Thematic Bias in Research and Media

• Thematic bias is a systematic distortion in representation and selection of themes across domains such as academic journals, news, and AI outputs.
• It encompasses phenomena from overrepresentation in mathematics journals to selective news coverage and algorithmic amplification in recommender systems.
• Studies use quantitative measures and formal models, including logarithmic scales and Bayesian methods, to diagnose and mitigate these asymmetries.

Thematic bias denotes a systematic distortion in how themes, topics, or subject matters are selected, represented, interpreted, or amplified. Across the literature, the term is not restricted to a single methodology or discipline. It refers, for example, to the publication of articles about certain mathematical subjects in quantities that are disproportionate to the production of papers about those subjects across all areas of mathematics; to the tendency of a news outlet to choose certain events to cover while ignoring others; to AI-induced thematic drift or over-simplification in qualitative coding; and to a disproportionate favoring or neglect of certain book themes in recommender system outputs (Grcar, 2010, Galeazzi et al., 2023, Turobov et al., 2024, Kalra et al., 21 Aug 2025). In adjacent work, thematic bias is also tied to framing effects, interpretive heuristics, and stereotype reproduction in generative models, indicating that the concept spans human editorial choice, computational inference, and machine generation (Sukumar et al., 30 Jun 2026, Asher et al., 2018, Porikli et al., 9 Jun 2025).

1. Definitions across domains

The term is operationalized differently depending on the object of analysis, but the common structure is asymmetry: some themes are over-represented, some are suppressed, and some are transformed by interpretation. In journal analysis, bias is statistical over- or under-representation relative to the full literature. In news analysis, the emphasis is often on selection bias, or gatekeeping bias, concerning what is covered rather than how it is framed. In qualitative research with LLMs, thematic bias concerns distortions in code generation, clustering, and theme interpretation introduced by training data, prompting, or model behavior. In recommender systems, it is a content-level bias affecting exposure to themes rather than a demographic attribute alone (Grcar, 2010, Galeazzi et al., 2023, Turobov et al., 2024, Kalra et al., 21 Aug 2025).

Domain	Operationalization	Source
Mathematics journals	Publication quantities disproportionate to subject production in all mathematics	(Grcar, 2010)
News reporting	Tendency to choose certain events to cover while ignoring others	(Galeazzi et al., 2023)
LLM-assisted thematic analysis	AI-induced thematic bias, thematic drift, over-simplification, hallucination	(Turobov et al., 2024)
Recommender systems	Disproportionate favoring or neglect of certain book themes	(Kalra et al., 21 Aug 2025)
Discourse interpretation	Selection of one “history” over another under purpose and belief constraints	(Asher et al., 2018)

A broader formalization appears in work on semantic and discourse interpretation, where interpretive bias is the set of features, constraints, and assumptions that lead an agent to select one “history” over another. In that literature, bias is not treated as an accidental defect alone; it is described as inescapable and purpose-driven, with even truth-seeking treated as a specific bias orientation rather than a neutral absence of bias (Asher et al., 2018).

2. Quantification and formal models

The literature uses several distinct measurement strategies. In generalist mathematics journals, bias is defined as the ratio between the fraction of journal papers on a subject and the fraction of all mathematics papers on that subject, and then displayed on a base-2 logarithmic scale. Positive values indicate over-representation, negative values under-representation, and zero a representative distribution (Grcar, 2010).

News studies separate narrative bias from selection bias. One line of work first classifies articles by event type and narrative stance using Google’s pre-trained BERT multilingual cased model, then estimates outlet-level latent stances and propensities with a Bayesian latent space model. The article counts $y_{ijk}$ are modeled as Poisson, with

$$\log \lambda_{ijk} = \alpha_{ik} - \| x_{ik} - z_j \|$$

where $x_{ik}$ is the outlet’s latent stance for event type $k$, $z_j$ is the ideal stance for category $j$, and $\alpha_{ik}$ is the propensity to cover event type $k$. Selection bias is then summarized through a Selection Index:

$$\text{SelectionIndex}_i = \left| \sin(\theta) PF^{(Adv)}_i - \cos(\theta) PF^{(Pos)}_i \right|$$

with $\theta = \pi/4$ (Galeazzi et al., 2023).

