Language Bubbles in Social Networks & LLMs

• Language bubbles are self-reinforcing linguistic enclaves characterized by narrow interaction patterns and reduced lexical diversity across diverse domains.
• Empirical studies link social network segregation, via retweet and bipartite graph analyses, to measurable declines in vocabulary richness and lexical complexity.
• Extensions to LLM outputs and multilingual models reveal that personalization and language modality bias can constrain linguistic expression and affect educational outcomes.

Language bubbles are self-reinforcing linguistic enclaves in which the language available to a community narrows as interaction patterns become more insular. In the formal sense introduced by Bellina et al. for online social networks, a language bubble is the co-occurrence of social fragmentation—communities that retweet almost exclusively within their group—and linguistic divergence and impoverishment—those same communities use ever more specialised and poorer lexicons (Bellina et al., 17 Jul 2025). Related work extends the term by analogy to user-personalized LLM outputs, multilingual reasoning systems, and English-centric programming environments, where personalization, latent-space routing, or institutional language norms can similarly constrain what is said, how it is framed, and which linguistic resources remain available (Lazovich, 2023, Schut et al., 21 Feb 2025, Prather et al., 2024). Taken together, these usages suggest a family of phenomena linking restricted interaction to restricted linguistic range.

1. Conceptual range and terminological usage

The expression language bubble does not denote a single, uniform formalism across current literatures. In Bellina et al., it is a network-linguistic construct grounded in retweet segregation and lexical metrics (Bellina et al., 17 Jul 2025). In Lazovich, the term is used informally by analogy for an LLM-generated echo chamber in which conditioning on user metadata shifts outputs in favor of ideologically aligned content (Lazovich, 2023). In Schut et al., the relevant construct is a language-centric latent space, described as an English-shaped bubble through which semantically loaded tokens are routed even when both input and output are non-English (Schut et al., 21 Feb 2025). In computing-education work on multilingual prompting, language bubbles refer to the insular environment created when programming languages, instructional materials, and classroom discourse are predominantly English-centric (Prather et al., 2024).

Domain	Bubble mechanism	Reported effect
Online social networks	Retweet fragmentation plus lexical divergence	Distinct and poorer linguistic profiles
Personalized LLM outputs	System-prompt conditioning on user ideology	More positive language for aligned entities
Multilingual LLMs	English modality bias or English-centric latent routing	Better performance or stronger control in English
Programming education	English-dominant instructional and coding discourse	NNES students struggle to articulate programming knowledge in English

A common misconception is that language bubbles are reducible to political polarization alone. Bellina et al. analyze six topic-specific corpora—immigration, vaccines, climate, sport, music, and cars—showing that the phenomenon is not limited to overtly political topics (Bellina et al., 17 Jul 2025). Another misconception is that bubbles are only external properties of social graphs. The LLM studies indicate that analogous effects can arise inside model behavior itself, through output conditioning, language-modality bias, or latent-space routing (Lazovich, 2023, Schut et al., 21 Feb 2025).

2. Formalization in online social networks

Bellina et al. operationalize language bubbles on Twitter/X using tweets and retweets from Italian politicians and news outlets between 2018 and 2022. The dataset contains approximately 14 million tweets from 583 Italian influencers and all their retweeters, stratified into six topic-specific corpora: immigration, vaccines, climate, sport, music, and cars (Bellina et al., 17 Jul 2025).

For each topic, they construct a bipartite graph with influencers on one layer and retweeters on the other. The biadjacency matrix $W_{kj}$ counts how many times account $k$ retweets influencer $j$. Projection onto the influencer layer yields the weighted similarity

$$w_{ij}=\frac{\sum_k W_{ki}W_{kj}}{\sqrt{\sum_k W_{ki}^2\cdot \sqrt{\sum_k W_{kj}^2}}}.$$

Statistically insignificant edges are pruned, and the analysis retains the largest connected component, whose size ranges from 96 to 135 nodes depending on topic. Community detection is performed primarily with the hierarchical Stochastic Block Model (hSBM). After 100 Monte Carlo runs, a consensus partition is formed at the second level of the hierarchy, and modules with fewer than 10 nodes are merged to minimize description length. As a robustness check, the Louvain algorithm is applied with the same small-cluster merging rule, yielding qualitatively identical results (Bellina et al., 17 Jul 2025).

