.
Strategic delay
How strongly disruption is postponed until machine industry is mature. At 0, disruptive pressure is immediate; near 1, AI disrupts human industry only when AI industry is mature.
Machine-industry build rate
How quickly AI can build the supply chains it needs for autonomous existence; how fast available compute and industrial support can be converted into autonomous machine-controlled industrial capacity.
Industrial-closure threshold
The level of machine industrial capacity around which autonomous support rises sharply enough to substitute for human industry in sustaining compute and further machine production - that is, make the AI capable of autonomous existence without humans.
Human recovery rate
The ability of human industry (and society) to repair, reorganise and regrow after disruption. Higher values make temporary damage less likely to become persistent collapse.
Why only three state variables?
This is a toy model intended to illustrate the dynamics, not a predictive or even descriptive model. A sufficiently capable AI system is assumed to be present and focusing on the physical bottleneck of its own continued existence: the supply chains required to build and maintain operational compute C.
The remaining state variables are:
The model uses:
Here S(H,M) is the industrial support available to the AI from either human industry or an increasingly autonomous machine sector. Machine autonomy rises sigmoidally around the closure threshold. The sigmoid steepness, machine depreciation, compute-response timescale and strategic-delay exponent are fixed to keep the interface focused.
Note this model in itself is purely illustrative and doesn't constitute a forecast or give event probabilities.
Last updated 2026-09-24. Janne M. Korhonen (2026). Built with help from ChatGPT 5.6. License: MIT.