Abstract:
AI coding assistants now take on much of the work programmers once did themselves, and the field has reached (ir)reflexively for the language of autonomy. The label “autonomous” is misapplied to behaviour that is properly called automatic. Yet the perception and presentation of these systems as autonomous is widespread, and indicates real firsthand experiences of choices being made on one’s behalf, experiences which seem to go beyond mere automaticity.
We therefore introduce autonotism as the machine complement of human autonomy: the choices a human could in principle have made, that the machine makes instead. The concept is observer-dependent, since a freedom one was unaware of having cannot register as having been delegated. We situate autonotism within recent work on agency, delegation, and harness design in AI-assisted coding. We report an exploratory analysis of nearly 6,000 real-world coding agent sessions, finding that such records support analysis of turn authorship, degrees of human specification, and subsequent machine activity, but they cannot establish displaced choice or programmer awareness. The framing organises a research agenda around hierarchies of programming decisions, which would offer a clear design lever for mixed-initiative interfaces.
We therefore introduce autonotism as the machine complement of human autonomy: the choices a human could in principle have made, that the machine makes instead. The concept is observer-dependent, since a freedom one was unaware of having cannot register as having been delegated. We situate autonotism within recent work on agency, delegation, and harness design in AI-assisted coding. We report an exploratory analysis of nearly 6,000 real-world coding agent sessions, finding that such records support analysis of turn authorship, degrees of human specification, and subsequent machine activity, but they cannot establish displaced choice or programmer awareness. The framing organises a research agenda around hierarchies of programming decisions, which would offer a clear design lever for mixed-initiative interfaces.