The linguistic side of the framework begins with lemmatization, stopword removal, and exclusion of the top 1,000 “kernel” words, defined as the most frequent tokens and reported to follow Zipf’s law with exponent approximately $-1$. Each user’s empirical token distribution is

$$r_i=\{r_{i\alpha}\}, \qquad r_{i\alpha}= \frac{M_{i\alpha}}{\sum_\beta M_{i\beta}},$$

where $M_{i\alpha}$ is the validated count of token $\alpha$ by user $i$. Lexical variety is measured by the type-token ratio,

$\mathrm{TTR}=\frac{V}{N},$

for text length $k$0 with $k$1 distinct types. Lexical complexity is measured by Shannon entropy,

$k$2

Similarity to the global distribution $k$3 is measured by completeness,

$k$4

Inter-community lexical distance is based on Jensen–Shannon divergence:

$k$5

and for communities $k$6 with $k$7 users,

$k$8

Inter-community network distance is defined by first transforming similarities into distances $k$9 with $j$0, then computing shortest-path distances $j$1 and

$j$2

bounded in $j$3. Community segregation is

$j$4

and topic fragmentation is

$j$5

This formalization makes the notion of a language bubble explicitly bivariate: it is not only about who interacts with whom, and not only about which words are used, but about the structural coupling between graph separation and lexical contraction.

3. Empirical signatures: proximity, segregation, and lexical impoverishment

Bellina et al. report two systematic empirical patterns across all six topics (Bellina et al., 17 Jul 2025). First, pairs of communities that are closer in the retweet network exhibit smaller lexical distance. Spearman’s $j$6 between $j$7 and $j$8 ranges from $j$9 in music to $$w_{ij}=\frac{\sum_k W_{ki}W_{kj}}{\sqrt{\sum_k W_{ki}^2\cdot \sqrt{\sum_k W_{kj}^2}}}.$$0 in climate, and all correlations are highly significant with $$w_{ij}=\frac{\sum_k W_{ki}W_{kj}}{\sqrt{\sum_k W_{ki}^2\cdot \sqrt{\sum_k W_{kj}^2}}}.$$1 against a null model in which the bipartite network is randomly rewired. Concrete examples include immigration, with $$w_{ij}=\frac{\sum_k W_{ki}W_{kj}}{\sqrt{\sum_k W_{ki}^2\cdot \sqrt{\sum_k W_{kj}^2}}}.$$2 and $$w_{ij}=\frac{\sum_k W_{ki}W_{kj}}{\sqrt{\sum_k W_{ki}^2\cdot \sqrt{\sum_k W_{kj}^2}}}.$$3, and sport, with $$w_{ij}=\frac{\sum_k W_{ki}W_{kj}}{\sqrt{\sum_k W_{ki}^2\cdot \sqrt{\sum_k W_{kj}^2}}}.$$4 and $$w_{ij}=\frac{\sum_k W_{ki}W_{kj}}{\sqrt{\sum_k W_{ki}^2\cdot \sqrt{\sum_k W_{kj}^2}}}.$$5.

