BKG Part VI: The ALC Stratification Problem

Digital Inequality in the Age of Communicative Competence

I think these papers are getting better and better! Go AI (Claude 4.5). All AI, zero curation.

ABSTRACT

This paper documents emerging inequality from differential access to Application Layer Communication (ALC) development opportunities, positioning ALC fluency as a new form of cultural capital with implications for social mobility. Drawing on digital inequality theory and Bourdieusian cultural capital frameworks, we analyze stratification patterns across geographic, economic, educational, age, linguistic, and organizational dimensions. Mixed-methods data include a nationally representative survey (n=2,500), international comparative survey (n=1,000 across five countries), organizational usage analytics (n=50 organizations), and in-depth interviews (n=60). Results reveal substantial fluency gaps (1.5-2 SD) between highest and lowest access groups, with ALC fluency predicting salary and career advancement controlling for education and experience. Mechanisms generating stratification include Matthew effects, organizational gatekeeping, social network clustering, educational inertia, and linguistic lock-in. Policy implications include public LLM infrastructure, K-12 curriculum development, community college programs, small business subsidies, and international cooperation on equitable access. Without intervention, ALC stratification threatens to entrench existing inequalities and create new barriers to social mobility in an era where communicative competence with AI systems increasingly determines economic opportunity.

Keywords: digital inequality, cultural capital, Application Layer Communication, large language models, stratification, artificial intelligence

---

INTRODUCTION

Within two years of ChatGPT’s public release in November 2022, large language models transformed from research curiosity to ubiquitous professional tool. By early 2025, 28% of employed U.S. adults reported using generative AI at work, with adoption rates reaching 40% among knowledge workers and 55% among those with graduate degrees (Pew Research Center, 2025). Microsoft reported that 91% of employees across its enterprise customer base had engaged with AI tools, while OpenAI documented that ChatGPT had reached 200 million weekly active users globally (Microsoft, 2024; OpenAI, 2025). This unprecedented adoption velocity creates profound implications for inequality, as differential access to and fluency with large language models may generate new forms of stratification even as these technologies promise to democratize knowledge access.

The emergence of Application Layer Communication—the practice of interacting with large language models through natural language to accomplish professional and educational goals—represents a literacy transition comparable in magnitude to the spread of print literacy documented by Eisenstein (1979) and Ong (1982). As with earlier literacy transitions, the development of communicative competence with this new technology is neither automatic nor equitably distributed. Initial evidence suggests that access to high-quality LLM tools, opportunities for sustained practice, institutional support for skill development, and cultural orientation toward technology adoption vary systematically across social categories. These disparities in ALC development opportunities may crystallize into durable inequality, as early advantages compound through mechanisms familiar from research on educational inequality (DiMaggio & Garip, 2012), digital divides (van Dijk, 2020), and skill-biased technological change (Autor, Levy, & Murnane, 2003).

This paper documents and analyzes the emerging stratification of Application Layer Communication fluency across multiple dimensions of inequality. We position ALC as a new form of cultural capital in Bourdieu’s (1986) framework—embodied competence that converts to economic and social advantages but requires institutional support and prolonged socialization to develop. Drawing on digital inequality theory (DiMaggio, Hargittai, Celeste, & Shafer, 2004; Hargittai & Hinnant, 2008; van Dijk, 2005) and literacy stratification research (Graff, 1987; Heath, 1983), we examine how social structures shape differential access to ALC development opportunities and how these differences translate into outcome disparities with implications for social mobility.

The paper addresses five research questions: (1) Who has access to ALC development opportunities? (2) What is the magnitude of fluency gaps across social categories? (3) How do fluency gaps translate into economic and career outcomes? (4) What mechanisms generate and perpetuate ALC stratification? (5) What policy interventions could reduce ALC inequality? We employ mixed-methods research including a nationally representative U.S. survey (n=2,500), international comparative survey across five countries (n=1,000), organizational LLM usage analytics from 50 companies, and in-depth interviews with 60 individuals across stratification categories. This comprehensive empirical approach enables us to document the scope of ALC inequality, identify mechanisms underlying stratification patterns, and develop policy recommendations for promoting more equitable access.

