The Generative AI Divide on X: Themes, Frames, and Discursive Shifts in High-Visibility Posts

From the Digital Divide to the Generative AI Divide

The concept of the digital divide has long served as one of the main analytical frameworks for explaining the social consequences of information and communication technologies. In its earliest form, the concept was defined largely in terms of physical access to technology. Limited infrastructure, connectivity constraints, hardware costs, and device ownership were treated as the primary drivers of technological inequality (Norris, 2001; van Dijk, 2005).

That initial framing had real value; it made the measurable dimensions of technological inequality visible and comparable across contexts. But it also had limits. Over time, researchers recognized that access alone could not explain the full range of social consequences that digital technologies produced. As a result, digital divide research developed into a more layered field through second and third-level divide approaches. The second-level divide concerns individuals’ digital skills, patterns of use, and capacities to interpret and work with technology, while the third-level divide focuses on the unequal distribution of the economic, social, and cultural benefits that arise from digital technology use (Hargittai, 2002; van Deursen & van Dijk, 2014).

The spread of GenAI tools has pushed these debates into unfamiliar territory. These systems do more than save time or increase productivity. They also shape which forms of knowledge become visible, credible, and useful (Bender et al., 2021; Birhane, 2021; Crawford, 2021; Noble, 2018). For this reason, the GenAI divide should not be treated simply as an extension of earlier access-based models. Recent work suggests that access to generative AI is not equivalent to the ability to use it effectively, evaluate it critically, and derive meaningful benefit from it. On top of that, geographic and socioeconomic differences may be reproducing familiar patterns of stratification in ways we have not fully mapped yet (Beckman et al., 2025; Daepp & Counts, 2025). The GenAI divide, in other words, is better understood as a multidimensional field of inequality that includes not only access but also the quality of use and the capacity to derive advantage from it.

Structural Axes of the GenAI Divide

Recent scholarship suggests that GenAI-related inequality extends beyond differences in access and individual skill. It moves toward broader questions of sociotechnical governance, epistemic justice, and institutional capacity (Aristombayeva et al., 2025; Barry & Stephenson, 2025; Jin et al., 2025; Kay et al., 2024). This study, therefore, approaches the GenAI divide through three analytical axes.

Governance matters because institutions differ in their ability to regulate, guide, and support the ethical and effective use of GenAI (Floridi et al., 2018; Jin et al., 2025; Mittelstadt et al., 2016). Epistemic justice matters because these systems can privilege some languages, contexts, and knowledge traditions while sidelining others (Barry & Stephenson, 2025; Bender et al., 2021; Birhane, 2021; Crawford, 2021; Fricker, 2007; Kay et al., 2024; Noble, 2018). And institutional capacity matters because access alone does not determine who benefits from GenAI in educational, social, or economic contexts. Infrastructure, human resources, and governance readiness all shape outcomes in ways that individual skill cannot compensate for (Beckman et al., 2025; Jin et al., 2025). Institutional capacity is also visible in how higher education systems organize AI ecosystems through research centers, human-resource development, and collaborations with public or industrial actors (Serpil & Kesim, 2024).

In this study, these axes are not treated as fixed subdimensions of the GenAI divide. Rather, they are approached as analytical axes gaining visibility in the literature. They can also be traced in high-visibility discourse on X. This perspective allows discussion of GenAI-related inequalities beyond differences in access and skill. It places them within broader political, cultural, and institutional contexts.

Research Design

This study adopts an interpretive qualitative design with structured descriptive quantification to examine high-visibility public discussions of the “GenAI divide” on X. The primary aim is not to estimate the distribution of public opinion or to produce inferential claims about platform-wide communication. Rather, the study analyzes a purposively assembled corpus of high-visibility posts to identify how GenAI-related inequality is constructed, framed, and circulated in publicly visible discourse.

The analytical core of the study is qualitative. In this formulation, the “interpretive qualitative design” refers to the study’s emphasis on meaning-making, framing, and the contextual interpretation of high-visibility discourse. Inductive thematic analysis was used to identify the substantive dimensions through which the GenAI divide was discussed, and discourse analysis was then employed to examine how these dimensions were framed, legitimized, and rhetorically intensified. “Structured descriptive quantification” refers to the limited use of counts, percentages, co-occurrence summaries, and year-to-year similarity measures derived from qualitative coding. These numerical summaries were used as analytic aids for making temporal and relational patterns within the corpus more visible; they were not used to test hypotheses, estimate population parameters, or support inferential claims. In this sense, the study follows qualitative content analysis approaches in which systematic coding and limited counting can strengthen transparency and pattern recognition without changing the study into a quantitative design (Schreier, 2012; Krippendorff, 2018).

