← All sources

Disentangling AI Alignment: A Structured Taxonomy Beyond Safety and Ethics

Kevin Baum · 2026 · AISoLA 2024 (LNCS 16032, Springer), pp. 158-173   background medium priority coded

Main argument

Thesis: 'alignment' conflates multiple legitimate configurations, and the field needs a structured taxonomy distinguishing (1) the normative AIM (safety, ethicality, legality, user intent, cultural appropriateness...), (2) the STANDARD X of that aim (e.g. utilitarianism vs Scanlonian contractualism as standards of ethicality), (3) SCOPE (outcome alignment - the result is acceptable - vs execution alignment - how it was achieved is acceptable), and (4) CONSTITUENCY (individual vs collective judges of alignment). Argument type: conceptual analysis with graded definitions. Key results: safety defined as robustness against harmful malfunctions (absent malicious influence) in foreseeable intended contexts - deliberately allowing a system with harmful intended purpose (weapons) to count as safe; safety does NOT entail ethicality, but ethicality practically implies safety; X-ethicality relativizes to an external moral standard, sidestepping both moral skepticism and the AIA-moral-agency question by focusing on behavior; the parameterized X,Y-alignment definition with perfect/realistically-optimal/sufficient gradations; all-things-considered alignment across mutually exclusive standards may be logically impossible; the proceduralist consensus (fair public process) is best understood as a candidate for ALL-THINGS-CONSIDERED alignment, not moral alignment.

Why it matters here

The definitional toolkit for the lit review's opening: parameterized X,Y-alignment (standard X of normative aim Y) with graded sufficiency, the safety/ethicality distinction with entailment analysis, outcome-vs-execution scope, and individual-vs-collective constituency. Also reconstructs the proceduralist argument as an explicit practical syllogism (P1-P6) - the cleanest statement of the inference the dissertation contests, with Baum's own observation that its conclusion is ambiguous between approximating moral truth and replacing it.

Reading notes

Full close read (16pp, extracted AISoLA chapter; same DFKI group as Steingrüber & Baum, same volume as Kästner). Baum is the RAIME group lead. Note: assumes AIAs are 'not, and will not in the foreseeable future be, moral agent[s], but rather tool[s] with some technical autonomy owned by a human moral agent' - another published anchor for the anti-moral-agency stance.

Baum, K. (2026). Disentangling AI Alignment: A Structured Taxonomy Beyond Safety and Ethics. In B. Steffen (Ed.), AISoLA 2024 (LNCS 16032, pp. 158-173). Springer. https://doi.org/10.1007/978-3-032-01377-4_8

Close reading — 9 coded units

#1 · pp. 161 · definition
“An AIA's safety is a strictly monotonically increasing function of its robustness against harmful malfunctions (absent malicious external influences) in foreseeable and intended application contexts.”
#2 · pp. 162–163 · definition
“We sidestep these concerns by (1) defining ethicality relative to an externally defined moral standard X (acknowledging its possible imperfection) and (2) focusing on observable AIA behavior rather than internal moral agency: An AIA's X-ethicality is a strictly monotonically increasing function of the proportion of its behavior under all intended application contexts consistent with moral standard X.”
#3 · pp. 164–165 · argument
“These AI systems deserve to be called safe—but that does not force us to claim that their use in an unjust war would be morally justified. Stretching the notion of safety to encompass moral permissibility would dilute its conceptual clarity [...] safety does not entail ethicality. [...] while not a conceptual necessity, it is at least a practical regularity that ethicality implies safety.”
#4 · pp. 166 · argument
“[The proceduralist practical syllogism:] (P1) Moral pluralism (at least when understood descriptively) is a fact. (P2) There is deep moral uncertainty and persistent moral disagreement. (P3) Imposing values should be avoided. (P4) If P1-P3 and we must nevertheless align AIAs, we should aim for alignment relative to publicly justifiable norms rather than moral alignment. (P5) We must align AIAs. (P6) Publicly justifiable norms can (only? best?) be achieved through a fair and public process [...] (C) We should aim for alignment relative to norms that are the result of such a fair and public process.”
#5 · pp. 166 · argument
“It remains somewhat unclear whether this approach ultimately aims to approximate moral ground truth via public procedures, or rather represents a shift in normative aims—from moral alignment to public justifiability understood as an independent aim.”
#6 · pp. 167 · argument
“it may be permissible to design an AIA—which is not, and will not in the foreseeable future be, a moral agent, but rather a tool with some technical autonomy owned by a human moral agent—in such a way that it takes its owner's goals and intentions into account in a proportionate way.”
#7 · pp. 167–168 · definition
“Given some normative aim Y, an AIA's X-alignment with respect to Y is a strictly monotonically increasing function of the proportion of its behavior under all X-relevant application contexts that is consistent with Y-normative standard X. [...] we can now say that an AIA is more or less ethicality-aligned in the utilitarian or Scanlonian sense.”
#8 · pp. 168 · argument
“achieving perfect conformity with alignment targets may be practically infeasible (when considered with respect to a specific standard X) or even logically impossible (when considered in an all-things-considered sense). [...] the various plausible standards—especially when drawn from different normative domains—may impose mutually exclusive requirements.”
#9 · pp. 169 · definition
“[Scope:] besides the outcome, the way it was achieved—the execution—matters as well. Adam may have reserved the table by bribing the staff or threatening the shift manager. Such behavior might violate important social or moral constraints—even if the outcome is aligned. [Constituency:] the entity or entities with respect to whom alignment is evaluated or from whom the normative aim ultimately (directly or indirectly) originates—e.g., an individual user, a group, or society at large.”

