NOTE-092: Quillien — Luck and intentional action

Contribution

The paper extends Quillien & German’s (2021) causal definition of “intentionally” to the luck effect, the finding that successes due to chance are judged less intentional. It derives from counterfactual theories of causal judgment a set of qualitative predictions: skill-independent luck effects, objective over subjective probability, alternative-option effects, the Kraemer means/ends asymmetry, and moral override. It then tests whether causal judgments of the form “E because the agent wanted E” move with intentionality judgments. The new empirical contributions are: (i) the Knobe (2003) marksman pattern holds for causal judgments; (ii) the Kraemer effect appears in causal judgments only in the morally bad vignette; (iii) a predicted Knowledge × Process interaction, which the know-how and probability-raising theses do not predict (note 18: they are “agnostic”).

Key insight

Calling an outcome intentional is, on this account, a causal judgment about the agent’s desire. That judgment is graded, and luck weakens it by making the desire–outcome link non-robust: it is easy to imagine the agent wanting E and E failing. The folk concept of intentional action is therefore “externalist”: it depends on how mental states connect to the world, not on the mental states alone. The paper compares it to the folk concept of knowledge in Gettier cases (Conclusion).

Assumptions

• The folk “A did X intentionally” is roughly equivalent to “X happened because A wanted X” (statements (1)–(2)). This is a simplification of Quillien & German (2021), which speaks of “attitude” and adds constraints against deviant causal chains (note 1).
• Causal judgment is graded and counterfactual. People simulate alternatives and weigh necessity and robustness (Causation and Counterfactuals). Counterfactual simulation is “anchored” in the actual world: in a mix-up, people imagine other mix-ups (note 17).
• The probability-raising thesis is read non-teleologically, as whether the action in fact raised the odds (note 7).
• The mediation analyses assume Condition → Causation → Intentionality is the right causal model. They are “confirmatory”, not discovery (Study 1 Results).
• “Morality” is a label for a vignette contrast confounded with difficulty (note 8; Study 2 Methods), and the paper does not rely on it as a clean manipulation.
• Samples: US Prolific workers with 50–1000 prior studies and ≥90% approval, excluded for failing comprehension or a white-text bot check.

Key results

• Study 1 (Knobe 2003 marksman; 2 skill × 2 morality, within-subjects; N = 198 of 200).
  • Intentionality: high skill M = 6.65 (SD 0.98) vs low M = 5.19 (2.14), F(1, 591) = 240; bad 6.51 vs neutral 5.34, F = 154; skill × morality interaction F = 94 (skill effect larger when morally neutral); all p < .001.
  • Causation: 6.56 vs 5.35, F = 194; bad 6.39 vs neutral 5.52, F = 100; interaction F = 79.
  • Know-how: skill F = 643, morality F = 46, interaction F = 54. Probability-raising: 127 / 66 / 18.
  • Mediation by causation: 64% of the skill effect [0.56, 0.73].
  • Mixed model standardized β: causation .46, know-how .33, probability-raising .14. Bootstrap CIs on differences: causation − know-how [.01, .25]; causation − probability-raising [.20, .43]; know-how − probability-raising [.10, .28].
• Study 2 (Kraemer; 2 Means/Ends × 2 morality, between-subjects; N = 673 of 695).
  • Intentionality: ends M = 4.26 (2.40) vs means 2.40 (1.87), F(1, 669) = 176; bad 4.22 vs neutral 2.44, F = 164; interaction F = 105.5.
  • Causation: ends 3.71 vs means 2.43, F = 78; morality F = 145; interaction F = 86.
  • Know-how: condition F = 178, morality F = 0.27 (p = .60), interaction F = 19. Probability-raising: 44 / 38 / 7.36 (p = .007).
  • Within the neutral vignette: intentionality ends 2.70 vs means 2.20, t(333.2) = 2.62, p = .009; know-how t = 6.26; probability-raising t = 2.75; causation 2.22 vs 2.20, t(350.3) = .12, p = .91 (no effect).
  • Within the bad vignette: intentionality 6.01 vs 2.62, t = 16.9; causation 5.39 vs 2.68, t(315) = 11.6; know-how t = 13.3; probability-raising t = 7.46.
  • Mediation: 45% [0.36, 0.55].
  • β: causation .56, know-how .20, probability-raising .16. Difference CIs: causation − know-how [.24, .47]; causation − probability-raising [.30, .49]; know-how − probability-raising [−.05, .14] (not reliable).
• Study 3 (gumball machine; Knowledge × Process, within-subjects; N = 142 of 198).
  • Intentionality: know 4.88 (2.22) vs don’t 2.91 (1.98), F(1, 423) = 202; normal 4.65 vs mix-up 3.14, F = 118; Knowledge × Process F = 32, p < .001 (Process matters more with knowledge).
  • Causation: 4.50 vs 2.96, F = 165; 4.25 vs 3.21, F = 75; interaction F = 36.
  • Know-how: Knowledge F = 800; Process p = .19; interaction p = .12. Probability-raising: Knowledge F = 8.5 (p = .004); Process F = 91; interaction p = .79.
  • Mediation of the Process effect: 49% [0.38, 0.61].
  • β: causation .53 [.46, .60], know-how .27 [.20, .33], probability-raising .10 [.04, .16].
  • Conditional Nakagawa R²: causation-only model .57 vs know-how × probability-raising model .44; difference CI [.07, .19].
• Item level (12 condition means across studies): intentionality correlates with causation r(10) = .98, with know-how r(10) = .82, and with probability-raising r(10) = .88.

