¿Ves algo que mejorar? Propón un cambio.
Esta página es pública.
Whether a comment is offensive depends on who reads it, and existing AI tools for detecting offensive text rely on labels that ignore that. A study tested it with 4,309 participants from 21 countries in eight cultural regions, who rated social media comments in English. People differed widely by region, and by individual moral values.
Offensiveness labels by country; annotators from China, Brazil and Egypt gave significantly different labels (Davani et al., CC BY 4.0).
How it worked
Participants rated 40 items each on a five-point scale from not offensive at all to extremely offensive, five of them quality-control checks; after those checks, 4,295 participants remained. The 4,554 items came from two public Jigsaw datasets of comments labeled for toxicity, a random, more ambiguous set, a set mentioning groups defined by gender, sexual orientation or religion, and a set with moral sentiments. Half the participants received a definition of "extremely offensive language"; the other half used their own sense of it. Each person also filled out the Moral Foundations Questionnaire, which scores six values: Care, Equality, Proportionality, Authority, Loyalty and Purity. Each item had labels from at least three people in every region.
Study 1: regions disagree
Offensiveness ratings differed significantly by region. Arab culture (average 1.19) and Latin America (1.13) rated most offensive, and the Sinosphere and Oceania (0.80 each) least. Of 28 pairs of regions, 25 differed significantly. The three pairs that did not were Sub-Saharan Africa and Indian culture, Western Europe and North America, and the Sinosphere and Oceania. The gap held after accounting for age, gender and socioeconomic status, and giving a definition narrowed it but did not remove it.
| Finding | Detail |
|---|---|
| Country level | Egypt, Brazil, India and Argentina all averaged above 1.2; China averaged below 0.6, though a rater pool skewed toward ages 30 to 50 may partly explain that |
| Cultural similarity to the US | The more similar a country to the United States, the fewer comments it called offensive, though the effect was not statistically significant |
| Cultural dimensions | Countries higher in uncertainty avoidance, individualism, femininity and short-term orientation rated more comments offensive; power distance and indulgence made no significant difference |
Study 2: moral values matter
Region and country explain little of the variation: individual differences account for 39.5 percent, country for 2.9 percent and region for 0.7 percent. So the team tested whether moral values explain the regional gaps. They found that Care and Purity significantly mediate them: regions whose people score higher on those values tend to rate more as offensive. Looking at individuals, someone who scores higher on Care than their country's average rates comments as more offensive, and the same holds for Purity, with a smaller effect. Care varies less between regions, since most people score high on it, while Purity varies more.
Care scores vary across regions and mediate regional differences in offensiveness ratings (Davani et al., CC BY 4.0).
Study 3: what it means for AI
The authors compared their ratings with the Jigsaw labels and the Perspective API, a commercial toxicity detector. Both match the regions' majority votes only weakly, best in Latin America, the Arab world and Sub-Saharan Africa and worst in the Sinosphere and Oceania. Perspective API also agrees more with annotators who score high on Care. The authors say that fits crowd-worker recruitment that leans on the Global South, but they flag the low match with North American raters as a concern, since the tool is probably used mostly by North American websites. Both measure toxicity rather than offensiveness, so some mismatch is expected.
Correlation between Perspective API labels and the majority vote of raters in each region (Davani et al., CC BY 4.0).
Why it matters
The authors conclude that judging offense is a moral judgment shaped by culture, and that datasets and models reflect some value systems over others. They call for culturally informed data collection, more diverse rater pools and community involvement in aligning AI with local norms, while warning that cultural values can conflict with the equity of marginalized groups within a society. The key question, they write, is whose values a model should follow.
Sources
- Aida Davani, Mark Díaz, Dylan Baker et al., Disentangling Perceptions of Offensiveness: Cultural and Moral Correlates, arXiv:2312.06861 (2023), published under CC BY 4.0. Changed here: condensed to a summary, with three of the authors' figures reproduced, and rewritten in hubnx's own words.
Comentarios