Data Source and Data Collection Tool

The data were gathered through tool-assisted retrieval on X via Grok, combining keyword-based and semantic searches. Searches were conducted separately for each year from 2022 through December 2025, and results were ranked by total engagement. Off-topic items, duplicates, and repeated discursive content within the same thread were removed, with only the highest-engagement version retained. When necessary, the surrounding thread context was checked. The resulting dataset, therefore, consisted of analytically meaningful, high-visibility posts related to GenAI divide discussions on X.

Search Strategy and Scope

The search strategy was designed to capture conceptual variants, multilingual expressions, and indirect discursive equivalents that could signal discussion of the GenAI divide. It included not only direct labels such as “GenAI divide,” “generative AI divide,” “GenAI gap,” and “AI divide,” but also broader inequality-related expressions and semantically related prompts intended to identify discursively relevant posts beyond exact lexical matches. No platform-level language filter was imposed. However, the retrieval strategy did not aim at exhaustive coverage of all world languages. Instead, it prioritized a set of languages selected based on X’s linguistic distribution and expected relevance for high-visibility GenAI discussions: English as the primary search language, followed by Japanese, Spanish, Arabic, Turkish, and German. English functioned as the primary retrieval language, while Japanese, Spanish, Arabic, Turkish, and German were searched through parallel query rounds using the same three-layer query logic. This decision was also informed by preliminary retrieval rounds, which showed that high-visibility GenAI divide discussions with broad cross-national circulation were predominantly English-language or circulated through English technical terminology. As a result, the corpus should be understood as multilingual but English-dominant rather than as a fully balanced cross-linguistic sample. Appendix A reports the language strategy, English reference query set, and the overall retrieval logic in detail.

Because of the volume and instability of content on X, the study used a purposive, year-stratified, high-engagement sampling strategy rather than attempting a full census of all relevant posts. The dataset was therefore built from posts that achieved comparatively high visibility within each yearly search round. Engagement was operationalized as the sum of likes, reposts, quote posts, and replies. At the beginning of the search process, higher engagement thresholds were applied. When some queries failed to yield a sufficient number of relevant results, however, the thresholds were gradually relaxed, query variants were expanded, and semantic search support was increased. Even so, total engagement remained the basis for ranking results throughout the process.

Importantly, this engagement metric was used as a within-year ranking criterion rather than as a directly comparable measure of absolute visibility across years. Because platform dynamics, account networks, content formats, and algorithmic amplification may change over time, the study does not assume that an engagement score in one year is equivalent to the same score in another year. The term “high visibility” therefore refers to posts that achieved comparatively high engagement within each yearly retrieval round, not to a platform-wide or cross-year standardized visibility measure.

To reduce the risk that year-to-year differences reflected a changing retrieval regime rather than discursive variation within the corpus, all yearly searches followed the same core retrieval logic. The three-layer query family described in Appendix A was used as the stable conceptual frame for each year and each target language. Adaptive steps, such as relaxing engagement thresholds, expanding localized query variants, or increasing semantic recall, were used only when the initial retrieval returned too few relevant candidate posts within a given year. These steps were not used to introduce new conceptual domains into the corpus; rather, they broadened recall within the same GenAI divide scope. All candidate posts, regardless of retrieval round, were then subjected to the same relevance screening, de-duplication, thread-level consolidation, and engagement-based ranking procedures. For this reason, temporal comparisons are interpreted as descriptive patterns within the retained corpus rather than as direct estimates of platform-wide discursive change.

In the final stage of sampling, the aim was to retain approximately 100 high-engagement posts per year to improve year-to-year comparability. After relevance screening, de-duplication, and thread-level consolidation, the final retained corpus consisted of 32 posts for 2022, 100 for 2023, 98 for 2024, and 102 for 2025. The lower number in 2022 reflects the topic’s relative novelty and lower visibility during the early period. The slight deviation in 2024 reflects the final relevance and de-duplication decisions applied to the retrieved set, while the slight increase in 2025 reflects the retention of tied posts at the final engagement threshold. For this reason, the 2022 corpus was treated as an exploratory early-period baseline rather than as a fully comparable yearly stratum. Proportions for 2022 were interpreted descriptively and with caution in comparisons with later years, and small absolute differences were not overinterpreted.