Claims

id	claim	strength	support
C1	Luck manipulations (skill, means/ends, process) move causal judgments of “E because the agent wanted E” in the same direction as intentionality judgments	strong (for the association)	experiments S1–S3, ANOVAs above
C2	Causal judgment predicts intentionality better than know-how or probability-raising	moderate	mixed-model β in all three studies with bootstrap CIs on differences; S1’s causation − know-how CI [.01, .25] barely excludes 0
C3	The Knowledge × Process interaction in intentionality, a novel prediction of the causal account, holds and is mirrored in causation but not in know-how or probability-raising	moderate–strong	S3, pre-registered; single vignette family
C4	Luck affects intentionality via causal judgment	weak–moderate	exploratory, non-pre-registered mediation covering only 45–64% of effects; the authors say the evidence is “mostly correlational”
C5	The causal account explains the Kraemer effect	weak (partial)	S2: holds in the morally bad vignette only; in the neutral vignette, intentionality shows the effect and causation does not
C6	The account also explains the objective-vs-subjective probability, alternative-option and moral-override patterns in prior literature	informal argument	Intro, re-reading of Ericson et al. 2023, Mele & Cushman 2007, Knobe 2003; not tested here
C7	The folk concept of intentional action is externalist, like that of knowledge	informal argument	Conclusion; analogy only
C8	Control-based accounts cannot explain S3, since the manipulation occurs after the agent’s action	informal argument	General Discussion

Method

Vignette experiments on Prolific with Likert (1–7) agreement ratings of four statements per scenario: intentionality, causation (“because [she] wanted”), know-how and probability-raising. Question order was partly fixed (S1, S3: intentionality first; S2: intentionality last). Analyses: (repeated-measures) ANOVA; Welch t-tests within vignettes; linear mixed-effects models with participant random intercepts and standardized β; participant-level bootstrap (1,000); exploratory mediation; item-level Pearson correlations over condition means. Data and code are on OSF.

Concepts

• Causal account of intentional action: “A did X intentionally” ≈ “X happened because A wanted X” (after Quillien & German 2021; Davidson 1980).
• Necessity: whether E would still have happened without C.
• Robustness / sufficiency: whether C would still have produced E under slightly different background conditions. The paper’s key lever for luck.
• Luck effect: lower intentionality for outcomes achieved by chance.
• Kraemer effect: the ends judged intentional, but the chancy means to them not.
• Know-how thesis (Pavese; Pavese & Henne 2023) and probability-raising thesis (Ericson et al. 2023; Doan et al. 2025), read non-teleologically: the rival accounts.
• Process (S3): whether the machine works normally or mixes up containers after the choice. It varies probability-raising while holding the agent’s mental states fixed.

Connections

It builds on Quillien & German (2021) and on counterfactual models of causal selection (Icard, Kominsky & Knobe 2017; Quillien & Lucas 2024). It replicates Knobe (2003) and Pavese & Henne (2023), and re-reads Ericson et al. (2023) and Doan et al. (2025). Among held works, LIT-206 (Goldstein et al., interpretationist desires for LLM “instance agents”, Deferred) is the closest neighbour on the question of attributing desires and agency. Quillien measures how people turn a desire attribution into an intentional-action attribution, and the answer is that the desire–outcome link must be robust. The record’s counterfactual-causation papers (LIT-144, Pernu on exclusion and downward counterfactuals) share the necessity/sufficiency vocabulary but are metaphysics, not folk psychology.

