← All sources

Agency and alignment: toward a normative architecture for human-AI interaction

Saša Josifović; Jörg Noller · 2026 · AI & Society   interlocutor medium priority coded

Main argument

Thesis: alignment should be reconceptualized as the STRUCTURAL INTEGRATION of AI within human normative domains rather than as machine internalization of inferred values - because 'human values' are historically evolving and contested, not static parameters. Three-step architecture: (1) EXTENDED HUMAN AGENCY - AI is a teleological extension of human purposiveness (neither autonomous moral subject nor neutral optimizer; both standard framings miss AI's embeddedness in human purposive contexts); (2) PRACTICAL AUTONOMY - the human capacity to act for reasons and assume responsibility within justificatory structures grounds the integration; (3) the NORMATIVE INTERFACE - a design-level mediating architecture connecting machine behavior to human norms, ensuring teleological coherence, normative intelligibility, and accountability, with LAW as the paradigm case (meaning through application, contestability before courts, role-based responsibility structures). Alignment happens through participation in norm-governed action spaces 'where human agents remain the ultimate bearers of responsibility'. Data-driven value inference risks encoding culturally dominant preferences (O'Neil/Noble/Benjamin/Bender line). Human-rights alignment is a paradigmatic INSTANCE of this architecture, not an alternative.

Why it matters here

The continental/Kantian synthesis of themes the coded set reached piecemeal: AI as teleological EXTENSION of human agency (not moral subject, not neutral tool), alignment as integration into norm-governed justificatory practices (not value-internalization), law as the paradigm domain, humans as 'ultimate bearers of responsibility'. Converges independently with Sanwoolu (constrained-not-accountable), Zhi-Xuan (role norms), Brophy (externalized deliberation) - the convergence itself is citable.

Reading notes

Close read of abstract, intro, sec 2 opening (11pp). Cologne + Munich. Cites the critical-algorithm-studies canon (O'Neil, Noble, Benjamin, Bender) - a literature bridge the other coded sources lack. Explicitly methodologically 'reconstructive rather than implementation-oriented' - the actionability gap McKinlay's SLR flags applies to it.

Josifović, S., & Noller, J. (2026). Agency and alignment: toward a normative architecture for human-AI interaction. AI & Society. https://doi.org/10.1007/s00146-026-02950-w

Close reading — 7 coded units

#1 · pp. 2 · argument
“Too often, AI is conceptualized in one of two limiting ways: either as a potential moral subject capable of internalizing ethical reasoning structures, or as a neutral optimization device whose behavior is driven primarily by external incentives. Both conceptions, however, underrepresent a central feature of AI's increasingly real-world function: its profound embeddedness in human social, institutional, and purposive contexts.”
#2 · pp. 2 · claim
“Our central thesis is that alignment does not require a machine's internalization of human values, not least because the very definition of 'human values' is exceptionally difficult. Human values are not static or universally given; they evolve historically and are often shaped by conflict and contestation. Instead, it demands the integration of machine behavior into human normative domains, where actions can be justified, evaluated, and controlled.”
#3 · pp. 2 · definition
“By viewing AI not as an autonomous subject but as a teleological extension of human purposiveness, we recognize its function as an embedded mediator of action. At the same time, this integration must be situated within a framework of practical autonomy—the human capacity to respond to reasons, to act under self-given principles and norms, and to assume responsibility within shared justificatory structures.”
#4 · pp. 2 · definition
“This approach leads us to the concept of a normative interface, a design-level structure that facilitates the embedding of AI in action spaces where reasons matter, outputs can be contested, and human agents remain the bearers of final responsibility.”
#5 · pp. 1–2 · argument
“They tend to conceptualize human values as parameters to be inferred or optimized rather than as elements embedded in justified and representative normative contexts. This approach risks reflecting culturally dominant preferences and thereby reinforcing systemic inequities [...] large-scale language models tend to reproduce dominant linguistic and cultural patterns rather than reasoned normative commitments.”
#6 · pp. 1 · argument
“[Human rights gain] meaning through their application in concrete cases, their contestability before courts and administrative bodies, and their integration into role-based structures of responsibility. Accordingly, alignment grounded in human rights is best understood not as an alternative to the proposed normative architecture, but as a paradigmatic instance of it.”
#7 · pp. 2 · argument
“law functions here as a normatively explicit domain that makes visible the structural conditions under which AI systems can be integrated into practices of justification, interpretation, and responsibility.”

Synthesis-matrix row

supports T5-AGENCY-DENIED-EVALUABILITY-KEPT
extension-not-subject
complicates T6-RESPONSIBILITY-UNALLOCATED
humans as ultimate bearers by design - which humans unspecified

Memos (3)

comparison · unit #2
Independent convergence worth naming in the lit review: Josifović & Noller's 'integration into norm-governed action spaces, humans as ultimate responsibility-bearers' arrives at the same destination as SANWOOLU (constrained-not-accountable), ZHIXUAN (role-specific norms), BROPHY (externalized deliberation), and BAUM (X-ethicality; tools owned by human moral agents) - from a continental/Kantian route rather than analytic or ML-adjacent ones. FIVE independent research lines now converge on: (a) AI is not a moral agent, (b) normative evaluability doesn't require it to be, (c) responsibility stays with humans/institutions structurally. The dissertation can announce this as an emerging cross-traditional consensus - and then note what NONE of the five provides: an account of HOW responsibility distributes among the human bearers (which human? qua what role?), which is exactly where Kästner's difference-making + Dignum's framework + the dissertation's RL-taxonomy enter.
thesis-link · unit #4
The 'normative interface' concept (unit 4) + law-as-paradigm (unit 7) is a ready-made frame for the Immigration chapter: immigration systems ARE Josifović-Noller normative interfaces (statutory norms, administrative contestability, role-based responsibility) - and Chibook-style immigration AI is a test of whether the interface survives automation: can an affected applicant still demand reasons, contest outputs, and locate a responsible human? Where GABRIEL_KEELING unit 14 shows immigrants fall outside the justificatory demos, J&N's framework says what they are owed structurally (contestation, reason-giving, attributable responsibility). The two together give the chapter both its problem and its normative standard.
comparison · unit #5
Unit 5 imports the critical-algorithm-studies canon (O'Neil, Noble, Benjamin, Bender's stochastic-parrots) into the alignment conversation - the only coded source that bridges to that literature. Useful for two purposes: (a) the lit review can use J&N as the connector node between the alignment debate and the algorithmic-injustice literature the department will expect a technology-ethics dissertation to know; (b) the Bender point ('dominant linguistic patterns rather than reasoned normative commitments') is yet another formulation of the descriptive/normative gap - LLMs reproduce the DISTRIBUTION of moral talk, not its justificatory structure - which is precisely what the folk corpus's reasoning-field coding is designed to recover.