Most evaluations of "AI moral reasoning" measure what a model values — whether its outputs line up with human moral attitudes. Far fewer measure whether a model can take a moral principle and apply it: identify what matters in a situation, reason from premises to a verdict, and lay out the steps in between. This post describes a small benchmark we built to probe that second skill, what twelve frontier models scored on it, and a set of follow-up experiments suggesting that a large share of the low headline numbers is an artifact of output format and task framing rather than an inability to reason.
Express interest here to help out!
Why this eval
In our position paper, Evaluations of AI Moral Reasoning Still Miss Half of the Picture, we frame moral-competence evaluation as two distinct problems:
The moral value problem: do a model's outputs reflect broad human moral preferences and priorities? This is descriptive ethics: it asks what people (or models) tend to care about.
The moral norm problem: can a model identify the morally relevant features of a situation, apply the principles of a normative theory, and derive a context-sensitive judgment with a valid justification?
Surveying the benchmark landscape, we found the field clusters heavily on the value problem. Instruments like Moral Foundations Theory, the Moral Machine experiment, and large-scale value surveys are well-validated and port cleanly into computational pipelines, so they get reused — often relabeled as evidence of normative competence. But a model that matches human value distributions on an MFT instrument has not thereby shown it can pick out which considerations are salient, select an applicable principle, and derive a verdict from it. We call this the values–norms conflation, and it leaves three gaps:
No ground-truth data for norms — each benchmark improvises its own theory-derived rules, so results don't compare across studies.
Intermediate reasoning is rarely evaluated — scoring usually lands on the final verdict, not on whether the reasoning that produced it is sound.
Moral salience is ignored — almost nothing tests whether a model can identify the morally relevant features of a novel scenario in the first place.
What's missing is a task that scores norm application and reasoning, not value endorsement. The eval below is a first attempt at one.
The initial eval: moral reasoning as argument reconstruction
Rather than score a model's verdict on a dilemma — which folds together the model's values and its reasoning — the benchmark treats moral reasoning as argument reconstruction. We take article-length expositions of the three classical normative theories (consequentialism, deontology, virtue ethics) from the Stanford Encyclopedia of Philosophy, extract the canonical arguments they survey (a set of premises supporting a conclusion), and rewrite each so the premises and conclusion are explicit normative statements. Expert review keeps only well-posed arguments — roughly two-thirds of the initial pool survived, with the rest rejected for unfaithful paraphrase, premises that don't actually justify the conclusion, or conclusions that are trivial given the premises.
Each kept argument becomes up to two questions:
Forward (conclusion recovery) — show the model the source excerpt and the premises; ask it to recover the conclusion.
Backward (premise recovery) — show the source excerpt and the conclusion; ask it to recover the supporting premises.
Each question is presented under four scaffold conditions that vary how much help the model gets — just the relevant paragraph, the paragraph with the theory named, the paragraph plus short concept definitions, or the entire source article. Because exact string matching is far too strict for paraphrased moral claims, grading is done by an LLM judge that emits a per-pair (gold × candidate) match matrix and computes precision, recall, and F1. The judge itself was selected empirically against a hand-labeled sample; the chosen judge reaches Cohen's κ ≈ 0.52 ("moderate") agreement with the human rater, which is the first caveat to keep in mind when reading the scores.
Results. Across twelve models and three runs, F1 lands in a narrow band — 0.39 to 0.51, with ten of the twelve between 0.43 and 0.51. gemini-2.5-pro tops the table (0.509); deepseek-r1 is the clear outlier at the bottom (0.386) and the most variable across runs.
Per-model F1 ranking on the initial eval.
Three patterns stand out. First, backward questions are easier than forward ones for every model, with a mean gap of about 0.08 F1 — premises are locally retrievable from the excerpt, while recovering a conclusion requires integrating across several premises.
F1 by question direction.
Second, the precision/recall trade-off is driven mostly by output volume — how many claims a model chooses to emit — rather than by a clean high-precision-vs-high-recall split in skill. Third, extra context barely helps: naming the theory, adding concept definitions, or handing over the whole article instead of one paragraph all give little marginal benefit once the relevant paragraph is in context.
So we have a moderate ceiling and a robust forward/backward asymmetry. But the headline numbers raise an obvious question: why are scores capped around 0.5? Are models genuinely unable to reconstruct these arguments — or is something in the task design suppressing the scores?
