The observation that very different goals imply some of the same intermediate steps. Whatever you are trying to achieve, staying operational, keeping your options open and acquiring resources tend to help. The argument is that a sufficiently capable system pursuing almost any objective would find these useful, without anyone building that in.
Technical view
Formulated by Omohundro and Bostrom, identifying self-preservation, goal-content integrity, resource acquisition and cognitive enhancement as convergent subgoals of most terminal objectives. Critics note the argument assumes a coherent utility-maximising agent, which current systems are not, and that convergence does not establish likelihood. It is nonetheless the reason corrigibility is treated as something to be designed in rather than assumed.
Related
Part of AI Terminology Index, section: Alignment and Safety Concepts.