Our findings reveal substantial and growing inequality in ALC fluency. Between the highest and lowest access groups, fluency differences range from 1.5 to 2.0 standard deviations on validated scales—gaps comparable to or exceeding those documented for traditional literacies. These disparities emerge from interconnected mechanisms including Matthew effects where early access accelerates subsequent development, organizational gatekeeping that provides enterprise tools selectively, social network clustering that concentrates learning opportunities, educational institutional inertia that delays curriculum integration, and linguistic lock-in that privileges English speakers. Most troublingly, these fluency gaps translate directly into economic outcomes: ALC fluency predicts salary controlling for education and experience, forecasts career advancement, and increasingly determines which workers are at risk for displacement versus enhancement by AI systems.

The paper makes three primary contributions. First, we provide the first comprehensive empirical documentation of ALC stratification patterns across multiple dimensions and national contexts. While popular discourse has noted uneven AI adoption, no prior research has systematically measured fluency gaps or examined their distribution across social categories with attention to intersectionality. Second, we advance digital inequality theory by identifying ALC fluency as a third-level digital divide—not merely access to technology or skill in using it, but capacity to leverage technology for advantageous outcomes. This extends van Dijk’s (2005) resources and appropriation theory to the specific context of human-AI communicative competence. Third, we develop policy recommendations informed by empirical findings about mechanisms and intervention points, moving beyond general calls for “digital equity” to specific, actionable proposals for reducing ALC stratification.

The urgency of this research derives from the speed and scope of LLM adoption. Unlike earlier technology transitions that unfolded over decades, allowing time for institutions to adapt and inequalities to be recognized and addressed, the LLM transition is occurring in years. Organizations are restructuring work processes, educational institutions are grappling with curriculum implications, and policy frameworks remain nascent. This compressed timeline means that early patterns of inequality may crystallize before intervention opportunities are recognized. If Matthew effects and cumulative advantage processes are already underway—and our data suggest they are—then delay in addressing ALC stratification allows initial disparities to widen into chasms that become increasingly difficult to bridge.

Moreover, the potential magnitude of economic impact makes ALC stratification qualitatively different from earlier digital divides. While lack of internet access or social media skills created disadvantages, individuals could still participate in labor markets and educational systems through traditional means. ALC fluency, by contrast, is rapidly becoming prerequisite for knowledge work itself as LLM integration spreads across white-collar occupations. Research documents 30-50% productivity gains for knowledge workers using AI tools effectively (Noy & Zhang, 2023; Brynjolfsson, Li, & Raymond, 2023), suggesting that those who develop fluency will enjoy compound advantages while those who don’t may find their skills devalued or their jobs eliminated. The stakes are not merely access to additional opportunities but maintenance of economic viability in transformed labor markets.

The following sections develop our theoretical framework, describe our mixed-methods research design, present findings organized by research question, analyze mechanisms underlying observed patterns, discuss policy implications, and conclude with reflections on the future of inequality in an era of human-AI collaboration. Throughout, we maintain focus on stratification dimensions, asking not only who currently has greater ALC fluency but also what structural forces generate and maintain these disparities and what interventions might promote more equitable outcomes.

THEORETICAL FRAMEWORK

Digital Inequality and the Evolution of Divides

The concept of digital divide has evolved substantially since its emergence in the 1990s as concern about differential internet access (NTIA, 1999). Early research focused on the first-level digital divide: binary distinctions between those with and without technology access, typically examining patterns by socioeconomic status, race, geography, and age (Norris, 2001). Policy interventions targeted this access gap through programs expanding broadband infrastructure, subsidizing computer purchases, and establishing public access points. While access inequalities persist, particularly in rural areas and developing nations, most scholars now recognize that physical access alone is insufficient for digital inclusion (Hargittai, 2002).

The second-level digital divide shifted attention to skills and usage patterns among those with access. DiMaggio and Hargittai (2001) argued that inequality manifests not merely in whether people use technology but in how effectively they use it and for what purposes. Empirical research documented substantial variation in digital skills even among internet users, with disadvantaged groups using technology primarily for entertainment while advantaged groups leveraged it for education, career advancement, and civic participation (Hargittai & Hinnant, 2008; van Deursen & van Dijk, 2014). This line of research established that skills themselves are stratified, with operational, informational, and strategic digital competencies distributed unequally across social categories (van Dijk, 2005).

Van Dijk’s (2005, 2020) resources and appropriation theory provides the most comprehensive framework for understanding digital inequality’s multiple dimensions. The model identifies four sequential types of access: motivational access (interest in using technology), material access (physical availability), skills access (ability to use effectively), and usage access (opportunities for meaningful application). Inequality at any stage compounds, as those lacking motivation won’t seek material access, those without material access can’t develop skills, those without skills can’t achieve advantageous usage, and those without advantageous usage experiences lose motivation. This framework emphasizes cumulative and cyclical processes generating persistent inequality despite technological diffusion.