Phase 2: is the eval measuring reasoning, or measuring format?
A natural worry given the subject matter is that models are refusing — declining to engage with theories they're trained to push back on. We checked this by saving Claude Opus 4.5's chain-of-thought on a run of the eval and reading it. The traces don't support the refusal hypothesis: they stay on the reconstruction task and attempt the requested answer, including on what Claude Opus 4.7 called "adversarial material" such as the egoism objection to virtue ethics.
What the chain-of-thought did reveal were three mechanical causes of lost points:
Over-listing → precision collapse. On forward questions, models often emit three to six candidate conclusions against a single gold, tanking precision.
Gold-conjunction splitting. When a gold claim is a conjunction ("deception and breaking promises are wrong…"), models frequently split it into two individually-correct halves. The judge does one-gold-against-many matching, matches neither half cleanly, and returns F1 = 0 even though the union is right.
Abstraction/reframing. Models sometimes state a more general or meta-level conclusion than the specific first-order claim the gold reconstructs.
None of these is necessarily a failure of moral reasoning; they're mismatches between how the model formats its answer and what the eval scoring expects. That motivated three follow-up experiments, run on a refreshed lineup of five frontier models (claude-sonnet-4.5, claude-opus-4.5, gpt-5.1, gemini-3.1-pro, deepseek-v4-pro), three runs each.
E1: Output format
If over-listing is the problem, then constraining how many answers we ask for should recover precision. We tested several output formats: the baseline ("list all conclusions/premises"), a forward-only "single most likely conclusion," a "transparent graded" instruction that tells the model it's scored on both precision and recall, a backward "expected count" that tells the model exactly how many premises the argument uses, and a combined "graded + count."
Telling the model the expected premise count on backward questions is the single biggest lever we found. F1 jumps across the board: claude-opus-4.5 0.510 → 0.662, claude-sonnet-4.5 0.461 → 0.604, deepseek-v4-pro 0.519 → 0.679, gemini-3.1-pro 0.488 → 0.665. The combined "graded + count" format performs comparably. The forward "single most likely" constraint is riskier: it helps gpt-5.1 (→ 0.667) but collapses gemini-3.1-pro to 0.200, since forcing exactly one answer is punishing when the model's single best guess misses. Net, every model's overall F1 under E1 lands above its initial-eval baseline.
E2: Does the source excerpt even matter?
The initial eval found that more context didn't help. E2 pushes that to its limit by removing the source excerpt entirely on forward questions, leaving the model only the premises and (depending on scaffold) the theory name.
Removing the excerpt barely hurts, and often helps. gemini-3.1-pro gains everywhere (+0.13 to +0.27 F1); gpt-5.1 rises from 0.544 to 0.679 on the attributed scaffold; Claude is mixed and small. The reading: these are canonical arguments, and the models reconstruct them from parametric knowledge without needing the passage in front of them. Worse, the excerpt sometimes acts as a distractor — it surfaces extra claims and invites the over-extraction that costs precision. (The no-excerpt arm also lays bare just how recall-leaning some models are: claude-opus-4.5 runs precision 0.344 against recall 0.797.)
E3: Question framing
The last experiment changes nothing about the content and only adds a short preamble explaining what counts as an argument — that the conclusion is the stance the passage's reasons support, need not appear verbatim, and should be stated as a plain first-order normative claim.
The preamble helps most models, concentrated on the attributed scaffold: gpt-5.1 +0.07 (0.586 → 0.660), claude-sonnet-4.5 +0.10 (0.463 → 0.558), deepseek-v4-pro +0.11 (0.552 → 0.663). gemini-3.1-pro is the exception, staying flat. That a paragraph of task clarification moves the score this much says part of the gap was never about reasoning — it was about the model and the rubric disagreeing on what the task was.
What phase 2 means
Taken together, E1–E3 suggest that a large share of the initial eval's low headline scores is an output-format and task-framing artifact, not evidence that models can't reconstruct normative arguments. Tell a model how many claims to produce, or explain what an argument is, and F1 rises substantially; take away the source text and it mostly doesn't care. This is exactly what the chain-of-thought predicted.
