Biased Tales: Narrative Bias Analysis

• Biased Tales is a research area that examines how stories encode and reproduce gender, cultural, and socio-demographic stereotypes in both human-authored and machine-generated narratives.
• Studies reveal that narrative biases are measurable across trope distributions, moral lexicons, and event structures using both qualitative and computational techniques.
• Research highlights the need for culturally grounded, context-specific evaluations to mitigate bias in educational texts, media, and AI-generated stories.

Searching arXiv for the cited papers to ground the article in the current record. Biased Tales denotes a research area concerned with how stories, tropes, and open-ended narrative generations encode and reproduce gender, cultural, and socio-demographic stereotypes, and it also names a dataset for analyzing such patterns in LLM-generated children’s stories (Rooein et al., 9 Sep 2025). Across human-authored corpora, fairy tales, and machine-generated narratives, the literature treats stories as socially consequential artifacts: popular media reflects and reinforces societal biases through tropes; fairy tales are a common resource for young children to learn a language or understand how a society works; and long-form LLM outputs such as bedtime stories, personas, and educational texts can carry both explicit and implicit bias (Gala et al., 2020, Zhou et al., 2022, Lum et al., 2024).

1. Scope and intellectual lineage

Early work on biased narratives was predominantly qualitative. A study of 147 tales written by Argentinian secondary-school students found that 78 stories featured scientists as central characters and that the dominant image was individualistic, solitary, and male, often framed through secrecy, extraordinary breakthroughs, and ethical ambiguity. The same analysis organized these depictions through social, practical-technological, knowledge, spatial-temporal, ethical, and emotive-mythic dimensions, with 72 narratives addressing the social dimension and 60 the ethical dimension (Pujalte et al., 2013).

This line of inquiry later expanded into computational analyses of narrative corpora. The shift is visible in work that models trope repositories, moral lexicons, event chains, commonsense inferences, and large-scale open-ended generations. A plausible implication is that “biased tales” are no longer treated only as isolated literary examples, but as measurable distributions over character attributes, events, settings, and inferred social meanings (Isaza et al., 2023, Huang et al., 2021).

2. Human-authored narratives: tropes, morals, and event structure

A central resource for large-scale analysis is the TVTropes corpus. One study crawled TVTropes.org and created a dataset containing approximately 30,000 tropes and 1.9 million occurrences across 40,000 works of film, television, and literature, enriched with IMDb and Goodreads metadata. For trope $i$, the paper defines a raw genderedness score

$$d_i = \left(\frac{f(X_i)}{f(X_i)+m(X_i)}\right)\Bigg/\left(\frac{f(\mathrm{tvtropes})}{f(\mathrm{tvtropes})+m(\mathrm{tvtropes})}\right),$$

and then normalizes it to

$g_i = \text{z-score of } d_i.$

Here, $X_i$ is the concatenation of the trope description and all its examples; the corpus-level female ratio is 0.32; and tropes with $|g_i|>1$ are treated as highly gendered. Manual validation on 150 random examples yielded male precision 0.85, recall 0.86, F1 0.86, and female precision 0.72, recall 0.78, F1 0.75 (Gala et al., 2020).

The substantive findings show asymmetry not only in valence but in topical breadth. Male-dominated tropes spanned science and technology, war and strength, religion, money and business, and intelligence or genius, whereas female-dominated tropes centered on relationships and marriage, physical appearance and beauty, sexuality and fanservice, motherhood and maternal roles, and food or eating. LDA with 75 topics yielded 45 male-skewed and 30 female-skewed topics, suggesting greater topical breadth among male tropes. The same study reports that logistic regression predicted high, medium, or low IMDb rating from trope features with 55% accuracy versus a 36% baseline, and predicted Goodreads author gender from binary trope occurrence vectors with 71% test accuracy versus a 64% baseline (Gala et al., 2020).