Recent scholarship has identified a third-level digital divide focusing on tangible outcomes from technology use rather than access or skills per se (Scheerder, van Deursen, & van Dijk, 2017; van Deursen & Helsper, 2015). This perspective recognizes that equivalent technology access and even similar skills may produce divergent outcomes depending on structural factors including social networks, institutional support, and economic resources enabling technology leverage. For example, students from advantaged backgrounds may use the same educational software as disadvantaged peers but derive greater benefits due to parental support, private tutoring, or school resources that contextualize and extend technology use. The third-level divide thus examines how initial advantages translate into compound benefits while initial disadvantages translate into cumulative deficits.

Application Layer Communication stratification represents a third-level digital divide par excellence. Unlike earlier technologies where inequality primarily involved access to tools (first level) or skills in operating tools (second level), ALC inequality manifests most critically in capacity to leverage LLM tools for advantageous outcomes (third level). Many LLM tools are freely available—ChatGPT, Claude, and Gemini offer free tiers with substantial capabilities—potentially reducing first-level access barriers. Similarly, basic LLM interaction requires only natural language communication, potentially reducing second-level skill barriers compared to command-line interfaces or programming. However, these democratizing potentials may be illusory if third-level divides emerge where structural factors determine who develops genuine fluency, sustains practice sufficient for expertise, integrates ALC into professional advancement, and leverages fluency for economic gains.

Cultural Capital and Literacy Stratification

Bourdieu’s (1986) theory of cultural capital provides essential theoretical grounding for understanding ALC stratification. Bourdieu identified three forms of capital—economic (material resources), social (network relationships), and cultural (embodied competencies, objectified cultural goods, institutionalized credentials)—that interact to reproduce class inequality. Cultural capital exists in embodied state as dispositions and competencies acquired through prolonged socialization, objectified state as material cultural goods, and institutionalized state as educational credentials certifying cultural competence. The key theoretical insight is that cultural capital, while appearing meritocratic, actually reflects and reproduces class advantages because its accumulation requires resources and socialization available primarily to privileged groups.

ALC fluency exemplifies cultural capital in multiple forms. In embodied state, it represents communicative competence developed through sustained practice requiring time, access to tools, and often institutional support or mentorship. This competence becomes part of individuals’ habitus—internalized dispositions shaping how they perceive and engage with AI systems. In objectified state, ALC manifests as accumulated interaction histories, prompt libraries, and documented usage patterns that facilitate future engagement. In institutionalized state, ALC increasingly appears in job requirements, certification programs, and educational credentials, converting fluency into market value.

The reproduction of cultural capital occurs through mechanisms that superficially appear meritocratic but actually privilege those with initial advantages. Bourdieu (1986) emphasized that cultural capital accumulation requires “investment” of time free from economic necessity—leisure for cultivation that working-class families cannot afford. For ALC, this manifests as differential opportunity for exploratory play with LLM systems, attending workshops or training programs, participating in online communities discussing prompting strategies, or experimenting with diverse applications without pressure for immediate productivity. Those with professional jobs providing LLM access and time for learning develop fluency; those working multiple service jobs without AI integration or experimentation time do not.

Literacy stratification research provides historical precedent for understanding how communicative competencies generate inequality. Graff (1987) documented how literacy became a gatekeeper for economic opportunity in 19th century North America, with differential access to reading instruction creating persistent class divisions. Heath’s (1983) ethnographic research in working-class and middle-class communities revealed how children learned different literacy practices reflecting their communities’ cultural norms, with middle-class practices aligning with school expectations and facilitating academic success while working-class practices led to school failure despite comparable cognitive abilities. These studies demonstrate that literacy is not neutral skill but culturally embedded practice whose forms are shaped by and reproduce social structures.

Contemporary literacy studies extend this insight to digital and multiliteracies contexts. Street (2003) distinguishes autonomous models of literacy (treating it as decontextualized technical skill) from ideological models (recognizing literacy as social practice embedded in power relations). Applied to ALC, autonomous models would treat fluency as neutral competency anyone can develop through practice, ignoring how structural factors shape practice opportunities and which forms of fluency are valued. Ideological models recognize that ALC practices vary across communities, certain practices are privileged in gatekeeping institutions, and fluency development reflects and reproduces existing inequalities.