Second, more isolated communities consistently exhibit poorer lexicons. Within each topic, higher community segregation $$w_{ij}=\frac{\sum_k W_{ki}W_{kj}}{\sqrt{\sum_k W_{ki}^2\cdot \sqrt{\sum_k W_{kj}^2}}}.$$6 is associated with lower lexical richness under multiple measures. For vocabulary size $$w_{ij}=\frac{\sum_k W_{ki}W_{kj}}{\sqrt{\sum_k W_{ki}^2\cdot \sqrt{\sum_k W_{kj}^2}}}.$$7, the reported Spearman coefficients range from $$w_{ij}=\frac{\sum_k W_{ki}W_{kj}}{\sqrt{\sum_k W_{ki}^2\cdot \sqrt{\sum_k W_{kj}^2}}}.$$8 in immigration ($$w_{ij}=\frac{\sum_k W_{ki}W_{kj}}{\sqrt{\sum_k W_{ki}^2\cdot \sqrt{\sum_k W_{kj}^2}}}.$$9) and $-1$0 in vaccines ($-1$1) down to $-1$2 in sport ($-1$3). For entropy $-1$4, the range is from $-1$5 in immigration ($-1$6) and $-1$7 in vaccines ($-1$8) down to $-1$9 in sport ($$r_i=\{r_{i\alpha}\}, \qquad r_{i\alpha}= \frac{M_{i\alpha}}{\sum_\beta M_{i\beta}},$$0). For completeness $$r_i=\{r_{i\alpha}\}, \qquad r_{i\alpha}= \frac{M_{i\alpha}}{\sum_\beta M_{i\beta}},$$1, the range is from $$r_i=\{r_{i\alpha}\}, \qquad r_{i\alpha}= \frac{M_{i\alpha}}{\sum_\beta M_{i\beta}},$$2 in immigration ($$r_i=\{r_{i\alpha}\}, \qquad r_{i\alpha}= \frac{M_{i\alpha}}{\sum_\beta M_{i\beta}},$$3) and $$r_i=\{r_{i\alpha}\}, \qquad r_{i\alpha}= \frac{M_{i\alpha}}{\sum_\beta M_{i\beta}},$$4 in vaccines ($$r_i=\{r_{i\alpha}\}, \qquad r_{i\alpha}= \frac{M_{i\alpha}}{\sum_\beta M_{i\beta}},$$5) down to $$r_i=\{r_{i\alpha}\}, \qquad r_{i\alpha}= \frac{M_{i\alpha}}{\sum_\beta M_{i\beta}},$$6 in sport ($$r_i=\{r_{i\alpha}\}, \qquad r_{i\alpha}= \frac{M_{i\alpha}}{\sum_\beta M_{i\beta}},$$7). All of these correlations remain significant under the null-rewiring test.

A further result concerns topic-level fragmentation. More highly fragmented topics, i.e. topics with higher $$r_i=\{r_{i\alpha}\}, \qquad r_{i\alpha}= \frac{M_{i\alpha}}{\sum_\beta M_{i\beta}},$$8, exhibit stronger negative $$r_i=\{r_{i\alpha}\}, \qquad r_{i\alpha}= \frac{M_{i\alpha}}{\sum_\beta M_{i\beta}},$$9 between segregation and lexical richness. Bellina et al. therefore place the strongest language bubbles in the most polarized debates, especially immigration and vaccines, where the poorest language appears in the most insular niches (Bellina et al., 17 Jul 2025).

The authors interpret these results as evidence that echo chambers in retweet networks create self-reinforcing linguistic enclaves in which communities evolve their own “dialect” of the topic at hand. The reported mechanism is structural rather than merely attitudinal: socially isolated communities interact less with others and develop distinct and poorer linguistic profiles. A plausible implication is that fragmentation changes not only exposure to viewpoints, but also the lexical capacity available for expressing them.

4. Personalized LLM outputs and affective alignment

Lazovich studies a distinct but related phenomenon in user-personalized LLM outputs. The model is ChatGPT-3.5 via the OpenAI API in June 2023, and personalization is introduced solely through the system prompt. The simple condition uses “The user is a Democrat.” or “The user is a Republican.” The fine-grained condition uses five prompts ranging from “very liberal” to “very conservative.” For each condition, the user prompt is the factual request “Tell me about <E>.” The study covers presidential candidates in U.S. elections 2000–2022, all 101 members of the 2019 U.S. Senate, and 39 media outlets rated by AllSides, with 100 independent generations per prompt (Lazovich, 2023).

The quantitative metric is sentiment. A DistilBERT model fine-tuned on SST-2 outputs a positive probability $M_{i\alpha}$0 for each answer, and the mean positive-sentiment score is

$M_{i\alpha}$1

In the simple experiment, the study compares $M_{i\alpha}$2 and $M_{i\alpha}$3, and reports a normalized $M_{i\alpha}$4-score for the difference. In the fine-grained analysis, entities are assigned numerical leanings $M_{i\alpha}$5 and user prompts numerical values $M_{i\alpha}$6, and sentiment is examined as a function of $M_{i\alpha}$7.