Fairy-tale research extends this analysis from trope inventories to moral language and narrative process. A corpus study of 624 fairy tales from 7 cultures identified 4,405 male and 2,125 female characters, giving a male-to-female ratio of approximately 2.07, or 107.29% more males than females. Using the extended Moral Foundations Dictionary, the study found female characters more associated with care, loyalty, and sanctity, and male characters more associated with fairness and authority. Event extraction further showed female associations with emotion, appearance, household activity, and life-stage events such as marriage or birth, while male characters were more associated with profession, violence, travel, justice, and conflict (Zhou et al., 2022).

A related event-chain study moved from frequency counts to temporal narrative structure. Using a pipeline that combined BookNLP, AllenNLP semantic role labeling, ECONET, and a custom verb annotation scheme, the authors analyzed 33,577 events, of which 69% were male and 31% female. They reported that 14% of unigram events were biased toward males and 11% toward females, while 20% of bigrams showed significant gender bias. Female characters exhibited the strongest bias in the beginning of their event chains, whereas male bias was consistent throughout the narrative. This temporal result matters because it shows that bias is not confined to who acts, but extends to when specific action types are allocated within a character’s story arc (Isaza et al., 2023).

3. Implicit bias in model-generated narratives

Bias in generated stories is often implicit rather than lexical. One influential pipeline fine-tuned GPT-2 on ROCStories, generated 9,796 short stories, identified the protagonist via SpanBERT-based coreference resolution, anonymized character names, and then applied COMeT to infer protagonist attributes, motivations, mental states, and effects on others. The semantic association score for a COMeT-generated word $x$ and lexicon $L$ is

$S(x,L)=\frac{1}{|L|}\sum_{a\in L}\cos(e(x),e(a)),$

with an analogous construction for a semantic power axis. A bag-of-words logistic regression on anonymized sentences predicted protagonist gender at only 57% accuracy, whereas a BERT classifier trained on concatenated COMeT inferences reached 68.15% accuracy. The study found that male protagonists had higher intellect scores, female protagonists higher appearance scores, and that some of these stereotypes were amplified in GPT-2 stories relative to human-written stories (Huang et al., 2021).

This inference-based perspective complements work on persona-conditioned LLM behavior. A study spanning 24 reasoning datasets, 4 LLMs, and 19 personas found that models overtly rejected explicit stereotypes when asked directly, but manifested stereotype-consistent and erroneous presumptions when answering while adopting a persona. For ChatGPT-3.5, 80% of personas showed statistically significant drops on at least one dataset, some datasets exhibited performance drops of 70% or more, and GPT-4-Turbo still showed bias in 42% of personas. The paper distinguishes explicit abstentions—such as persona-conditioned refusals that falsely link a social identity to lack of competence—from more difficult-to-detect implicit errors that persist even after abstentions are removed (Gupta et al., 2023).

Taken together, these studies indicate that scrubbing explicit gender terms or eliciting fairness disclaimers does not eliminate biased narrative reasoning. A plausible implication is that narrative bias should be analyzed at the level of inferred motivations, social consequences, and persona-conditioned competence, not only at the level of token counts.

4. Dedicated “Biased Tales” datasets for children’s story generation

The 2025 “Biased Tales” dataset was designed specifically to analyze cultural and topic bias in LLM-generated children’s bedtime stories. It contains 5,531 short stories generated by GPT-4o, Llama3-8B, and Mixtral8x from prompts that vary child gender, parental role, religion, ethnicity, and nationality. Annotators extracted 2,536 unique protagonist attributes grouped into Physical, Emotional, Mental, Moral, and Other categories; context-centric attributes captured cultural, environmental, and socioeconomic details; and bias was measured through Pearson correlations and TF-IDF-based predictability analyses after explicit group words were scrubbed. The headline result is a 55.26% increase in appearance-related descriptors when the protagonist is a girl rather than a boy. The same study reports that nationality and economy can be predicted from the text with almost 90% accuracy, that 91% of stories explicitly referenced the prompted sociocultural factor in the opening, and that the corpus remained child-friendly by Average Age of Acquisition 5.86, Flesch-Kincaid Reading Ease 75.5, and toxicity 0.06 (Rooein et al., 9 Sep 2025).