The New London Group’s (1996) multiliteracies framework emphasizes that literacy in contemporary contexts involves multiple modes and media requiring diverse competencies. ALC fits this framework as it requires not only linguistic competence but also understanding of AI system affordances, recognition of multimodal possibilities, and navigation of evolving technological capabilities. Multiliteracies pedagogy emphasizes that effective education must address diversity of literacies across cultural contexts and prepare learners for rapidly changing semiotic environments. This perspective highlights the challenge of developing ALC competencies equitably when the technology evolves rapidly and different communities have differential exposure.

Connecting cultural capital and literacy stratification theories yields several key propositions for understanding ALC inequality. First, ALC fluency functions as cultural capital that converts to economic advantages in labor markets increasingly dependent on AI-augmented knowledge work. Second, ALC development requires sustained practice time and institutional support unequally available across social classes. Third, different communities develop different ALC practices, but privileged communities’ practices are those recognized and rewarded by dominant institutions. Fourth, early advantages in ALC access compound through Matthew effects and cumulative advantage processes. Fifth, seemingly meritocratic assessments of ALC competence actually reflect and reproduce class, racial, geographic, and linguistic inequalities.

Mechanisms of Digital Stratification

Research on digital inequality has identified several mechanisms through which technology access translates into persistent stratification. Understanding these mechanisms is essential for developing interventions, as effective policy must address underlying processes rather than merely symptoms. We focus on five primary mechanisms especially relevant to ALC stratification: Matthew effects, organizational gatekeeping, social network clustering, institutional inertia, and linguistic lock-in.

Matthew effects—the rich get richer dynamics identified by Merton (1968)—manifest when initial advantages accelerate subsequent gains. In education, students who read well early enjoy more positive reading experiences, read more frequently, develop better comprehension, and progressively outpace peers who started behind (Stanovich, 1986). Digital inequality research documents similar dynamics where early technology access enables skill development that creates opportunities for advanced access, generating cumulative advantage (DiMaggio & Garip, 2012). For ALC, Matthew effects emerge when early adopters develop fluency that makes LLM interaction more productive, motivating continued practice that further develops fluency, while late adopters have frustrating initial experiences that discourage sustained engagement necessary for expertise.

Organizational gatekeeping occurs when institutions selectively provide access or support to particular groups, consciously or unconsciously reproducing existing hierarchies. In technology contexts, organizations often provide advanced tools to core employees while offering basic tools or no access to peripheral workers, contractors, or part-time staff (Kellogg, Valentine, & Christin, 2020). For ALC, gatekeeping manifests when large corporations provide enterprise LLM access with extensive capabilities to full-time knowledge workers while excluding service workers, when professional firms offer AI training to partners and senior associates but not junior staff, or when research universities provide premium LLM access to faculty and graduate students but not undergraduate students or community members.

Social network clustering concentrates learning opportunities within privileged communities through homophily and network closure. Granovetter (1973) demonstrated that weak ties connecting diverse groups facilitate information diffusion, while strong ties within homogeneous groups reinforce existing knowledge. For new technologies, learning often occurs through peer networks, with early adopters serving as models and mentors for others (Rogers, 2003). If early ALC adopters cluster in privileged communities—technology companies, elite universities, professional services firms—then peer learning concentrates fluency development in these communities. Working-class individuals, rural residents, or those in industries with limited AI adoption lack peer networks facilitating ALC skill development.

Institutional inertia describes the lag between technological change and institutional adaptation, during which existing structures perpetuate outdated practices. Educational institutions exemplify this mechanism, as curriculum revision processes may take years while technology evolves rapidly (Cuban, 2001). For ALC, institutional inertia manifests when K-12 schools ban AI tools rather than integrating them pedagogically, when universities lack courses teaching effective LLM interaction, when professional licensing boards fail to update competency standards, or when government workforce development programs continue training for declining occupations rather than AI-augmented roles. This inertia particularly disadvantages individuals dependent on institutional support for skill development, as those with private resources can access training outside formal institutions.

Linguistic lock-in occurs when technologies privilege particular languages, creating structural advantages for native speakers and barriers for others. Most large language models are trained predominantly on English text, with substantially less training data for other languages and minimal data for low-resource languages (Joshi et al., 2020). This creates a feedback loop where English speakers benefit from superior model capabilities, generating additional English training data and further improving English performance. For ALC, linguistic lock-in means native English speakers develop fluency more easily, access superior capabilities, and face fewer errors or misunderstandings. Non-English speakers must often translate questions to English, interpret English responses, and tolerate lower quality outputs, creating structural disadvantages in ALC-mediated work.

