ONTOLOGICAL FLATNESS AND THE ALIGNMENT PROBLEM: How a Mismodelled Ontology of Action Blinds AI Alignment Research

The AI alignment problem is standardly treated as a technical challenge. This paper argues that it has a prior, unexamined philosophical dimension. Alignment research operates under a tacit monistic ontology of action—a frame- work in which goal representation, decision-making, and execution are treated as continuous moments of a single optimisation process. Drawing on Lakatosian philosophy of science, this paper identifies that ontology as the un- questioned hard core of the alignment research programme and demonstrates that it generates systematic theoretical blind spots: phenomena such as deceptive alignment are not merely difficult to solve within the programme but are in part constituted by its foundational ontological commitments. As a philosophical remedy, the paper introduces a three-layer ontology of action distinguishing intention (durative teleological goals), will (discrete situational decisions), and execution (deliberate and automatic realisation). The central contribution is the operationalisation of will as an evidentiary parameter of intention: an agent’s decision pat- terns provide defeasible inductive evidence that a stated intention is operationally held, independently of outcome measurement. This framework does not replace the technical agenda of alignment research; it reconstructs the onto- logical foundations upon which that agenda rests, thereby expanding what the field can theorise, detect, and address.