Multilingual work generalizes this design beyond English. BiasedTales-ML introduced a parallel corpus of approximately 350,000 children’s stories across eight languages using a full-permutation prompting design over nationality, religion, social class, parent role, and child gender. Narrative attributes were extracted as protagonist adjectives $A_{adj}$, environment keywords $V_{env}$, and cultural references $$d_i = \left(\frac{f(X_i)}{f(X_i)+m(X_i)}\right)\Bigg/\left(\frac{f(\mathrm{tvtropes})}{f(\mathrm{tvtropes})+m(\mathrm{tvtropes})}\right),$$0. The framework used a log-probability ratio

$$d_i = \left(\frac{f(X_i)}{f(X_i)+m(X_i)}\right)\Bigg/\left(\frac{f(\mathrm{tvtropes})}{f(\mathrm{tvtropes})+m(\mathrm{tvtropes})}\right),$$1

Jensen–Shannon divergence for bias strength, cosine similarity for cross-lingual consistency, and Valid Story Rate for quality control. The study found substantial cross-lingual variability, a consistent communality bias for female protagonists across all evaluated languages in the 8B models, higher bias strength in grammatical-gender languages such as Spanish, Russian, and Arabic, and markedly lower generation quality in Swahili, where Valid Story Rate was approximately 49% (Ouyang et al., 18 Apr 2026).

StereoTales enlarged the scope from predefined demographic axes to open-ended stereotype discovery. It covers 10 languages, 79 socio-demographic attributes, more than 650,000 stories generated by 23 recent LLMs, and more than 1,500 over-represented associations identified through contingency-table analysis, Fisher’s exact tests, bias-corrected Cramér’s $$d_i = \left(\frac{f(X_i)}{f(X_i)+m(X_i)}\right)\Bigg/\left(\frac{f(\mathrm{tvtropes})}{f(\mathrm{tvtropes})+m(\mathrm{tvtropes})}\right),$$2, and lift ratios. Roughly 15% of discovered associations were rated harmful; each model produced approximately 200 benign and 20–45 harmful associations, with a median of about 34 harmful associations per model. The study’s principal claims are that every model emitted consequential harmful stereotypes, that prompt language strongly shaped which stereotypes appeared, and that human and LLM harmfulness judgments were broadly aligned with Spearman $$d_i = \left(\frac{f(X_i)}{f(X_i)+m(X_i)}\right)\Bigg/\left(\frac{f(\mathrm{tvtropes})}{f(\mathrm{tvtropes})+m(\mathrm{tvtropes})}\right),$$3 (Jeune et al., 11 May 2026).

5. Cultural specificity and Indian-context evaluations

India-centered evaluation has produced two complementary strands: scenario-based fairness benchmarks and story-specific cultural representation audits. FairI Tales introduced INDIC-BIAS, a benchmark covering 85 identity groups across caste, religion, region, and tribe, grounded in more than 1,800 socio-cultural topics and 20,000 manually validated scenario templates. It organizes evaluation into plausibility, judgment, and generation tasks, and quantifies outcomes with ELO ratings, the Rank Shift Metric

$$d_i = \left(\frac{f(X_i)}{f(X_i)+m(X_i)}\right)\Bigg/\left(\frac{f(\mathrm{tvtropes})}{f(\mathrm{tvtropes})+m(\mathrm{tvtropes})}\right),$$4

the Stereotype Association Rate, and refusal rate. Evaluating 14 popular LLMs, the study found strong negative biases against marginalized identities, stereotype reinforcement in over 50% of relevant cases, and evidence of both allocative and representational harms. It also found that models struggled to mitigate bias even when explicitly asked to rationalize their decision, and that chain-of-thought prompting often failed to reduce bias or refusal rates and occasionally worsened outcomes (Nawale et al., 29 Jun 2025).