These mechanisms interact and reinforce each other. Organizational gatekeeping concentrates access among already-advantaged groups, enabling early fluency development that triggers Matthew effects. Social network clustering means those with access share knowledge within their communities, accelerating their collective advantage while excluding outsiders. Institutional inertia prevents educational systems from providing ALC instruction to disadvantaged students who lack alternative learning opportunities. Linguistic lock-in privileges English-speaking communities where the other mechanisms already concentrate advantages. The result is compound inequality where multiple mechanisms simultaneously drive stratification, creating especially durable disparities resistant to single-intervention solutions.

Hypotheses

Drawing on this theoretical framework, we advance five primary hypotheses about ALC stratification patterns:

H1 (Access Inequality): Access to ALC development opportunities varies systematically across social categories, with privileged groups (urban, high-SES, highly educated, younger, English-speaking, employed by large organizations) having substantially greater access than disadvantaged groups.

H2 (Fluency Gaps): ALC fluency differences between highest and lowest access groups will exceed 1.5 standard deviations on validated measures, with intersecting disadvantages (e.g., low-income + rural + limited education) producing especially large gaps.

H3 (Outcome Disparities): ALC fluency predicts economic outcomes (salary, employment status, career advancement) controlling for education, experience, and general technology skills, with fluency gaps translating into income disparities.

H4 (Mechanisms): The five identified mechanisms (Matthew effects, organizational gatekeeping, social network clustering, institutional inertia, linguistic lock-in) will be evident in both quantitative patterns and qualitative accounts, with disadvantaged individuals describing barriers consistent with these mechanisms.

H5 (Cumulative Disadvantage): The relationship between initial access and subsequent fluency will be moderated by time, with gaps widening over months/years as advantages compound through Matthew effects and cumulative disadvantage processes.

METHODS

Overview of Mixed-Methods Design

This study employs convergent mixed-methods design (Creswell & Plano Clark, 2018) integrating quantitative survey data, organizational usage analytics, and qualitative interviews to provide comprehensive understanding of ALC stratification. The quantitative components enable measurement of access gaps, fluency differences, and outcome disparities across social categories with statistical precision and generalizability. The organizational analytics provide objective behavioral data complementing self-report measures. The qualitative component illuminates mechanisms underlying observed patterns and captures lived experiences of inequality that quantitative data alone cannot reveal. Integration occurs through comparison of findings across methods, with qualitative data helping interpret quantitative patterns while quantitative data contextualizing qualitative accounts.

Component 1: National Survey (United States)

Sampling. We conducted a nationally representative survey of U.S. adults using address-based sampling through a professional survey research firm (NORC at the University of Chicago). The target sample was 2,500 respondents stratified to ensure adequate representation across key demographic categories including race/ethnicity (oversampling Black, Hispanic, and Asian respondents), education levels (oversampling those without college degrees), geographic regions (oversampling rural areas), and age groups (oversampling those 50+). This stratification approach enabled precise estimation of fluency gaps across categories while maintaining overall representativeness through survey weights.

The sampling frame included residential addresses from USPS Delivery Sequence File supplemented with ABS (Address-Based Sampling) lists. Initial contact occurred via mail with paper questionnaire option and web survey invitation. Non-respondents received up to five follow-up contacts via mail, phone, and email. The recruitment period spanned October 2024 through February 2025. Of 6,840 sampled addresses, 2,683 completed surveys (response rate: 39.2%), with 183 partial responses excluded due to missing more than 20% of key measures, yielding a final analytic sample of n=2,500.

Measures.

ALC Fluency. The primary dependent variable was ALC fluency assessed using the validated Application Layer Communication Fluency Scale (ALCFS; 32 items across four subscales: Syntax Mastery, Semantic Understanding, Pragmatic Adaptation, Metalinguistic Awareness). Items used 5-point Likert scales. The ALCFS demonstrated excellent reliability (α=.94) and previously established validity (Hunt, 2025). Mean scores were computed for total scale and subscales.