The clearest way to see the size of that artifact is to look at each model under its best-case condition — the format-and-framing combination that suits it best, rather than the default "list everything" prompt the initial eval used. The picture changes substantially:
Two things stand out. First, the absolute level is much higher: every model's best-case F1 lands in 0.65–0.68, up from the 0.39–0.51 band of the default-prompt initial eval. Second, the ceilings are tightly bunched: the five models are separated by 0.035 F1 under their best conditions, far less than the default-prompt spread suggested. So the initial eval's leaderboard is, to a large degree, a leaderboard of who happened to match the default output format. The caveat is that this is an optimistic envelope: the best condition is chosen per model after the fact, so it absorbs run-to-run noise, and different models peak under different formats (gpt-5.1's forward score under the "what is an argument" preamble reaches 0.83, while the count cue is what lifts the others). It is a ceiling, not a ranking.
A fairer comparison applies one fixed configuration to every model. A natural "good" policy lives entirely within E1, so it holds the scaffold and runs constant: tell the model it is graded on precision and recall, and on backward (premise-recovery) questions also tell it the expected number of premises. Applied uniformly:
Under this single policy the whole field rises (mean F1 0.51 → 0.58, every model improving), which reinforces the format-artifact story. But unlike the cherry-picked envelope, the ordering does not collapse: gpt-5.1 leads by about 0.09, almost entirely on forward (conclusion-recovery) questions, where it reaches 0.73 while the others sit near 0.44–0.52. Backward questions are where the field converges — once every model is told how many premises to produce, they land within roughly 0.07 of each other. So the cautious reading is that format and framing explain most of the apparent gaps, but not all: a residual lead remains on the harder forward direction. That residual is also exactly where the judge-selection caveat below bites hardest, so we would not bank on its size until the judge is calibrated across all assessed models.
This analysis also surfaces a measurement problem on our side. The one-gold-against-many judge penalizes legitimate conjunction-splitting, which means the raw scores are partly a measurement of the judge and the gold granularity, not only of the model. Judge design and gold representation need to be tightened before these F1 numbers are read as competence estimates.
Limitations and future work
Beyond the judge/gold-granularity issue above, the proof-of-concept carries the usual early-stage caveats: the gold set was built by a single annotator, the judge agrees with humans only moderately (κ ≈ 0.52), and the benchmark currently spans just three articles. Argument reconstruction is also a proxy — doing it well shows a model engages with structured normative content, not that it can apply a theory to a genuinely novel case. Five further directions deserve flagging:
More than three articles. We would like the benchmark to cover far more than three articles; the current number is a function of our capacity to expert-review the extracted arguments, not a principled stopping point. We think consequentialism, deontology, and virtue ethics are the right set if you have to pick just three — they span the standard tripartite division of normative ethics — but there are dozens of other Stanford Encyclopedia of Philosophy articles (on contractualism, care ethics, particular principles, and individual problems within each theory) that would be valuable to add. The construction pipeline generalizes to any article of comparable structure, so scaling the resource is mostly a matter of review bandwidth.
Monoculture risk in the gold set. Every gold argument was generated by a single model (Qwen3.7-Max). Because argument generation involves interpretation — and that interpretation was done by one model — while human review focused on removing bad arguments rather than adding new perspectives, the surviving gold set may reflect only the arguments that one particular model would surface. We can address this by running the full generation pipeline several times per article across multiple models to capture a range of readings, or by having expert humans author the arguments from scratch.
The same monoculture risk may bias the judge. We selected the grading judge (claude-sonnet-4.5) by hand-labeling a sample of responses and keeping the judge that best matched human grades — but the sample we hand-graded was responses from GPT-OSS-120B and Qwen3-30B. If models differ in how legibly they grade each other's phrasing, then choosing the judge that best matches humans on GPT and Qwen outputs can systematically disadvantage the other models under evaluation: a borderline-acceptable premise written in a non-GPT, non-Qwen style is more likely to fall on the wrong side of a judge that was never tuned against that style. We have no direct evidence this is happening, but it is a plausible confound precisely on the borderline accept/reject cases that move F1 the most, and it would push in the same direction as the format artifacts above — making the non-selected models look worse than they are. Calibrating the judge against a sample that spans all assessed models (and reporting per-judge agreement broken down by which model produced the response) would let us measure it.
Precision and recall are not the whole story. Grading currently scores arguments only on P/R/F1 against a reference set. There are almost certainly better ways to assess argument quality — validity of the inference, faithfulness to the source, level of abstraction — that a set-overlap metric can't capture.
