NOTE-034: Curry et al. 2019, morality-as-cooperation in 60 societies

Contribution

• A game-theoretic taxonomy of seven types of cooperation, each paired with a moral domain:
  • kin allocation → family;
  • coordination → group loyalty;
  • social exchange → reciprocity;
  • hawkish contest displays → bravery;
  • dovish displays → respect;
  • division → fairness;
  • possession → property. From this the paper derives the prediction that each cooperative behaviour will be judged morally good in every culture where it arises (pp. 48–50).
• A holocultural content analysis of eHRAF for the 60-society HRAF Probability Sample, with a published codebook (Table 1) and search syntax (Table S2), reporting valence and cross-cultural presence of the seven.

Key insight

If morals are solutions to recurrent cooperation problems, then each solution should be called moral wherever the problem arises. So the empirical signature is not merely that cooperation exists but that it is always valenced positively. The paper finds no society in which any of the seven is valenced negatively.

Assumptions

• Sample. The eHRAF Probability Sample Files (60 societies, six regions, at least 1,200 pages each by trained ethnographers) are representative enough, and independent enough (Galton’s problem), to host counterexamples. The authors concede that it “does not completely solve” non-independence (p. 58).
• Ethnographic proxy. Moral valence, as described by (mostly Western, mid-20th-century) ethnographers, stands in for the society’s moral judgments.
• Coding scheme. Presence is coded as binary. Absence of evidence is read as underreporting, not absence (p. 55).
• Operationalizations (Table 1):
  • hawkish traits are coded only as bravery;
  • dovish traits as respect for elders, chiefs and superiors;
  • division only when division resolves or forestalls a dispute, with generic “sharing” excluded.
• Fn. 3. A cooperative behaviour is predicted to be judged bad only when pursued at the expense of a larger form of cooperation. Such cases, and the 18 value conflicts found (fn. 11), were not coded as negative.

Key results

• Table 2 (paragraph observations, positive/negative): kin 214/0, group 127/0, reciprocity 151/0, bravery 101/0, respect 133/0, fairness 17/0, property 218/1. Total 961/1.
• Table 3 (societies with a positive observation, of 60): kin 47, group 44, reciprocity 43, bravery 32, respect 35, fairness 9, property 54.
• Regional variation. One-way ANOVAs by region were non-significant (F from 0.5 to 1.28; P from .29 to .78). With about 10 societies per region, the authors acknowledge low power (fn. 14).
• Coverage. Morals observed per society correlate with the society’s total eHRAF paragraph count (r = 0.43, P = .001).
• Conflicts. 18 examples of two of the seven in conflict (fn. 11), 9 of them family vs group. They were portrayed as conflicts, not as one being bad.

Claims

id	claim	strength	support
C1	In 60 societies, the seven cooperative behaviours are morally valenced positively in 961/962 observations, with no society valencing any of them negatively	moderate (as a description of the coded record)	content analysis, Tables 2–3
C2	These seven are “plausible candidates for universal moral rules”	weak–moderate	follows from C1 only if negative valence could have been detected. The behaviours are defined in virtue terms (“helping”, “respecting”, “being fair”) and conflicts are set aside (fn. 3, fn. 11), so the test is lopsided
C3	The results “provide strong support” for morality-as-cooperation, and “Hume was right, and Locke was wrong” (p. 55)	weak	C1 cannot discriminate MAC from any theory predicting that these behaviours are commonly praised; no competing domain (purity, sexual norms, harm) was coded for comparison, which Bloom and Smith & Kurzban press on
C4	The seven morals occur “with equal frequency across all regions”	weak	non-significant ANOVAs at n ≈ 10 per region: absence of evidence, stated as equality
C5	Most of these morals are present in most societies	moderate	6 of 7 in more than 50% of societies; fairness in 15% (Table 3), attributed by the authors to narrow coding
C6	Much variation in observed morals reflects ethnographic coverage, not culture	moderate	r = 0.43 with paragraph count (p. 55)
C7	MAC is more precise and broader than Moral Foundations Theory, since MFT lacks kin, reciprocity, bravery and property and includes the non-cooperative “purity”	informal argument	theoretical comparison (pp. 49–50); not tested here
C8	Sexual morality, food taboos and moral externalization are also cooperative (Reply)	weak	post-hoc hypotheses offered in response to commentators, explicitly untested

Concepts

• morality-as-cooperation (MAC) — morality is “a collection of biological and cultural solutions to the problems of cooperation recurrent in human social life” (abstract; Curry 2016).
• moral valence — a paragraph describes the behaviour as good, right, moral, virtuous, an obligation or duty (positive), or as bad, wrong, shameful, taboo (negative) (fn. 8).
• Probability Sample Files — the HRAF 60-society stratified random sample, designed to reduce Galton’s problem.
• hawkish / dovish traits — displays of formidability (bravery, fortitude) and of submission (deference, respect) that settle contests without fighting (Maynard Smith & Price).

