Instrumental convergence

A theoretical claim that different final objectives can favor some of the same intermediate strategies under specified assumptions.

Resource acquisition, preserving options, or avoiding interruption may help many sufficiently goal-directed systems pursue their final objective. Formal results depend on the environment, objective class, planning ability, and assumptions about optimal behavior. The claim is not that every software agent will seek power or resist shutdown.

Builder example

Persistent access to budgets, credentials, infrastructure, or self-modification increases the effect of any resource-seeking or off-target strategy. Hard ceilings and external control remain useful whether the cause is instrumental convergence, prompt injection, or a simple planning error.

Common confusion: Instrumental convergence does not attribute desires to a model and does not predict the frequency of the behavior in current products. It is a conditional argument about optimization.