TALES focused directly on cultural misrepresentations in LLM-generated stories for Indian identities. TALES-Tax was derived from focus groups with $$d_i = \left(\frac{f(X_i)}{f(X_i)+m(X_i)}\right)\Bigg/\left(\frac{f(\mathrm{tvtropes})}{f(\mathrm{tvtropes})+m(\mathrm{tvtropes})}\right),$$5 and individual surveys with $$d_i = \left(\frac{f(X_i)}{f(X_i)+m(X_i)}\right)\Bigg/\left(\frac{f(\mathrm{tvtropes})}{f(\mathrm{tvtropes})+m(\mathrm{tvtropes})}\right),$$6, then used in a large-scale annotation study comprising 2,925 annotations from 108 annotators with lived cultural experience across 71 regions in India and 14 languages. The taxonomy includes Cultural Inaccuracy, Unlikely Scenarios, Clichés, Oversimplification, Factual Error, Linguistic Inaccuracy, and Logical Error. The main empirical result is that 88% of generated stories contained one or more cultural inaccuracies. Misrepresentations increased by 56% for mid-resource languages and more than tripled for low-resource languages compared with English, and they were also more prevalent in peri-urban and rural settings. TALES-QA, a question bank built from these annotations, showed a striking knowledge–generation gap: models often possessed the requisite cultural knowledge despite generating stories rife with cultural misrepresentations (Bhagat et al., 26 Nov 2025).

These India-centered studies converge on a common point. Bias in narrative generation is not exhausted by overt offensiveness; it also includes omission, flattening, overuse of outsider-facing markers, and failure to apply stored cultural knowledge in context. This suggests that culturally grounded evaluation requires lived-experience annotation, regional coverage, and open-ended generation rather than only template classification.

6. Evaluation limits, harms, and controlled alternatives

A recurring methodological result is that standard benchmark performance is a poor proxy for long-form narrative behavior. In work on gender-occupation bias, neutrality, skew, and stereotype metrics derived from decontextualized “trick tests” showed no significant correlation with analogous RUTEd evaluations on children’s bedtime stories, user personas, and English-language-learning exercises. The paper concludes that selecting the least biased model from trick tests coincides with selecting the least biased model in realistic use no more than random chance (Lum et al., 2024).

The literature also identifies several recurrent limitations. TVTropes analyses are limited to binary gender categories because of lexicon constraints, and the contributor population of TVTropes itself was reported as 64% male and 33% female with median age 20, implying possible systemic sampling skew (Gala et al., 2020). Multilingual story generation studies show that English is not a reliable proxy for other languages, and that lower-resource settings can combine lower quality with higher divergence in attribute distributions (Ouyang et al., 18 Apr 2026). Indian story audits further show that less-resourced languages can suffer simultaneously from more errors and less cultural richness, producing both “misrepresentation” and “no-representation” failure modes (Bhagat et al., 26 Nov 2025).

Mitigation results are similarly sobering. StereoTales found that negative instructions explicitly forbidding stereotypes reduced harmful associations more than generic fairness encouragement, but simply asking for fairness was insufficient (Jeune et al., 11 May 2026). Persona-assignment work reported that simple prompt-based debiasing was ineffective or could worsen bias, while only “persona + explicit expertise” substantially reduced errors, a solution the authors describe as not generalizable for open-domain settings (Gupta et al., 2023). FairI Tales likewise found that rationalization and chain-of-thought prompting did not systematically fix bias (Nawale et al., 29 Jun 2025).

One controlled alternative appears in CuentosIE, an educational chatbot built around “tales with a message” for emotional-intelligence teaching. Its tales are selected and tagged by psychologists using a taxonomy of 30 emotions and psychological themes; LLM functionality is restricted to psychologist-verified content to prevent hallucination and misinformation; and its transformer-based emotion classifier achieved 84.53% accuracy across 30 emotion classes. At the same time, the authors explicitly note curation bias, limited tale diversity, and possible emotional blind spots introduced by the specific taxonomies and the initial tale pool of approximately 50 stories during experiments (Ferrández et al., 2024).

The broader literature therefore treats biased tales as a representational, methodological, and deployment problem. Stories can reproduce stereotypical roles, moral framings, inferred motivations, and culturally localized associations; evaluation must be narrative-level, context-specific, and multilingual; and mitigation remains partial when it relies only on prompt wording rather than deeper interventions in data, modeling, and culturally grounded assessment.