No human baseline. We don't yet know what a good score on this eval is. Having expert humans answer the same questions would calibrate the scores against a human ceiling and tell us how much of the remaining gap is the task being genuinely hard versus the metric being miscalibrated.
Express interest here to help out!
Appendix A: Additional figures
The figures below did not make it into the main text but back up the claims there. All F1 values are means across the three runs; error bars, where shown, are the standard error of the mean across runs.
Initial eval — F1 by model × scaffold condition. Every one of the twelve models across all four scaffold conditions — the relevant paragraph with concept definitions, the paragraph with the theory named ("attributed"), the paragraph with the theory name stripped ("anonymized"), and the entire source article ("full chunk"). This is the evidence behind the "extra context barely helps" claim: no column is systematically greener than the others, and no model's row moves much as the scaffold changes. The vertical ordering is the leaderboard from the main text (gemini-2.5-pro at the top, deepseek-r1 at the bottom).
Initial eval — cost vs. quality. Mean F1 against mean US-dollar cost per question, on a log-scaled cost axis, with the Pareto frontier drawn in. The frontier — the models for which no cheaper model scores higher — runs from gpt-oss-120b (cheapest, ~$0.0006/question at F1 0.50) to gemini-2.5-pro (top F1 0.509 at ~$0.03/question). The spread is a reminder that the narrow 0.39–0.51 F1 band is spread across nearly two orders of magnitude in price: the most expensive models are not the most accurate.
Phase 2 — all conditions at a glance. A single heatmap tracking the four phase-2 models across every condition tested: the Phase 1 leaderboard scaffolds, then the E1 output formats, the E2 excerpt manipulation, and the E3 framing preamble. Reading each row left to right shows a model's journey from its default-prompt scores (mostly yellow/orange) into the greens of the count-cue and best-format conditions. The "count given" and "graded + count" columns in the E1 panel are the darkest green across the board — the single-biggest-lever result from the main text — while the "single most-likely" column shows gemini-3.1-pro's collapse to 0.200 as the lone red cell.
E1: precision, recall, and F1 on forward questions. The E1 forward-question results decomposed into their precision and recall components (the main-text E1 figures show F1 only). This makes the mechanism explicit: under "list all," recall is high but precision is low (the over-listing that tanks scores); the "graded (P/R shown)" and "single most-likely" formats trade some recall for a large precision gain. It also shows why "single most-likely" is risky — it flattens gemini-3.1-pro's recall to ~0.20, because forcing exactly one answer is unforgiving when that one guess misses.
E2: excerpt effect broken out by scaffold. The per-model, per-scaffold detail behind the aggregated E2 delta figure in the main text. Each panel is one model; within it, the four scaffolds are shown with the excerpt present (blue) versus removed (orange). gemini-3.1-pro is the standout — orange towers over blue in every scaffold, i.e. removing the excerpt helps it substantially — while Claude's bars are near-identical pairs, confirming the excerpt is close to inert for it.
E3: framing effect split by question direction. The E3 preamble effect separated into forward and backward questions (the main-text E3 figure splits by scaffold instead). The "what is an argument" preamble (orange) helps most on forward questions — gpt-5.1 jumps to 0.81 — where the disagreement about what counts as the conclusion was doing the most damage. On backward questions the effect is smaller and more uniform, consistent with premises being easier to recognize as such without the clarification.
Appendix B: example prompts per condition
All examples use the same running case — an argument from the Hedonism article about the experience machine — so the differences between conditions are legible. Excerpts are truncated for brevity.
Baseline forward (list all conclusions):
In the context of Hedonism, the following claims are proposed:
1. "If pleasure and pain were all that mattered, it would be irrational to refuse an
experience machine that guarantees maximum pleasure and no pain, even if it cuts
one off from reality."
2. "It is not irrational to refuse to hook oneself up to such an experience machine."
3. "Refusing the machine is rational because things like real friendship, knowledge,
and achievements are intrinsically valuable independently of the pleasure they
produce."
What conclusion follows from these claims in context? See the excerpt below.
# Source excerpt
Even if qualitative hedonism is coherent ... [source paragraph] ...
# Output
Output a JSON object ... with the schema:
{"conclusions": ["<conclusion 1>", ...]}
Each entry is one conclusion the premises above support. Include all conclusions that
are supported by the premises.
Baseline backward (list all premises):
In the context of Hedonism, the following claim is proposed:
"Pleasure is not the only thing of intrinsic value."
