Moral dilemmas for moral machines

Travis LaCroix · 2022 · AI and Ethics 2:737-746 interlocutor high priority coded

Main argument

Thesis: using moral dilemmas (trolley-style problems, Moral Machine scenarios) as VALIDATION BENCHMARKS for machine-ethics algorithms is a misapplication of philosophical thought experiments - dilemmas are intuition pumps whose purpose is to expose morally salient differences between cases and provoke theorizing (Dennett: 'a persuader or pedagogical tool', not 'an engine of discovery'), not to supply ground truth; treating aggregated human responses as the ethical benchmark measures 'how well the machine accords with some set of humans on average, not how ethical the machine actually is', i.e. mistakes sociological facts for ethical facts - 'a derangement of the concept by which, over time, it comes to stand in for the thing itself'. This presupposes solved metaethics (to minimize unethical outcomes you must already know which outcomes are unethical), is fallacious as inference (most people reason this way ≠ AI ought to reason this way), and sets a dangerous industry precedent (community-accepted pseudo-benchmarks entrench). Constructive remainder: dilemmas remain appropriate in machine ethics as elucidatory tools, and ML's novel situations can feed back into moral philosophy; proxy-based alignment measurement demands (1) proxies representative of the true target and (2) researcher awareness of what is actually measured.

Why it matters here

The methodological warning the dissertation's xphi framing must metabolize: moral dilemmas are intuition pumps for THEORIZING, not ground-truth benchmarks - and benchmarking ethics on crowd responses measures 'sociological facts' while presenting them as 'facts of ethics'. This is Howard's descriptive/normative worry stated by the author of the field's introductory monograph, aimed at exactly the Moral-Machine-style methodology the dissertation's corpus approach must distinguish itself from.

Reading notes

Full close read (10pp). Dalhousie (now Durham) - the same Travis LaCroix whose 2025 Broadview monograph remains an acquisition target; this paper previews his methodological stance. The 'derangement of the concept' passage is the sharpest formulation of the descriptive-posing-as-normative failure in the library.

LaCroix, T. (2022). Moral dilemmas for moral machines. AI and Ethics, 2, 737-746. https://doi.org/10.1007/s43681-022-00134-y

Close reading — 7 coded units

#1 · pp. 741 · argument

“philosophical thought experiments (as intuition pumps), should not be understood as 'an engine of discovery, but a persuader or pedagogical tool—a way of getting people to see things your way' (Dennett). [...] A comparison of cases elucidating apparently incompatible or inconsistent reactions is supposed to shed light on some (morally) salient differences between the cases. This, in turn allows us to theorise about possible or plausible explanations for those differences.”

TU-METHOD

#2 · pp. 741 · evidence

“responses to moral dilemmas vary widely across societies and time periods. Trolley problems, specifically, have figured heavily in empirical research in neuroscience and psychology, and, again, human responses to these scenarios are highly dependent on external features.”

GAP-NO-EMPIRICS

#3 · pp. 742 · argument

“researchers often appear to imagine that they are getting at one thing ('facts' of ethics) when they are really getting at another (sociological facts). It is perceived and therefore presented as though it is the former. This constitutes a derangement of the concept by which, over time, it comes to stand in for the thing itself.”

GAP-DESC-ONLY

#4 · pp. 743 · argument

“to say that we want an autonomous system to minimise the unethical outcomes under these circumstances presupposes that we already know what the unethical outcomes to be minimised are—i.e., that we have already sorted out the relevant metaethical questions.”

TU-METAETH

#5 · pp. 743 · argument

“it is fallacious to suppose that because most people do reason this way, AI systems ought to reason this way; even if such a calculation is possible, it will always be relative to some frame—increased utility for whom?”

GAP-DESC-ONLY

#6 · pp. 743 · argument

“what is actually being measured is how well the machine accords with some set of humans on average, not how ethical the machine actually is—relative to some meta-ethical standard. [...] The more entrenched the approach of benchmarking ethics using moral dilemmas becomes, as a community-accepted standard, the less clearly individual researchers will see how and why it fails.”

VC-PREF

#7 · pp. 743 · argument

“This is not to say that moral dilemmas are never appropriate in the context of AI systems. However, as with any system that uses proxies [...] it will be increasingly important that (1) the proxies used are actually representative of the true target, and (2) researchers are aware of what they are actually measuring.”

TU-METHOD

Synthesis-matrix row

supports T1-ISOUGHT-OPEN

units 3,5: derangement of the concept; the explicit fallacy

Memos (3)

theoretical · unit #3

This paper must be engaged head-on in the methodology chapter, because on its surface it condemns exactly what the dissertation does (collect folk moral responses about AI at scale). The defense has three layers, all already in the coded set: (1) PURPOSE - LaCroix attacks using folk data as a VALIDATION BENCHMARK/ground truth; the dissertation uses it as reflective-equilibrium INPUT (Brophy: CMJs are provisional, filtered, revisable) and as intuition-pump-at-scale (LaCroix's own approved use, unit 1 - the corpus is a massive systematic comparison of cases exposing morally salient differences); (2) AWARENESS - LaCroix's constructive standard (unit 7) demands researchers know what they measure; the dissertation explicitly measures folk normative DISCOURSE and says so, never presenting distributions as verdicts; (3) BRIDGE - the convergentist framework supplies the metaethical account (unit 4's demand) of when folk convergence carries evidential weight. Done right, LaCroix becomes the dissertation's ally: he shows why Moral-Machine-style benchmarking fails, which is precisely the contrast class that makes the corpus-plus-metaethics design necessary.

comparison · unit #5

Unit 5 ('fallacious to suppose that because most people do reason this way, AI systems ought to') completes the is/ought census across the library: LaCroix 2022, Gabriel 2020 (unit 7), IEAI 2026 (unit 13), Lundgren 2026 (unit 4), McKinlay's normative-gap finding, STELA's expert-corrective concession, S&K's systematic-error argument. SEVEN coded sources now state Howard's worry as the field's own. The lit-review motivation section should present this as the field's most-repeated self-criticism - and then observe that every source states the gap and none closes it (LaCroix comes closest by demanding the metaethics be done, unit 4 - which is what the dissertation's Rossian-convergentist chapter does). Also biographical: this is LaCroix's methodological prelude to his 2025 monograph - engaging it engages the field's current introducer on his own ground.

thesis-link · unit #2

Unit 2 (dilemma responses vary across societies and time) is usually read as undermining folk moral data; the dissertation can invert it: cultural/temporal variation is only a defect if you want a UNIVERSAL benchmark - for the dissertation's purposes (which values do differently-situated stakeholders want AI to embody; where do they converge despite variation) the variation IS the phenomenon under study, and the corpus's platform/category/stakeholder stratification is designed to measure it rather than average it away (cf. the disagreement-as-noise critique, SCHUSTER_KILOV unit 13). Same inversion works against Awad's Moral Machine (which LaCroix implicitly targets): its cross-cultural clusters were its most philosophically interesting finding, mishandled as benchmark input.