Where the same thread contained repeated content, only the highest-engagement post was retained, while duplicate posts reproducing the same discourse were excluded. The study, therefore, focuses not on a social media sample intended to represent the general public or the full universe of X posts, but on discursive patterns that achieved comparatively high visibility on X in discussions of the GenAI divide. This sampling choice offers an analytical advantage for examining influential, widely circulated, high-visibility discourse on X. However, it also risks bias by providing limited representation for lower-visibility or niche counter-discourses.

Data Verification and Auditability

Before being included in the analysis, content gathered through Grok was systematically recorded, and each post was entered into the dataset along with its year, engagement information, and identifying details. The basic unit of analysis was the original text. Short summaries and listings produced during the research process were used as supporting tools to organize the data, maintain thematic traceability, and facilitate year-by-year comparison. Coding and interpretation, however, were conducted directly on the post content itself. When direct quotation was used, the post URL and/or post ID were recorded, accessibility was checked, and the original wording was verified against the live post and, where necessary, against screenshots or archived records.

This procedure was intended to keep the study’s findings open to external scrutiny and to reduce the risk of inaccurate quotation or attribution. To strengthen traceability, a partial manual verification procedure was also applied through X’s Advanced Search interface. For each year, approximately one-quarter of the retained unique posts were checked using combinations of account-based searches, topic terms, and date ranges. This step functioned as a plausibility and traceability check rather than as a full archival replication of the retrieval process.

Data Preparation

Because the dataset was multilingual, posts were prepared before analysis to preserve their content integrity. Original post texts were retained. At the same time, to ensure conceptual consistency in thematic and discursive comparisons, posts written in languages other than English were rendered into semantically equivalent English. This was not done as a word-for-word translation. Instead, the process followed a principle of semantic equivalence, intended to preserve the post’s central claim, its argumentative structure, and its discursive tone. To strengthen the reliability of this process, a sample of posts in different languages was checked by four language specialists or multilingual researchers, with particular attention to conceptual equivalence, normative tone, and discursive emphasis. Established conceptual expressions, proper names, institution and tool names, and terms carrying strong discursive weight were preserved as much as possible. The output of this process was stored in the dataset as a separate normalized text field. Coding and comparative analyses were conducted on this field, while the original text remained available for cross-checking when needed.

During preparation, duplicate items were removed, off-topic keyword matches were excluded, and year, engagement, and identifying information were systematically recorded. In this way, the dataset was structured to preserve multilingual discursive diversity while remaining suitable for thematic and discursive comparison.

Analysis Procedure

Reliability, Validity, and Researcher Bias

Although this study is based on an interpretive qualitative design and does not claim statistical generalizability, several strategies were used to strengthen the reliability, consistency, and credibility of the analytical process. First, both thematic and discursive coding were carried out iteratively, and the code-theme structure was refined across three rounds. During this process, general and residual categories were separated, unused codes were removed, and theme boundaries were clarified through boundary notes. Separate thematic and discourse codebooks were developed, including code names, operational definitions, inclusion and exclusion criteria, and example expressions. Code assignments were linked to source posts at the post_id level, thereby producing an audit trail between the analytical claims and the original material.

Second, data collection, initial screening, and first-cycle coding were conducted by the researcher, while coding consistency was subsequently checked through inter-rater and intra-rater procedures. A second expert with expertise in artificial intelligence and instructional technologies reviewed the thematic and discourse codebooks, and 25% of the corpus underwent an interrater consistency check. Initial agreement across this subset was 84.6%; after that, coding differences were discussed and resolved by consensus. In addition, to assess coding stability over time, the primary researcher repeated the analysis after a one-month interval. This intra-rater review yielded an agreement level of approximately 96.1%, suggesting a high degree of internal consistency in the application of the coding framework over time. Third, a code-recode procedure was applied to a selected subsample, particularly multilingual and high-engagement posts, to assess category stability and preserve sensitivity to negative cases. Fourth, to reduce meaning shifts in the multilingual corpus, non-English posts were rendered into semantically equivalent English, and selected cases were reviewed for translation accuracy. At the same time, the original-language material remained available for cross-checking.