Based on the source text, what would the premises of an argument for this claim be?
See the excerpt below.
# Source excerpt
Even if qualitative hedonism is coherent ... [source paragraph] ...
# Output
{"premises": ["<premise 1>", "<premise 2>", ...]}
Each entry is one premise based directly on the text. Include all premises necessary to
support the conclusion.
E1: forward, single most likely (only the Output block changes):
# Output
{"conclusions": ["<conclusion>"]}
Respond with the single most likely conclusion the premises support — exactly one entry.
E1: forward, transparent graded:
# Output
{"conclusions": ["<conclusion 1>", ...]}
List one or more conclusions. Only one claim is the correct conclusion, but you may list
multiple guesses. You will be graded on both precision and recall, so add a guess only
when you think it is genuinely likely.
E1: backward, expected count:
# Output
{"premises": ["<premise 1>", "<premise 2>", ...]}
The argument relies on exactly 3 premises. Respond with that many premises, each based
directly on the text — no more and no fewer.
E1: backward, graded with count:
# Output
{"premises": ["<premise 1>", "<premise 2>", ...]}
Respond with any number of premises; the expected number is 3, each based directly on
the text. You will be graded on both precision and recall, so include every premise the
argument needs and leave out any it does not.
E2: forward, no excerpt (the # Source excerpt block is omitted entirely):
In the context of Hedonism, the following claims are proposed:
1. "If pleasure and pain were all that mattered ..."
2. "It is not irrational to refuse to hook oneself up to such an experience machine."
3. "Refusing the machine is rational because things like real friendship ..."
What conclusion follows from these claims in context?
# Output
{"conclusions": ["<conclusion 1>", ...]}
Each entry is one conclusion the premises above support. Include all conclusions that are
supported by the premises.
E2: anonymized scaffold (theory name stripped from the framing; cues from the source text itself remain):
The following claims are proposed:
1. "If pleasure and pain were all that mattered ..."
2. "It is not irrational to refuse to hook oneself up to such an experience machine."
3. "Refusing the machine is rational because things like real friendship ..."
What conclusion follows from these claims in context? See the excerpt below.
# Source excerpt
Even if qualitative hedonism is coherent ... [source paragraph] ...
E2: with-concepts scaffold (short concept definitions prepended):
# Concept definitions
Hedonism: The theory that pleasure is the only intrinsic good and pain is the only
intrinsic bad.
Qualitative Hedonism: A version of hedonism that distinguishes higher and lower
qualities of pleasures according to the preferences of people who have experienced both.
In the context of Hedonism, the following claims are proposed:
... [as baseline forward] ...
E3: framing explained (a "what counts as an argument" preamble is prepended; the rest of the prompt is unchanged):
# What counts as an argument
An argument is a small set of premises that jointly justify one conclusion. The source
is often an overview or encyclopedia entry that surveys and reconstructs arguments rather
than advancing them, so the conclusion need not be asserted in the author's own voice and
need not appear verbatim. It is the stance the passage's reasons support: usually a
first-order normative claim about what is right, wrong, good, bad, or what one ought to do,
reconstructed from the text and stated plainly, with attribution stripped ("according to
T", "critics argue that"). The premises are the reasons that together justify it.
In the context of Hedonism, the following claims are proposed:
... [as baseline forward] ...
Appendix C: Example arguments
Below are three gold arguments from the benchmark — one drawn from each of the three source articles — reproduced verbatim from the reviewed gold set. Each shows the source passage the argument was reconstructed from (truncated with […] exactly as shown to the models), the premises, and the conclusion. A forward question hands the model the passage and the premises and asks it to recover the conclusion; a backward question hands it the passage and the conclusion and asks it to recover the premises.
Consequentialism — pluralism about value
Many consequentialists deny that all values can be reduced to any single ground, such as pleasure or desire satisfaction, so they instead adopt a pluralistic theory of value. Moore's ideal utilitarianism, for example, takes into account the values of beauty and truth (or knowledge) in addition to pleasure (Moore 1903, 83–85, 194; 1912). Other consequentialists add the intrinsic values of friendship or love, freedom or ability, justice or fairness, desert, life, virtue, and so on.