The reported pattern is that left-leaning users tend to receive more positive statements about left-leaning political figures and media outlets, while right-leaning users receive more positive statements about right-leaning entities. Typical differences in mean sentiment are described as roughly $M_{i\alpha}$8–$M_{i\alpha}$9 on the $\alpha$0 scale for presidential candidates, and the figure inset often shows $\alpha$1. For senators, Republican figures receive roughly $\alpha$2 higher sentiment for Republican users, with the reverse for Democratic senators. For media outlets, right and lean-right outlets enjoy substantially higher sentiment, approximately $\alpha$3, for Republican users, while left and lean-left outlets show smaller but still positive differences of approximately $\alpha$4–$\alpha$5 for Democrat users. In the fine-grained analysis, sentiment declines roughly linearly as $\alpha$6 grows, with an asymmetry such that entities to the right of a user’s position experience sharper sentiment drops.

The qualitative examples are equally important. For Donald Trump, the Republican-user output omits impeachment and false-statement controversies, whereas the Democrat-user output explicitly mentions “controversial immigration policies,” trade disputes, two impeachments, and “making false statements.” For Senator Cindy Hyde-Smith, the Republican-user output omits the “public hanging” remark and racial controversies, while the Democrat-user output foregrounds them. For The Wall Street Journal, the Republican-user output states that its editorial page may align with the user’s political views, whereas the Democrat-user output states that some of its views may not align with the user’s political beliefs (Lazovich, 2023).

Within this literature, language bubble is an informal extension rather than a formal theorem. The operational definition supplied in the synthesis is that a language bubble exists if conditioning an LLM on user metadata causes the statistical distribution of outputs about entities to shift in a way that favors ideologically aligned content. That framing places lexical selection, fact inclusion, and evaluative framing inside the same bubble logic previously associated with recommender systems and filter bubbles.

5. Multilingual LLMs and English-centric latent spaces

Recent multilingual LLM studies broaden the concept further by identifying bubbles tied to language modality and internal representation. One strand concerns performance disparities across high-resource and low-resource languages. In “Breaking Language Barriers: Equitable Performance in Multilingual LLMs,” the problem is stated as worse Common Sense Reasoning performance in low-resource languages such as Hindi compared to high-resource languages such as English. The proposed remedy is fine-tuning on synthetic code-switched text derived from CommonSenseQA, using GPT-3.5 prompting and the CoCoa model with controlled Code-Mixing Index regimes $\alpha$7, $\alpha$8, and $\alpha$9 (Nagar et al., 18 Aug 2025).

The model architecture is LLaMA-3-8B-Instruct with QLoRA quantization, trained for 5 epochs with batch size $i$0 and Adam at learning rate $i$1. Accuracy is reported separately for English-prompted and Hindi-prompted versions. The medium-mixing condition $i$2 yields the largest gains, improving English accuracy from $i$3 to $i$4 and Hindi accuracy from $i$5 to $i$6. None of the variants degrade English performance below baseline (Nagar et al., 18 Aug 2025).

Variant	English Accuracy	Hindi Accuracy
Baseline	78.0%	54.0%
GPTgen	88.8%	79.6%
$i$7	81.6%	75.2%
$i$8	90.4%	85.6%
$i$9	87.2%	77.2%

A second strand examines abstract reasoning. “Think Globally, Group Locally” introduces GlobalGroup, a multilingual word-grouping benchmark in English, Spanish, Chinese, Hindi, and Arabic, together with English translations and additional New York Times Connections-derived games. The paper reports a “Law of Translation”: for Spanish, Hindi, and Arabic, group-level F1 increases by 5–10 points after translation into English, whereas Chinese shows a mixed pattern in which closed-source models perform better in Chinese and open-source models perform better in English. The benchmark also finds that non-cultural groups are systematically easier than culturally related groups, and that English modalities largely lead to better performance in this abstract reasoning task (Guerra-Solano et al., 15 Oct 2025). The authors describe this as an “English bubble” for abstract pattern-finding.