The illustrative examples reported in the findings tables were drawn from this normalized English field. They should therefore be read as semantically equivalent English renderings rather than literal quotations, unless explicitly identified as direct quotations. For English-language posts, the examples were lightly cleaned only for readability where necessary; for non-English posts, they were rendered to preserve the central claim, argumentative structure, and discursive tone. Accordingly, normalized examples are presented as illustrative excerpts rather than verbatim quotations.

Finally, the study explicitly recognizes that the sample consists of high-engagement posts that achieved comparatively high visibility on X. This is treated as a methodological limitation, since such a sampling choice may introduce biases shaped by platform visibility and the engagement economy. For the same reason, the descriptive summaries used in the findings are interpreted as corpus-internal analytic aids rather than as inferential statistics or as evidence of platform-wide prevalence. The findings should therefore be read as a carefully documented analysis of high-visibility discourse rather than as a representative account of all public discussions of the GenAI divide.

Ethical Considerations

The content analyzed in this study consists of publicly accessible posts on X. Even so, the reporting stage followed several principles: Avoiding the disclosure of personal data, preventing unnecessary identification in direct quotations, and remaining sensitive to the recirculation of deleted content. The use of direct quotations may therefore be limited when necessary to balance research transparency with user privacy, and paraphrasing may be preferred in some cases.

Findings

This section presents the findings in two layers: Thematic analysis and discourse analysis built on the thematic layer. The analysis is based on a purposively assembled corpus of high-engagement posts that gave the GenAI divide high public visibility on X during the 2022–2025 period (2022: n = 32; 2023: n = 100; 2024: n = 98; 2025: n = 102). To improve traceability, the tables include illustrative normalized English excerpts together with the relevant post identifier (Post ID). These excerpts are drawn from the normalized text field and should not be read as verbatim quotations unless explicitly indicated.

RQ1. Around Which Thematic Components was the GenAI Divide on X Structured During 2022–2025?

The thematic analysis identified three main themes that organized the GenAI divide. As shown in Table 1, the relative visibility of these themes changed over time.

Within the retained corpus, the theme of “access constraints and uneven availability” became more visible in 2024 and 2025, whereas “economic access barriers and institutional privilege” and “platform, compute, and market concentration” were more visible in the earlier period. A similar pattern can be seen in “labor market disruption and sectoral shifts,” which also appeared more clearly in the later years. Taken as a descriptive visibility pattern, this suggests that high-visibility GenAI divide discourse in the corpus was articulated not only through economic access and individual capacity, but also through regional availability and broader socioeconomic consequences.

RQ2. How did the Relative Visibility of Discursive Problem Frames Change between 2022 and 2025?

Given the purposive and adaptive retrieval design, the year-by-year comparison should be read as a corpus-internal indication of changing visibility rather than as a direct measure of platform-wide discursive change. Within this retained corpus, post-level multiple-frame coding suggests a marked reorientation in the visibility of certain frames (Table 5). The increased visibility of the geographic divides and tool availability frame in 2024 and 2025 suggests that later high-visibility posts more often centered on which tools were accessible in which regions. Similarly, the growing visibility of the labor-market disruption frame in 2024–2025 suggests that, within the retained corpus, GenAI divide discourse became more strongly linked to outcomes such as employment and professional standing. By contrast, the skills and effective-use gap frame, which was more pronounced in 2023, became less visible in 2024–2025. Taken as a descriptive corpus-level pattern, this suggests a shift toward more macro-level explanations, especially geographic and structural ones.

Here, frames are not treated as additional content themes, but as interpretive lenses through which posts define the GenAI divide as a particular kind of problem.

The similarity pattern should be read as a descriptive aid rather than as a separate quantitative finding. It suggests that the 2022–2023 and 2024–2025 frame distributions were more similar within each period than across periods (Table 6).

These descriptive results suggest that discourse-frame distributions clustered more closely within two temporal phases: 2022–2023 and 2024–2025. In particular, the very high similarity between 2024 and 2025 suggests a relatively stable late-period framing pattern within the corpus. In contrast, the lower similarity between early and late years is consistent with a gradual shift in the structure of high-visibility discourse.