In recommender systems, thematic bias is quantified through theme extraction and exposure analysis. Themes are derived from book descriptions using BERTopic, each book is assigned a single dominant theme for quantitative analysis, and algorithmic amplification is measured by the Exposure Ratio:

$$\log \lambda_{ijk} = \alpha_{ik} - \| x_{ik} - z_j \|$$0

alongside item coverage, user-theme profile alignment, Gini Coefficient, and Chi-squared tests (Kalra et al., 21 Aug 2025).

In Twitter event mining, thematic context is modeled through uncertainty-weighted vectors rather than exposure imbalance. The uncertainty of keyword $$\log \lambda_{ijk} = \alpha_{ik} - \| x_{ik} - z_j \|$$1 with respect to event $$\log \lambda_{ijk} = \alpha_{ik} - \| x_{ik} - z_j \|$$2 is

$$\log \lambda_{ijk} = \alpha_{ik} - \| x_{ik} - z_j \|$$3

with $$\log \lambda_{ijk} = \alpha_{ik} - \| x_{ik} - z_j \|$$4, and the thematic context vector is

$$\log \lambda_{ijk} = \alpha_{ik} - \| x_{ik} - z_j \|$$5

This does not define bias directly, but it provides a formal way to distinguish certain from uncertain thematic associations in noisy informal text (Khatavkar et al., 2023).

3. Human institutions, editorial systems, and media selection

The classical case is topical selectivity in “generalist” institutions. In mathematics, the Proceedings of the AMS showed strong positive bias for subjects such as commutative rings, abstract harmonic analysis, and algebraic topology, while several highly active subjects were under-represented, including probability theory, combinatorics, statistics, numerical analysis, and operations research. The paper reports that among the ten most heavily published subjects, the journal is strongly biased against seven, and that a paper about algebraic topology is about 6.5 times as likely to appear in Proceedings as a combinatorics paper (Grcar, 2010).

The proposed explanation is self-perpetuation: authors submit where their fields have a history of acceptance, and editors may be more comfortable evaluating familiar topics and styles. The paper’s concern is not merely descriptive imbalance but institutional fragmentation, since generalist journals may give readers and the community a misleading view of which topics dominate mathematics (Grcar, 2010).

In news reporting, thematic bias is often operationalized as selection bias. A vaccine-debate study distinguishes narrative bias, concerning how events are framed, from selection bias, concerning what gets covered. It finds that third-party source classifications closely follow the narrative bias dimension, while they are much less accurate in identifying the selection bias. It also reports a nonlinear relationship between bias and engagement, with higher engagement for extreme positions, and shows that outlets with similar ideological positions share common audiences on Twitter (Galeazzi et al., 2023).

A complementary headline study distinguishes bias from diversity. Using multiple correspondence analysis on 1.8 million U.S. headlines, it concludes that discrepancies in domestic politics and social issues can be attributed to a certain degree of media bias, that foreign-affairs discrepancies are largely attributable to diversity in individual journalistic styles, and that economic coverage is highly consistent across outlets. The same study uses frequent $$\log \lambda_{ijk} = \alpha_{ik} - \| x_{ik} - z_j \|$$6-grams to show how framing differences can align with ideological stance, such as “abortion rights” versus “abortion law” (Pan et al., 2023).

4. LLM-mediated thematic interpretation

In qualitative thematic analysis, thematic bias shifts from researcher subjectivity to a hybrid problem involving model priors, prompting, and verification. A study using GPT for initial coding in qualitative thematic analysis states that ChatGPT may reproduce biases present in its training data, over-rely on explicit content instead of deeper context, and generate “descriptive” rather than “interpretive” outputs. The paper reports quotation errors, paraphrased rather than verbatim quotes, and code-naming bias, including a case where a passage about “the design and accountability of autonomous intelligent systems” was labeled “AI Governance” even though the authors judged “AI Design” or a more precise code more appropriate (Turobov et al., 2024).