Connections

• The theory’s lineage. It extends kin selection (Hamilton), reciprocal altruism (Trivers), coordination (Lewis), hawk–dove contests (Maynard Smith & Price), bargaining (Nash, Skyrms) and possession (Gintis 2007). It is positioned against Moral Foundations Theory (Haidt & Graham), which it criticizes as ad hoc. The questionnaire follow-up is Curry, Jones Chesters & Van Lissa 2019 (J. Res. Pers.), cited in the Reply as testing the seven-factor structure.
• The programme and its rivals, now read. Curry 2016 (LIT-503) states the prediction this paper did not test: behaviour unrelated to cooperation will be “morally neutral”. The questionnaire (LIT-496), “Moral Molecules” (LIT-509) and the 256-society study (LIT-512) do not run the blind, all-content test either, and the last cannot see valence. The purity objection has three cooperative replies on record, none worked out inside the seven domains: Fitouchi, André & Baumard’s self-control account (LIT-101), Rai & Fiske’s Unity motive (LIT-507), and the sexual-restraint-as-fairness sketch in Baumard, André & Sperber’s Response (LIT-508). Moral Foundations Theory keeps Sanctity as a foundation (LIT-499, LIT-500).
• Bloom.
  • It is unclear what would falsify MAC.
  • It is silent on proximate mechanisms.
  • It does not explain why cooperation is moralized rather than merely liked.
  • A test would also need to show that non-cooperative behaviour is not moralized, and sex morality and battery look non-cooperative.
• Gintis. Supportive. He places MAC within his account of hominin politics: lethal weapons, reverse dominance hierarchy, socially constructed rules.
• Smith & Kurzban.
  • Morality is nonconsequentialist, as trolley judgments show.
  • Many norms are unrelated to cooperation or destructive (cow-slaughter taboo, kidney-sale bans).
  • People endorse the fairness rule that benefits themselves.
• Wong.
  • Morality also serves intrapersonal coherence and ideals of a worthwhile life, including autonomy and aesthetic ritual.
  • The domain-frequency variation may not be explicable.
  • He argues for pluralism rather than universalism.
• Reply. The authors:
  • recast MAC as a functional theory that yields proximate predictions;
  • propose cooperative accounts of sexual morality across all seven domains and of the cow taboo;
  • argue that self-serving, nonconsequentialist norms follow from morals being conditional strategies.
• In this record. A companion to LIT-035 (Peters & Adamou), a different, variance-reduction account of why cooperation pays, and to the collective-behaviour works LIT-075 and LIT-046. It gives the cooperation thread its normative side.

Bearing on the record

There is no instruction for ML practice. For this record’s ethics and cooperation thread, it is the standard empirical citation for “cooperative behaviours are moralized positively across cultures”. Any THEORY document citing it for universality should carry the caveat that the design could hardly find the opposite, and that fairness was seen in only 15% of societies. No anthology document is known to cite it.

Note of 2026-10-16: THEORY-468 and THEORY-653 now cite LIT-047 as a source.

Limitations

• Asymmetric detectability. Negative valence of a behaviour defined as “helping”, “respecting” or “being fair” is near-contradictory. Negatives that do occur can be reassigned to “conflict” (fn. 3, fn. 11) or explained as one cooperation trumping another (the Chuuk case). Bloom’s falsifiability question is not answered in the paper or the Reply.
• Non-blind screening. The step from 3,460 to 1,426 paragraphs was done by the authors, with no reliability reported. Blind agreement on the remainder was moderate (κ = .58), and poor for fairness (κ = .14). The final codes are consensus.
• No comparison domain. The paper does not measure how often non-cooperative behaviours (purity, sexual conduct, treatment of the dead, harm) are moralized. So it cannot show that morality is specifically about cooperation, which is the paper’s stated test.
• Unit of analysis. Paragraphs are clustered within societies and sources (603 sources over 300 years). Only presence/absence per society is analysed, so strength and ranking of values are not measured, as the authors acknowledge (p. 58).
• Power. Regional equality rests on non-significant ANOVAs at about 10 societies per region.
• Scope. Seven of MAC’s cooperative behaviours were tested. Others (forgiveness, generosity) and the many non-cooperative moral topics the authors met in the corpus (industry, honesty, chastity, hospitality, gossip, religious obligations) were not coded (p. 58).