Agency. The work holds a causal-externalist account of attributed agency. What makes an outcome someone’s intentional doing, in the folk concept, is that their desire was a robust (not merely necessary) cause of it. It is a theory of the concept of intentional action, not of what agents are. The agency tag is justified: intentional action and control are in the tag’s blurb, and a reader browsing agency would expect this. It should stay first, with cognition second.

Bearing on the record

No THEORY in the record rests on it that I know of. It could support a theory note on the folk attribution of intentional action, if one is written. For ML practice it carries nothing: there is no instruction for training or evaluating models. One tangential relevance, which is not a practice: the finding that people withhold intentionality when the desire–outcome link is non-robust bears on how people will attribute intent to automated agents whose success is chancy. The paper does not study machines.

Limitations

• The evidence is correlational. Mediation is exploratory and partial (≈ half).
• The author raises two worries himself: asking causation and intentionality about the same vignette invites task-demand and consistency effects, and the argument against this (differential tracking) is indirect.
• The causal-judgment item (“because [she] wanted”) is close in wording to the intentionality item, and is not a measure of causal strength per se.
• Few vignettes: one scenario family per study. The morality contrast is confounded with difficulty and wording.
• The item-level r = .98 rests on 12 condition means.
• The neutral Kraemer case dissociates intentionality from causation (also Varghese & Henne 2025), which the account does not explain.
• No computational counterfactual model is fitted, so “robustness” is never measured, only inferred from the manipulations.
• The SI was not read (see read:).

Open questions

• Fit an explicit counterfactual causal-strength model (e.g. Quillien & Lucas 2024) to the vignettes and test whether its predicted desire–outcome strengths predict intentionality quantitatively, beyond the rated causation item.
• What drives the residual half of the luck effect, and the neutral Kraemer divergence?
• Do the same patterns hold for attributions to artificial agents, and across cultures beyond US Prolific?

Corrections to the seeded skim

• There is no formal or computational model. The dossier and skim describe a “causal theory” as if it had a model, and the request asked for “the model’s exact form”. The account is verbal: “people judge that an agent did X intentionally if the agent’s desire toward X caused X to happen” (Our Proposal), where causal strength is graded. The two ingredients, necessity (“would E still have happened if C had not happened?”) and robustness/sufficiency (“would C still have led to E even if background circumstances had been slightly different?”), are taken informally from counterfactual theories (Icard et al. 2017; Quillien & Lucas 2024). Note 2 says the mind need not compute them explicitly. No equation is fitted and no model predictions are computed. The only quantitative models in the paper are statistical: repeated-measures and 2×2 ANOVAs; linear mixed-effects regressions of intentionality on causation, know-how and probability-raising with participant random intercepts (standardized β via effectsize::standardize_parameters, “refit”); participant-level bootstrap (1,000 resamples) CIs on β differences; and exploratory mediation of the form Condition → Causation → Intentionality.
• “Subsumed by” overstates the paper. The skim says know-how and probability-raising are “presented as consistent with, and subsumed by” the causal account. The paper says they are “consistent with” it (Other Recent Accounts). The General Discussion concedes that know-how and probability-raising stay “weakly or moderately associated with intentionality even when controlling for causal judgments”, and that the findings “do not foreclose” independent effects.
• Causal judgment explains only about half of the luck effect. The dossier and abstract do not say this. The mediated shares are 64% [.56, .73] in S1, 45% [.36, .55] in S2 and 49% [.38, .61] in S3. All the mediation analyses are exploratory and were not pre-registered (note 11).
• Study 2 was not pre-registered, although it follows the pre-registrations of pilots S2a and S2b. Studies 1 and 3 were pre-registered.
• The Study 2 vignette is not the intro’s. The intro’s Kraemer example (a lever opens one of ten boxes, box eight holds poison) is not the Study 2 stimulus. Study 2 used Pavese & Henne’s arrow-down-path-eight “moral” vignette and a game-show green-ball vignette.
• Study 3 lost 28% of its sample to comprehension checks (56 of 198 excluded; final N = 142). The dossier does not mention this.
• Typo in the published text. The Study 3 causation mix-up mean is printed as “(SD = 3.21, SD = 1.98)”. It is presumably M = 3.21.
• Otherwise the dossier’s bibliographic data (Open Mind vol. 10, 857–883; 2026-06-17; DOI 10.1162/OPMI.a.361) and its reading of the Study 2 divergence (the neutral Kraemer case) agree with the text.