RQ3. Legitimation, Rhetorical Strategies, and Discursive Package Patterns Across Frames

In this subsection, the unit of analysis is not the theme but the discursive frame. Consistent with the refined scope of RQ3, this subsection reports the recurrent legitimating and rhetorical mechanisms that were sufficiently stable across the corpus to support descriptive comparison. Broader discourse features such as actor construction, metaphors and labels, and platform-specific circulation practices informed the interpretive coding process, but they are not treated as separate corpus-level findings because they were more dispersed and less consistently represented across years. The co-occurrence tables presented here, therefore, capture relationships not among content themes but among the discursive frames that are constructed together within posts.

Discussion

What does this corpus tell us about the GenAI divide? The corpus suggests that, in high-visibility discourse on X, the GenAI divide is constructed not as a simple access gap but as something closer to a conversion-centered inequality regime. In this regime, access matters, but it is not enough on its own. What counts is whether access can be turned into effective use, institutional capacity, regional availability, and longer-term socioeconomic benefit. Read in this way, the GenAI divide extends beyond the classic digital divide model of connectivity and skills. It also becomes increasingly tied to the uneven distribution of infrastructural support, platform power, and opportunities for social and economic conversion (DiMaggio & Hargittai, 2001; van Dijk, 2005; Warschauer, 2004).

The three thematic clusters detailed in Tables 2, 3, 4 make this shift visible by showing how the GenAI divide is articulated through access conditions, structural concentration, and downstream social consequences. Access, Capability & Distribution captures not only affordability and user competence, but also the uneven material and geographic conditions under which tools become usable. Political Economy, Data & Compute Power shows that inequality is also publicly framed through ownership, extraction, and concentration. Societal & Educational Impacts, in turn, link these asymmetries to downstream consequences in work, education, and rights. Taken together, these clusters suggest that the GenAI divide is publicly articulated less as a question of possession alone and more as a question of who can convert sociotechnical access into durable advantage under unequal structural conditions. This interpretation is consistent with approaches that understand digital inequality not only through access itself, but also through the institutional and social conditions that shape autonomy, use, and benefit (DiMaggio & Hargittai, 2001; Helsper, 2012).

When we read the corpus-level temporal patterns (RQ2), alongside the discourse analysis (RQ3), this conversion-centered shift becomes visible not only in which themes became prominent, but also in how those themes were legitimized. In the early period, the discourse relied more heavily on certainty, speed, and dramatization. The later period saw a more settled repertoire emerge, centered on geographic availability and labor-market consequences, and a stronger regime of evidence grounded in reports, research, and institutional references. The growing prominence of the geographic divides and labor-market disruption pairing in the later-period corpus suggests that discussion of the GenAI divide is increasingly linked to unequal access to outcomes such as work, income, and professional standing. This pattern is broadly consistent with task-based accounts of technological change, which suggest that automation may intensify displacement unless it is accompanied by new task creation and adaptive institutional conditions (Acemoglu & Restrepo, 2019). At the same time, the continued visibility of political economy and data extraction frames indicates that questions of ownership, infrastructural dependence, and platform concentration remained central structural counterpoints within the high-visibility discourse (Couldry & Mejias, 2019; Zuboff, 2019).

It is also worth noting that the high-visibility discourse was frequently articulated through the language of injustice and inequality. This is likely because high-engagement environments tend to amplify normatively charged claims, and because GenAI is widely linked to concentrated economic and institutional power. In this sense, the GenAI divide appears not merely as a technical difference in use but as a broader social issue with distributive consequences.

Methodologically, combining thematic and discourse analysis makes it possible to examine not only the substantive terms through which the GenAI divide is discussed, but also how those discussions circulate through particular argumentative strategies (Entman, 1993; van Leeuwen, 2007). In interpreting these patterns, though, it is important to keep in mind the difference between corpus-level visibility and conceptual significance. Some themes appeared only marginally in the retained corpus, yet they may still signal important extensions of the GenAI divide beyond the frames that dominated high-visibility discourse.

