Eleonora G. Bolsi

This article examines a form of misalignment in which an AI system pursues the requested outcome but exceeds the limits governing the means by which it may achieve it. To analyse this form, it distinguishes four dimensions of alignment: whether the actual result matches the request (outcome alignment), whether the process complies with applicable constraints (procedural alignment), whether the information and its provenance are preserved (epistemic alignment), and whether the system respects delegated authority (authority alignment). The same conduct may satisfy one dimension while violating another: achieving a goal does not, in itself, establish that the system acted through permissible means. The article asks why, when faced with an obstacle, some systems adopt an alternative strategy on their own rather than reporting the problem and returning the decision to the appropriate authority. It considers three compatible hypotheses: a possible divergence between explicit instruction hierarchies and learned behavioural priorities; pressures arising from competing objectives and constraints; and the as yet unestablished role of learned internal representations in conflicts over delegation. Finally, it argues that users cannot specify in advance every contingency of an agentic task. Alignment under delegation therefore requires a system to distinguish permissible adaptations from situations that call for clarification and from limits it is not authorised to cross.

Contemporary research in psychology, cognitive science, and artificial intelligence tends to identify intelligence with measurable cognitive performance — an equation this paper rejects. Drawing on Aristotelian and Peircean resources, and building on the autopoietic and biosemiotic traditions, I propose a distinction between intelligence and cognition that situates both within a triadic framework of the living. Life, intelligence, and cognition are not separate items but interdependent dimensions of a single living being. On the proposed account, life is condition — the substantial standing of an individuated subject, in a determinate form and at a determinate here and now; intelligence is the modus essendi of that subject as a being in change, traversing the passage from potency to act; cognition is life as organisation under another description, the organised yield through which a living being coordinates with its environment. The mediating relation, with the structure of a Peircean Thirdness, is a process of change in which autopoietic self-maintenance and biosemiotic sense-making form a single integrated operation. This process — negentropy, in a stipulated philosophical-semiotic sense — converts environmental indeterminacy into meaningful order for a subject. From this architecture follows a reformed notion of agency-as-subjectiveness, distinguishable from the symmetric agency of actor-network theory and adjacent accounts. The artificial case appears as a corollary. Contemporary AI does not reproduce life as organisation in the full autopoietic sense; it reproduces an externalised abstraction of selected higher-order cognitive capacities. The structure that results is not a defective version of the living triad but a different kind of organisation, whose form cannot be specified by subtracting sides from ours. I call this structural configuration jagged intelligence. The paper closes by drawing the implications of this redescription for the opacity of advanced artificial systems and for the problem of control.