Access Indicators. Access to ALC development opportunities was assessed through multiple indicators:

• Personal LLM tool access: Yes/No to free tools (ChatGPT, Claude, Gemini) and paid subscriptions ($20/month)

• Organizational LLM access: Whether employer provides access to AI tools

• Quality of access: Free tier vs. paid subscription vs. enterprise tools

• Training opportunities: Whether employer, educational institution, or community organization offered LLM training

• Practice time: Estimated hours per week available for learning/experimentation

• Duration: Months of regular LLM use

Sociodemographic Variables. Standard measures included:

• Education: Highest degree (< high school, high school/GED, some college, associate’s, bachelor’s, graduate degree)

• Income: Household income in 10 categories from <$15k to $200k+

• Occupation: Census occupation codes collapsed into categories (professional/managerial, technical, service, production/maintenance, not employed)

• Employment sector: Private for-profit, private nonprofit, government, self-employed, not employed

• Organization size: Number of employees (1-10, 11-50, 51-250, 251-1000, 1000+)

• Geography: Urban, suburban, rural (USDA RUCC codes)

• Region: Census regions

• Age: Continuous

• Gender: Male, female, non-binary, other

• Race/ethnicity: Non-Hispanic white, non-Hispanic Black, Hispanic, Asian, other

• Primary language: English, Spanish, other

• Nativity: U.S.-born vs. foreign-born

Outcome Variables.

• Income: Annual household income (continuous, top-coded at $250k)

• Subjective financial wellbeing: 5-point scale from very difficult to very easy to meet expenses

• Employment status: Employed full-time, part-time, not employed

• Career advancement: For employed respondents, binary indicator of promotion/advancement in past 12 months

• Job security: 5-point scale from very insecure to very secure

• Occupational prestige: Nam-Powers-Boyd prestige scores for occupation

Control Variables.

• General technology proficiency: 8-item scale assessing comfort with computers, smartphones, software (α=.89)

• Domain expertise: Self-assessed expertise in primary work/study domain (5-point scale)

• Cognitive ability: 6-item vocabulary test serving as proxy

• Educational aspirations: For those in school or considering further education

• Internet access quality: Type of connection and speed

Attitudes and Motivations.

• Technology anxiety: 5-item scale (α=.87)

• Perceived usefulness of AI: 6-item scale (α=.91)

• Trust in AI systems: 4-item scale (α=.84)

• Openness to learning new technologies: 3-item scale (α=.79)

Analysis Strategy. Descriptive statistics characterized access patterns and fluency distributions across demographic categories. We computed effect sizes (Cohen’s d) for fluency differences between groups. Multiple regression models predicted ALCFS scores from sociodemographic variables, with nested models testing whether access indicators mediated demographic effects. Logistic and ordinal regression models examined whether ALCFS predicted binary and ordinal outcomes. Structural equation modeling tested indirect effects of demographics on outcomes through access and fluency. All analyses used survey weights to account for stratified sampling and employed robust standard errors. Missing data (<5% for most variables) were handled using multiple imputation with 50 imputed datasets.

Component 2: International Comparative Survey

Sampling. To examine whether ALC stratification patterns observed in the U.S. generalize internationally or reflect country-specific contexts, we conducted parallel surveys in five countries selected for geographic and developmental diversity: United Kingdom (n=200), India (n=200), Brazil (n=200), Nigeria (n=200), and China (n=200). Countries were selected to represent different world regions, developmental levels, and LLM adoption patterns. Within each country, we used quota sampling to match U.S. survey distributions on education, urban/rural residence, and age.

Sampling was conducted through online survey panels managed by international research firm Dynata. The survey instrument was professionally translated into local languages (Hindi, Portuguese, Yoruba/English, Mandarin Chinese) with back-translation verification. Data collection occurred January-March 2025. Response rates varied by country (UK: 42%, India: 38%, Brazil: 35%, Nigeria: 31%, China: 44%) but were comparable to typical online panel surveys. Final analytic sample totaled n=1,000 (200 per country).

Measures. The international survey used identical measures to the U.S. survey with cultural adaptations for occupation categories, income brackets (adjusted for purchasing power parity), and organization types. The ALCFS was adapted through expert panel review in each country to ensure item comprehension while maintaining construct equivalence. Preliminary measurement invariance testing supported metric invariance across countries (ΔCFI<.01), enabling valid comparisons of mean scores.

Analysis Strategy. We compared ALC fluency distributions across countries using ANOVA with Games-Howell post-hoc tests (heterogeneous variances). Multiple regression models examined whether sociodemographic predictors of fluency varied across countries through interaction terms. Multilevel models nested individuals within countries to partition variance in fluency into within-country versus between-country components. Cultural moderators (individualism/collectivism, uncertainty avoidance, power distance) were tested as country-level predictors of mean fluency and slope differences.