That study’s workflow is explicitly hybrid. GPT assists with initial coding, while theme development and interpretation remain the researcher’s responsibility. Mitigation is based on manual review, highly structured prompting, a stepwise process, transparency in instructions and workflow, and cross-validation with Latent Dirichlet Allocation topic modeling to check whether AI-derived themes align with, or diverge from, statistical groupings of concepts (Turobov et al., 2024).

A later workshop study in Software Engineering extends these concerns. Participants warned that LLMs may produce themes based on previous data, that different LLMs have different biases, that short-quote reading can cause loss of context, and that prompting strategy directly affects code and theme generation. The recommendations emphasize hybrid human-LLM workflows, comparative use of multiple LLMs, manual coding followed by LLM theme generation and human refinement, full prompt and model-version documentation, and a skeptical stance toward model outputs (Ornelas et al., 18 Nov 2025).

Bias also appears in abstractive interpretation. In a study of life-narrative summarization, LLMs are treated as interpreters whose summaries differ across race and gender. The proposed positionality portrait pipeline compares summaries with and without explicit demographic prompts using ROUGE-1/L, BERTScore, LIWC, Valence-Arousal-Dominance scores, Stereotype Content Model projections, and human evaluation by psychologists. The study reports race and gender bias with the potential for representational harm, including lower semantic overlap for Black participants and altered emotional framing for men, especially Black men (Subbiah et al., 22 Apr 2026).

A related topic-model evaluation study situates thematic bias at the level of human interpretation itself. Using reflexive thematic analysis on topic rationales, it finds that users interpret topics through availability and representativeness heuristics rather than probability and proposes anchoring-and-adjustment as a theory of topic interpretation. On this account, topic interpretation is a judgment under uncertainty by an ecologically rational user, implying that user disagreement is not merely noise but a property of thematic interpretation (Hingmire et al., 25 Jul 2025).

5. Algorithmic amplification in recommendation and generation

In recommender systems, thematic bias is framed as a system-level exposure problem. A book-recommendation study using the Book-Crossing dataset finds that thematic bias originates from content imbalances and is amplified by user engagement patterns. By segmenting users along popularity orientation and thematic diversity, it reports that users with niche and long-tail interests receive less personalised recommendations, while users with diverse interests receive more consistent recommendations. The most affected group is Specialist readers with long-tail patterns; recommendation outputs tend to homogenize users toward dominant themes rather than preserve their original thematic profiles (Kalra et al., 21 Aug 2025).

Generative systems exhibit a different but related phenomenon: themes become carriers of stereotypes. A text-to-image study curated 160 unique topics across occupations, traits, actions, ideologies, emotions, family roles, place descriptions, spirituality, and life events, then generated over 16,000 images with Stable Diffusion 1.5 and Flux-1. It reports strong disparities in the representation of gender, race, age, somatotype, and geography, including >90% Middle Eastern depictions for “terrorist,” 99% Western settings for positive place prompts, and systematic gender skew in occupations and actions (Porikli et al., 9 Jun 2025).

Story generation exhibits analogous thematic effects. The “Biased Tales” dataset analyzes 5,531 LLM-generated children’s stories and finds that when the protagonist is described as a girl rather than a boy, appearance-related attributes increase by 55.26%. It also reports that stories featuring non-Western children disproportionately emphasize cultural heritage, tradition, and family themes, while Western or White protagonists are more associated with magical settings and adventurous framing (Rooein et al., 9 Sep 2025).

Bias can also be measured in cultural corpora rather than in model outputs alone. A song-lyrics study clusters 537,553 English songs with BERTopic and then uses SC-WEAT to quantify gender bias in topic- and genre-specific embeddings. It reports a historical thematic shift from romantic themes toward a heightened focus on the sexualization of women, substantial misogynistic and profane content in the largest thematic cluster, and persistent male associations with intelligence and strength versus female associations with appearance and weakness (Chen et al., 2024).