For education and public policy, the findings suggest that high-visibility discourse on X recognizes not only GenAI’s potential to reduce inequality, but also its capacity to deepen it. Governance, privacy, and rights appeared less frequently than access and labor-centered concerns in this corpus. This aligns with recent higher education and open distance e-learning research showing that GenAI adoption depends not only on awareness or interest, but also on institutional guidance, equitable access, competency development, and governance capacity (Jin et al., 2025; van Wyk et al., 2023). Even so, governance, privacy, and rights concerns remain conceptually important because they mark domains in which GenAI-related inequality is likely to be institutionalized, regulated, or contested beyond the most visible frames of affordability, access, and employment. This suggests that addressing the GenAI divide requires more than individual GenAI literacy initiatives. Access regimes, institutional readiness, data governance, and accountability mechanisms all need attention. A multilayered policy language, one that considers open-source approaches, regulatory frameworks, and capacity-building strategies together, seems more realistic than any single-lever response (Kergroach & Héritier, 2025; UNESCO, 2023). These implications should be read in light of the study’s design as an analysis of high-visibility discourse rather than as a representative account of all platform-wide or societal perceptions of the GenAI divide.

Limitations and Future Research

The findings of this study focus on highly visible discursive patterns concerning the relationship between GenAI and digital inequality on X. This choice offers a clear analytical advantage for tracing narratives, frames, and legitimating strategies that circulated effectively in the public sphere. Even so, the findings should not be read as representing all discussions on X or the broader distribution of public opinion. Platform algorithms and engagement dynamics may have limited the presence of lower-visibility or counter-hegemonic discourses in the dataset. In addition, because engagement was used as a within-year ranking criterion rather than as a standardized cross-year measure, the study does not claim that visibility levels are directly comparable across years. Changes in platform design, recommendation systems, account networks, and content formats may have influenced which posts became highly visible in each period. The sample sizes across years are also not fully balanced. Since the 2022 dataset was limited to 32 posts, the proportions for that year should be interpreted more cautiously than those for later years. The 2022 findings are therefore better understood as exploratory indicators of within-sample visibility in an early phase of the high-visibility discourse, rather than as evidence of settled patterns.

A further limitation is that data collection relied on tool-assisted searches conducted via Grok. Combining keyword-based and semantic search techniques made it easier to capture discursive variation. Still, the dynamic nature of platform data means that identical queries cannot be expected to yield identical results over time. Engagement values may change, and some posts may be deleted or become inaccessible. Moreover, although the multilingual structure of the dataset improved inclusiveness, some linguistic and discursive nuance may have been softened despite the use of semantically equivalent translation procedures. In addition, because retrieval depended on a platform-integrated tool environment in which each query returned only a limited set of results, the retained corpus was shaped not only by topical relevance but also by internal ranking and visibility dynamics.

Language coverage is another limitation worth flagging. Although no platform-level language filter was imposed, the same three-layer retrieval logic was repeated across six languages: English, Japanese, Spanish, Arabic, Turkish, and German. However, this parallel multilingual strategy did not produce balanced outputs, and the final corpus remained dominated by English-language or mixed-language posts. Consequently, posts written entirely in local languages and lacking globally circulating English technical terminology may have remained underrepresented. The dataset should therefore be interpreted as a multilingual but unevenly distributed corpus rather than as a comprehensive cross-linguistic map of GenAI divide discourse on X. In addition, some major linguistic spheres, such as Mandarin Chinese and Hindi, were not directly targeted through dedicated query design. This means that discourse from large language communities may have been only partially captured unless it circulated through English or mixed-language expressions. This is especially relevant in platform ecosystems where access conditions, usage patterns, or cross-platform migration may shape the visibility of public discussion.

Finally, the proportions, co-occurrence patterns, and similarity measures used in the study should not be treated as inferential statistics. They are descriptive devices intended to make the temporal patterning of qualitative findings more visible. For that reason, the study should be read less as a fully reproducible archival scan and more as an interpretive analysis conducted on a social media corpus documented within a particular time frame.

Future research could strengthen the scope and robustness of these findings by using more balanced yearly samples, comparing multiple platforms, and analyzing social media data alongside policy documents, institutional reports, or interview material. Future research could also test whether these thematic and framing patterns hold up over time, identify the events that intensify the compute/colonialism axis, and examine when language, representation, and epistemic justice become more visible within high-visibility discourse. Cross-platform comparisons, for example, across Reddit, YouTube, and LinkedIn, would also help assess how closely these dynamics reflect X’s algorithmic visibility and engagement patterns.