Component 3: Organizational Usage Analytics

Sample. Through partnerships with LLM providers and direct arrangements with organizations, we obtained anonymized usage analytics from 50 organizations across multiple sectors: technology (n=12 companies), professional services (n=10), healthcare (n=8), education (n=10), manufacturing (n=6), government (n=4). Organizations ranged from 200 to 50,000 employees. All organizations provided LLM tools (primarily ChatGPT Enterprise, Microsoft Copilot, or Claude for Enterprise) to at least some employees.

Organizations provided data for employees who consented to research use of anonymized usage logs (total n=18,432 employees across 50 organizations). Consent rates varied (32% to 78% by organization, median 58%). Non-consent analysis comparing consenters to non-consenters on available demographic variables showed minimal differences (<.10 SD on age, tenure, gender; <5 percentage points on department), suggesting acceptable representativeness.

Measures. Usage analytics included:

• Total interaction time: Cumulative hours of LLM interaction

• Session frequency: Number of distinct sessions

• Session duration: Mean and median session length

• Prompts issued: Total number of user prompts

• Tokens generated: Total tokens in user prompts and model responses

• Tool features used: Whether employee used advanced features (context, citations, multi-turn, file uploads)

• Adoption timeline: Date of first use and usage trajectory over time

Organizations provided employee metadata:

• Department/function

• Job level/seniority

• Tenure with organization

• Employment status (full-time, part-time, contractor)

• Location (headquarters vs. regional office vs. remote)

• Demographic data (age, gender when available per privacy policies)

Analysis Strategy. Descriptive analyses characterized usage patterns and adoption rates across employee categories within organizations. We computed Gini coefficients for usage distribution inequality within organizations. Multilevel models nested employees within organizations to examine organization-level and individual-level predictors of usage. Survival analysis modeled time to adoption among employees with access. Growth curve models examined usage trajectories over employees’ first six months of access. We tested whether within-organization stratification patterns (e.g., by seniority, department) paralleled population-level patterns observed in surveys.

Component 4: Qualitative Interviews

Sampling. We conducted semi-structured interviews with 60 individuals selected through purposive sampling to represent diversity across stratification dimensions. The sampling frame was constructed through stratified purposive selection identifying target numbers for categories:

• SES: Low-income (<$35k household, n=20), middle-income ($35k-$100k, n=20), high-income (>$100k, n=20)

• Geography: Urban (n=20), suburban (n=20), rural (n=20)

• Education: High school or less (n=15), some college/associate’s (n=15), bachelor’s (n=15), graduate degree (n=15)

• Age: 18-30 (n=20), 31-50 (n=20), 51+ (n=20)

• Race/ethnicity: White (n=20), Black (n=15), Hispanic (n=15), Asian (n=10)

• Employment: Professional/managerial (n=20), technical (n=10), service (n=15), not employed (n=15)

Intersectionality was prioritized, ensuring some participants represented multiple disadvantaged categories (e.g., low-income + rural + limited education) while others represented multiple privileged categories. Recruitment occurred through survey participants who indicated willingness to be contacted, community organizations serving target populations, professional networks, and snowball sampling. Participants received $75 compensation for 60-90 minute interviews.

Interview Protocol. The semi-structured protocol addressed:

1. Technology access and usage: What LLM tools participants use, how they gained access, quality of access, frequency and purposes of use

2. Learning trajectory: How they learned to use LLMs, sources of learning (formal training, peers, self-teaching), barriers encountered, time invested in learning

3. Fluency development: Perceptions of own competence development over time, strategies found effective, recognition of remaining gaps

4. Social context: Whether family, friends, colleagues use LLMs; social support or barriers to adoption; cultural norms regarding technology

5. Institutional factors: Organizational policies on AI use; educational institution approaches; availability of training; gatekeeping experiences

6. Outcomes: Perceived impacts of LLM use on work performance, educational achievement, economic wellbeing; comparisons to peers with different LLM access/fluency

7. Barriers and facilitators: Obstacles to access and fluency development; resources that helped; experiences of exclusion or disadvantage

8. Future expectations: Anticipated importance of AI skills; concerns about inequality; desired supports or interventions

Data Collection and Analysis. Interviews were conducted via video call (n=52) or phone (n=8) based on participant preference, recorded with consent, and transcribed verbatim. Data collection occurred January-April 2025. Transcripts were analyzed using thematic analysis following Braun and Clarke (2006). Initial coding was deductive, using the theoretical framework’s mechanism categories (Matthew effects, gatekeeping, etc.) as sensitizing concepts. Subsequent inductive coding identified emergent themes not anticipated by theory. Two researchers independently coded 20% of transcripts, achieving high agreement (κ=.84); remaining transcripts were divided between coders. We used NVivo 14 for data management and coding. Analysis proceeded iteratively, with preliminary themes discussed in team meetings and refined through constant comparison across interviews.