6. Related constructs: coherence, discrepancy, framing, and consistency control

Several adjacent constructs clarify what thematic bias is and is not. Thematic coherence concerns whether textual units belong together around a common subject. In fake-news detection, coherence is measured by comparing topic distributions between an article’s opening sentences and the remainder using Chebyshev, Euclidean, and squared Euclidean distances. Across seven cross-domain datasets, fake news shows greater thematic deviation between opening and body than truthful news, making thematic deviation a useful detection feature rather than a direct measure of bias (Dogo et al., 2020).

In microblogs, thematic coherence is defined as the extent to which posts belong together so that experts can easily extract and summarize the common stories underpinning them. Journalist annotations and automated evaluation show that text generation metrics such as BERTScore, BLEURT, and MoverScore are more reliable than topic coherence metrics and less sensitive to time-window effects. This suggests that semantic coherence and thematic bias are related but distinct: the former measures internal topical unity, whereas the latter concerns selective amplification, omission, or distortion (Bilal et al., 2021).

Thematic framing is another neighboring concept. In visualization research, thematic framing foregrounds broader patterns or aggregate statistics, whereas episodic framing foregrounds a specific vivid event. In a preregistered experiment on U.S. mass-shooting visualizations, episodic framing elicited significantly more negative emotional valence than both thematic conditions, and the episodic condition was perceived as more pro–gun control despite identical data. The study therefore treats framing as a route by which the same evidence can appear more or less neutral (Sukumar et al., 30 Jun 2026).

Finally, some systems are designed explicitly to reduce thematic drift rather than to diagnose bias post hoc. In thematic collection design, iCONTRA uses an interactive concept-transfer interface and a zero-shot image-editing algorithm to anchor generated objects to an original object’s structure, color, style, and background. The stated goal is thematic consistency across a collection, with foreground masking and fade-in attention used to minimize background changes and avoid broken thematic continuity (Vo et al., 2024).

7. Mitigation, transparency, and open problems

Across these literatures, mitigation strategies converge on transparency, comparative validation, and preservation of human interpretive authority. In LLM-assisted thematic analysis, the recurring recommendations are hybrid workflows, manual verification of codes and quotations, prompt engineering, stepwise documentation, and triangulation with alternative methods such as LDA topic modeling or multiple LLMs. In life-narrative summarization, the recommended instrument is a positionality portrait that makes the model’s perspective visible through quantitative and human evaluation. In topic interpretation research, bias-aware user models, multi-construct evaluation, and “Topic Cards” are proposed to document interpretive diversity rather than suppress it (Turobov et al., 2024, Ornelas et al., 18 Nov 2025, Subbiah et al., 22 Apr 2026, Hingmire et al., 25 Jul 2025).

For generative and retrieval systems, the recommended interventions operate earlier in the pipeline. Text-to-image work emphasizes more inclusive datasets and standardized multi-category evaluation. Recommender-system work argues for theme-centric fairness analysis, user-aligned calibration, and subgroup analysis rather than global averages. In journal and media settings, the proposed remedy is public discussion of whether existing topical selectivity is appropriate or desirable, along with awareness that current evaluative systems may miss selection bias even when they track narrative bias (Porikli et al., 9 Jun 2025, Kalra et al., 21 Aug 2025, Grcar, 2010, Galeazzi et al., 2023).

A plausible synthesis is that thematic bias is best understood as a family of asymmetries in thematic representation whose mechanisms vary by setting: editorial gatekeeping, institutional self-reinforcement, prompt-induced abstraction, latent training-data priors, engagement amplification, or stereotype-laden generation. The cited work does not reduce these mechanisms to a single causal theory. Instead, it shows that rigorous analysis of thematic bias typically requires domain-specific operationalization, explicit comparison baselines, and documentation of the interpretive pipeline through which themes are produced, selected, or made visible.