Open questions

• In the same 60 societies, how often are non-cooperative domains (purity, sexual conduct, harm to strangers) moralized, coded blind with the same procedure? That is the discriminating test between MAC and pluralist accounts.
• Does the fairness rarity (15%) survive a broader coding that includes sharing norms, or does it mark a real gap?
• Does domain prioritization vary with ecological indicators of cooperation value (kin density, mobility, warfare), as MAC’s further prediction states (pp. 58–59)? This needs graded, not binary, codes.

Corrections to the seeded skim

• Publication date. The version of record reads “submitted 13 V 16, accepted 22 XII 17, and electronically published 8 II 19” (p. 47), in Current Anthropology 60(1), February 2019. That settles the dossier’s date conflict for 2019-02-08, matching ORA. Crossref’s 2019-02-02 was not traced and is presumably an issue or metadata date. The May 2016 date on the supplement matches the submission, which explains the dossier’s “2016 version” hint: a submitted manuscript, not a separate publication.
• Corpus. The dossier describes the corpus as “2,519 Ethics paragraphs from 400 documents”. That is only the first Ethics extraction. The full pipeline:
  • Ethics 2,519 paragraphs, plus Norms 1,605 (58 societies), plus a keyword/subject search 1,737;
  • after de-duplication, 3,460 paragraphs, 606,556 words, from 603 sources dated 1704–2009 (median 1967);
  • authors’ screening leaves 1,426 germane paragraphs;
  • 761 paragraphs carry valence information, giving 962 observations (Methods, fn. 7).
• Coding independence, which the dossier asked to check.
  • Raters 1 and 2 are authors Curry and Mullins, who were not blind. They coded all 3,460 paragraphs and conferred, which cut the set to 1,426.
  • A hypothesis-blind rater 3 coded only those 1,426.
  • Agreement between the two sets was κ = .58 overall, by domain: kin .52, group .47, reciprocity .66, bravery .65, respect .56, fairness .14, property .75.
  • Discrepancies were then “reviewed and resolved until complete agreement was reached (κ = 1)”.
  • So the κ = 1 is a consensus, not a reliability figure. No reliability is reported for the non-blind 3,460 → 1,426 screening step, which decided what could be counted at all.
• Coverage. The dossier’s “most of these morals were observed in most societies” understates the variation.
  • Fairness appears in only 9/60 societies (15%); bravery in 53% and respect in 58% (Table 3).
  • Property is significantly more frequent than bravery, respect and fairness, and fairness is significantly less frequent than every other domain (repeated-measures GLM, Bonferroni, P ≤ .001).
  • Mean 4.4 of 7 per society (SD 1.5). The reported “mode = 5.5” is impossible for integer counts; a bimodal 5/6 is presumably meant.
• Commentaries. The dossier read only “the start” (Bloom). All four and the Reply are summarized under Connections. The sharpest objections come from Bloom (what would falsify it; sex and harm morality) and from Smith & Kurzban (non-cooperative and destructive norms).
• Regions. The text says all seven were observed in all six regions except “dividing disputed resources in Central America”. Central America is not one of the six Probability Sample regions (Sub-Saharan Africa, Circum-Mediterranean, East Eurasia, Insular Pacific, North America, South America; Methods). This is an internal inconsistency, and I could not check it against Table 3’s layout.
• Hawk–dove is not in Maynard Smith & Price 1973. The paper, as the lineage above reports it, credits “hawk–dove” contests to Maynard Smith & Price, and Curry 2016 (LIT-503) does the same. The 1973 paper (LIT-505, read in NOTE-404) has five strategies, Mouse, Hawk, Bully, Retaliator and Prober-Retaliator; “dove” does not occur, and its contestants are identical, so it has no display-then-defer asymmetry either. What it grounds is restraint backed by retaliation. The display and deference domains need a later source; which one is open here.
• C7 is half right. Read against the Moral Foundations texts (LIT-502, LIT-499, read in NOTE-396), “more precise” holds as a claim about how the lists were made: MAC derives its domains from games, and MFT’s five criteria screen candidates but do not generate them. “Broader” does not. MFT grounds Care in kin selection and Fairness in reciprocal altruism, and names ownership as a candidate foundation; only bravery is absent. In the other direction MAC leaves out Sanctity and care for non-kin, so neither list contains the other. The narrower charge stands: MFT folds reciprocity into fairness, and no MFQ item, in either version, asks about returning a favour.