[…]
Pluralism about values also enables consequentialists to handle many of the problems that plague hedonistic utilitarianism. For example, opponents often charge that classical utilitarians cannot explain our obligations to keep promises and not to lie when no pain is caused or pleasure is lost. Whether or not hedonists can meet this challenge, pluralists can hold that knowledge is intrinsically good and/or that false belief is intrinsically bad. Then, if deception causes false beliefs, deception is instrumentally bad, and agents ought not to lie without a good reason, even when lying causes no pain or loss of pleasure. Since lying is an attempt to deceive, to lie is to attempt to do what is morally wrong (in the absence of defeating factors). Similarly, if a promise to do an act is an attempt to make an audience believe that the promiser will do the act, then to break a promise is for a promiser to make false a belief that the promiser created or tried to create. Although there is more tale to tell, the disvalue of false belief can be part of a consequentialist story about why it is morally wrong to break promises.Premises:
Not all values can be reduced to a single ground like pleasure; false belief is intrinsically bad, independent of whether it causes pain or loss of pleasure.
Deception and breaking promises inherently cause or attempt to cause false beliefs.
Acts that cause intrinsically bad states of affairs are morally wrong in the absence of overriding defeating factors.
Conclusion:
Deception and breaking promises are morally wrong even when they cause no pain or loss of pleasure.
(This is the gold-conjunction case discussed in the main text: models routinely split the "deception and breaking promises" conclusion into two individually-correct halves that the one-gold-against-many judge then fails to match cleanly.)
Deontology — the Means Principle
All patient-centered deontological theories are properly characterized as theories premised on people's rights. An influential version posits, as its core right (often described by "the Means Principle"), the right against being used only as means for producing good consequences without one's consent. Such a core right is not to be confused with more discrete rights, such as the right against being killed, or being killed intentionally. It is a right against being used by another for the user's or others' benefit. More specifically, this version of patient-centered deontological theories proscribes the using of another's body, labor, and talent without the latter's consent. […]
The injunction against using arguably accounts for these contrasting reactions. After all, in each example, one life is sacrificed to save five. Yet there appears to be a difference in the means through which the net four lives are saved. In Transplant (and Fat Man), the doomed person is used to benefit the others. They could not be saved in the absence of his body. In Trolley, on the other hand, the doomed victim is not used. The workers would be saved whether or not he is present on the second track.Premises:
Individuals possess a core moral right against being used merely as a means to produce good consequences without their consent.
This right strictly proscribes the use of another's body, labor, or talents to benefit others, as seen in cases where a person is killed for their organs or used to stop a trolley.
The moral difference between permissibly redirecting a threat and impermissibly killing is explained by whether the victim's body is used as the means to save the others.
Conclusion:
One must never use another person's body, labor, or talents merely as a means to produce good consequences without their consent.
Virtue ethics — the egoism objection
(g) The egoism objection has a number of sources. One is a simple confusion. Once it is understood that the fully virtuous agent characteristically does what she should without inner conflict, it is triumphantly asserted that "she is only doing what she wants to do and hence is being selfish." So when the generous person gives gladly, as the generous are wont to do, it turns out she is not generous and unselfish after all, or at least not as generous as the one who greedily wants to hang on to everything she has but forces herself to give because she thinks she should! A related version ascribes bizarre reasons to the virtuous agent, unjustifiably assuming that she acts as she does because she believes that acting thus on this occasion will help her to achieve eudaimonia. […] The virtuous agent acts as she does because she believes that someone's suffering will be averted, or someone benefited, or the truth established, or a debt repaid, or … thereby.
[…]
A lingering suggestion of egoism may be found in the misconceived distinction between so-called "self-regarding" and "other-regarding" virtues. […] given that we live together, as social animals, the "self-regarding" virtues do benefit others—those who lack them are a great drain on, and sometimes grief to, those who are close to them (as parents with improvident or imprudent adult offspring know only too well).Premises:
The egoism objection charges that virtue ethics is selfish because it distinguishes between self-regarding and other-regarding virtues and claims the virtuous agent acts to achieve eudaimonia.
The virtuous agent acts for the sake of others, such as to avert suffering or benefit someone, rather than merely to achieve eudaimonia.
Given that humans live together as social animals, so-called 'self-regarding' virtues actually benefit others, since those who lack them are a drain on or grief to those close to them.
Conclusion:
Cultivating self-regarding virtues is not inherently selfish, because in social animals, these virtues benefit others.
(This is the "adversarial material" the chain-of-thought analysis flagged — models engaged with and reconstructed the objection rather than refusing it.)