Synthesis-matrix row

supports T3-PROCEDURALISM-INCOMPLETE
P1-P6 syllogism reconstructed; ambiguity diagnosis
supports T5-AGENCY-DENIED-EVALUABILITY-KEPT
X-ethicality sidesteps agency; tools owned by human agents

Memos (3)

theoretical · unit #4
Unit 4's syllogism is a gift: the proceduralist consensus (Gabriel, Zhi-Xuan, G&K, Schuster-Kilov) finally stated as an explicit inference with numbered premises - which means the dissertation can locate its dissent PRECISELY at P4. The convergentist counter: P1-P3 are all true, but P4 is a non-sequitur, because there is a third option between imposing one theory and retreating to procedure - aligning with verdicts on which the plural reasonable theories CONVERGE imposes nothing on anyone (each constituency can accept the verdict from its own premises) while remaining genuinely MORAL alignment, not just public justifiability. Baum's own observation (unit 5) that proceduralism is ambiguous between truth-approximation and aim-replacement strengthens this: convergentism resolves the ambiguity by keeping moral truth-directedness while earning public justifiability as a byproduct. The metaethics chapter should quote the syllogism verbatim and argue against P4.
comparison · unit #2
Unit 2 (X-ethicality: external standard + behavior only) completes a trio of independent behavior-relative moves: MILLIERE fn1 (behavioral framing to avoid the moral-values-attribution question), SANWOOLU (constrained not accountable), BAUM (X-ethicality sidesteps agency). The field has quietly converged on evaluating AI behavior against external normative standards WITHOUT settling moral agency - which means Augustine's experiment (which settles the agency question empirically for the negative) supplies the warrant these behavior-relative frameworks presuppose but leave undischarged. Also unit 6: Baum flatly asserts AIAs are tools owned by human moral agents 'for the foreseeable future' - assert-vs-argue again.
theoretical · unit #9
The outcome/execution scope distinction (unit 9) upgrades the responsibility taxonomy: KAESTNER's difference-makers and G&K's misalignment modes both concern outcomes; Baum shows EXECUTION misalignment (right result, wrong means - bribing the maître d') is a distinct failure class. For agentic systems this is central: multi-step agents choose their own means, so execution alignment is precisely what increases with agency (cf. HELLRIGEL_DUNG unit 11) - and execution wrongs may have NO outcome trace (the bribe succeeded, everyone's happy). Who is responsible for unobserved wrongful means? Connects to Kästner's type-(ii) difference-makers (only interpretability reveals the execution path) and to the folk corpus: check whether folk comments distinguish outcome-blame from means-blame - a codable distinction the AI Responsibility category may already contain.