A large-scale investigation of everyday moral dilemmas

Abstract

Questions of right and wrong are central to daily life, yet scientific understanding of everyday moral dilemmas is limited. We conducted a data-driven analysis of these phenomena by combining state-of-the-art tools in machine learning with survey-based methods. We extracted and analyzed 369,161 descriptions ("posts") and 11 M evaluations ("comments") of dilemmas from the largest known online repository of everyday moral dilemmas: Reddit's "Am I the Asshole?" Users described a wide variety of everyday dilemmas on topics ranging from broken promises to privately held emotions. Dilemmas involving relational obligations were the most frequently reported, while those pertaining to honesty were the most frequently condemned. The types of dilemmas people experienced depended on the interpersonal closeness of the interactants, with some dilemmas (e.g. politeness) more prominent in distant-other interactions and others (e.g. relational transgressions) more prominent in close-other interactions. A preregistered follow-up investigation showed that similar dilemmas are reported in a census-stratified representative sample of the US population (n = 510). Overall, this paper highlights the diversity of moral dilemmas experienced in daily life and contributes to the development of a moral psychology grounded in the vagaries of everyday experience.

Keywords: closeness; dilemmas; harm; honesty; judgments; morality; relationships.

Fig. 1.

Overview of the AITA dataset (n = 369,161). A) Frequency of posts and comments across time. In this figure, posts whose negative evaluation score (weighted proportion of YTA + ESH comments) was less than one-third are labeled NTA; those whose negative evaluation score was greater than two-thirds are labeled as YTA; and those that did not reach either threshold are labeled as controversial. B) Frequency distribution of comments per post. C) Frequency distribution of word counts by post. D) A semantic map of the situations described across all posts. To create this visualization, we extracted an embedding (a 768-dimensional vector) of each post using the RoBERTa sentence encoder (32). We then reduced the embeddings to two dimensions and projected them onto a coordinate plane using UMAP (33). Then, we used the HDBSCAN (34) algorithm to identify clusters of posts and labeled each cluster according to its most prominent words (as reflected in an analysis of term frequency, inverse document frequency), “or TF-IDF.” The most frequently occurring situations pertained to family (21%), the office (8.4%), and weddings (4.2%) (35).

Fig. 2.

Overview of the 29 most common dilemma types present in the AITA dataset, grouped according to moral theme (n = 369,161 posts). Word clouds reflect the bigrams with the highest relative frequency for each moral theme. The name of each dilemma type is displayed along with the title of that dilemma type's most “exemplary” post (i.e. the post for which that dilemma type was most prevalent). The fact that the post titles aptly characterize their corresponding dilemma type supports the claim that our algorithmic approach was successful in detecting subtle differences in moral language.

Fig. 3.

Average negative evaluation of each of the 29 dilemma types and six moral themes identified in the AITA catalog of everyday dilemmas (n = 369,161). In the main panel, dilemma types are ordered according to evaluation; dot size corresponds to relative frequency of occurrence of each dilemma type across all posts with range (1.2%, 44.2%). As shown in the figure, several forms of dishonesty, including cheating, misrepresentation, and secret violation, are among the most negatively evaluated dilemma types; the most common dilemma types are various forms of relational obligation as well as behavioral overreaction. In the right panel, moral themes are similarly plotted according to frequency and evaluation. The inlaid panel on the top left indicates the frequency of occurrence of the different moral themes.

Fig. 4.

Prevalence of each dilemma type in each relational context. For this analysis, we used only the posts (n = 298,632, or 80.1% of data) that contained a common and identifiable relationship (see method 6 for additional details). Prevalence values reflect the mean predicted likelihood of each dilemma type appearing in posts containing each relational context (theoretical range: [0, 1], M = 0.051, SD = 0.020). Bar graphs provide example comparisons of prevalence values between the mother and boyfriend relational context; error bars = 95% CI.

Fig. 5.

Associations between the prevalence of each dilemma type and the estimated interpersonal closeness of the interactants. Intervals = 95% CIs around standardized beta coefficients. Estimates colored red are significantly below zero, blue significantly above zero, and gray not significantly different from zero.

Fig. 6.

Correlations between features of dilemma types extracted from the AITA database and those observed in the representative sample (study 2). The top left panel shows the relationship between the frequency of the various dilemma types in the AITA sample (n = 369,161) and that in the representative sample (n = 510), r(27) = 0.73. The top right panel shows the relationship between the average negative evaluation (that is, proportion rated “YTA”) of dilemmas in the AITA sample and the average wrongness judgments of dilemma in the representative sample, r(27) = 0.68. The bottom panel shows the correlation between the beta weight reflecting the association between dilemma prevalence and relational closeness in the AITA sample, and the average closeness of the relational contexts in which the various dilemma types occurred in the representative sample, r(27) = 0.75.