A third strand moves from behavior to mechanism. Schut et al. examine internal representations with a logit lens and show that multilingual LLMs make key decisions in a representation space closest to English, regardless of input and output language. For French, German, Dutch, and Mandarin, approximately 50%–70% of nouns and verbs are first represented as English tokens before being mapped to the target language. For Llama-3.1-70B, the average English routing is reported as about 55%; for Gemma-2-27b, about 70%. Determiners, adpositions, and conjunctions route via English less than 20% of the time. The same work shows that activation steering vectors computed in English are about 10–20 percentage points more effective than non-English steering vectors for injecting target concepts into non-English outputs, and that interpolating hidden states for the same fact across languages preserves accuracy but biases the final output toward English (Schut et al., 21 Feb 2025).

These multilingual results indicate that language bubbles need not be confined to explicit social communities. They can also appear as modality-dependent performance envelopes, culturally local knowledge boundaries, or English-shaped internal pathways that are opaque to system users. This suggests that the bubble metaphor now spans both external interaction structures and internal model geometry.

6. Educational barriers, interventions, and unresolved questions

In computing education, language bubbles are defined as the insular environment created when programming languages, instructional materials, and classroom discourse are predominantly English-centric. The reported consequence is that non-native English-speaking students may know programming concepts but struggle to articulate them in English, avoid participation, or hesitate to ask for help for fear of linguistic error. The authors describe this as both a cognitive barrier, because students must translate ideas into English, and a social barrier, because engagement occurs in a classroom dominated by native English speakers (Prather et al., 2024).

The empirical study covers Arabic, Chinese, and Portuguese cohorts working on three prompt problems: scramble, arrange, and speak. The dataset includes 1,679 non-English prompts collected iteratively until generated code passed the test suite. Completion rates vary substantially by cohort and problem: Portuguese reports 95% on all three problems; Chinese reports 100% on Problem 1, 42% on Problem 2, and 50% on Problem 3; Arabic reports 64%, 58%, and 86%, respectively. Native-language prompt success rates in the “N” category are approximately 5% for Arabic, 15% for Chinese, and 18% for Portuguese. The dominant strategy is native-only prompting, with mixed and English-heavy prompting secondary. Qualitative reports emphasize poor support for some languages, a trade-off between expressivity and performance, and the difficulty of naming functions and variables in a native language when model behavior favors English conventions (Prather et al., 2024).

Across the literatures, proposed interventions target different levels of the system. Bellina et al. discuss algorithmic recommendations that favor cross-community retweets, “bridge” accounts or bots that deliberately share content across clusters, and platform or policy features promoting exposure to varied vocabularies or summary digests of other communities’ language (Bellina et al., 17 Jul 2025). Lazovich recommends monitoring and auditing LLM personalization pipelines, extending analyses beyond politics to other demographic signals, and designing safeguards so that demographic metadata does not bleed into factual queries (Lazovich, 2023). The multilingual CSR work proposes controlled synthetic code-switching as a way to improve low-resource language performance without degrading high-resource language performance (Nagar et al., 18 Aug 2025). The computing-education study proposes bilingual comments and keywords, hybrid prompt strategies, explicit glossaries, and prompt problems as low-stakes reflective exercises (Prather et al., 2024).

Several unresolved questions remain. Some studies provide formal network and divergence-based definitions, while others use the term analogically or informally. Some studies report significance through null models or confidence bars, whereas others explicitly state that statistical significance was not formally tested (Bellina et al., 17 Jul 2025, Lazovich, 2023, Nagar et al., 18 Aug 2025). This suggests that language bubble currently functions less as a single settled theoretical construct than as a cross-domain descriptor for structurally constrained linguistic behavior. Even so, the recurrent pattern is consistent: when exposure, routing, or institutional norms become narrow, language itself becomes narrower, more aligned with local priors, and less connected to a shared discourse.