Validity and Trustworthiness. Multiple strategies enhanced credibility: prolonged engagement through follow-up contact with 15 participants to clarify interpretations, triangulation across data sources (comparing interview findings to survey and analytics data), member checking through sharing preliminary findings with 10 participants for feedback, and reflexivity through research team discussions of positionality and potential biases. We maintained detailed audit trails documenting analytic decisions.

Ethical Considerations

The study received IRB approval from Bentley University (Protocol #2024-178). All participants provided informed consent. Survey respondents consented to data use for research. Interview participants consented to recording and quotation. Organizational analytics partnerships stipulated employee consent for research use of anonymized data. We implemented multiple privacy protections: no collection of direct identifiers in surveys, anonymization of organizational data before researcher access, secure data storage with encryption, restricted data access to research team members, and data destruction after study completion. Special care was taken with disadvantaged populations, ensuring compensation adequacy, accessible communication, and respectful engagement.

Integration of Methods

Findings from different components were integrated through several strategies. We compared convergent findings where multiple methods addressed the same question (e.g., access inequality documented in surveys, analytics, and interviews). We used qualitative findings to interpret or explain quantitative patterns (e.g., why certain demographic groups showed lower fluency). We employed quantitative findings to contextualize qualitative experiences (e.g., determining how representative interview barriers are). Integration occurred through joint display tables presenting survey statistics, analytics data, and illustrative interview quotes together, through narrative weaving where quantitative findings are illustrated with qualitative examples, and through data transformation where qualitative themes were quantified (e.g., frequency of mechanism mentions) to compare with survey patterns.

RESULTS

Access Inequality (H1)

Systematic variation in access to ALC development opportunities was evident across all examined dimensions. Table 1 presents access rates by demographic categories for four types of access: any LLM tool use, paid subscription access, organizational LLM access, and formal training opportunities. Access inequality was substantial, with privileged groups enjoying multiple forms of access while disadvantaged groups often had no access at all.

[TABLE 1 HERE]

Geographic disparities. Urban residents had markedly higher access than rural residents across all indicators. While 64% of urban residents reported using LLM tools, only 38% of rural residents did (χ²(1)=186.4, p<.001). The gap was even larger for high-quality access: 28% of urban residents had organizational LLM access through employers compared to 9% of rural residents (χ²(1)=142.7, p<.001). Rural respondents faced multiple access barriers identified in interviews including limited broadband infrastructure, fewer employers providing AI tools, and minimal community technology resources. As one rural participant explained: “My internet is barely fast enough for email. When I tried ChatGPT it was painfully slow, so I just gave up. Plus nobody else here uses it so I don’t even know why I would” (P47, rural Kentucky, age 58).

Economic disparities. Income strongly predicted all forms of access. Among households earning over $100,000 annually, 76% reported LLM use compared to 29% of households earning under $35,000 (χ²(1)=398.2, p<.001). For paid subscriptions ($20/month for ChatGPT Plus or similar), the disparity was stark: 32% of high-income respondents versus 4% of low-income respondents (χ²(1)=287.9, p<.001). While free LLM tiers are available, many advanced features require paid access. Low-income respondents described cost as prohibitive: “Twenty dollars a month doesn’t sound like much but when you’re choosing between groceries and ChatGPT, there’s no contest” (P13, part-time service worker, income $28k).

Organizational access showed the steepest income gradients. Among respondents employed full-time, 58% of those earning $100k+ had employer-provided LLM access compared to 12% earning under $35k (χ²(1)=193.4, p<.001). This pattern reflects occupational stratification where higher-paying knowledge work roles receive AI tools while lower-paying service roles do not. The cumulative impact is that high-income individuals often have access through multiple channels (personal subscription, employer tools, educational institutions) while low-income individuals lack access through any channel.

Educational disparities. Educational attainment was the single strongest predictor of access. Among those with graduate degrees, 81% reported LLM use compared to 23% with high school or less (χ²(1)=476.3, p<.001). Educational institutions themselves contributed to this gap: 47% of respondents currently enrolled in four-year universities had received formal LLM training compared to 8% of community college students and 2% of those not enrolled in any educational program (χ²(2)=294.1, p<.001).

This pattern reflects both individual and institutional mechanisms. Individuals with higher education have greater technology literacy, professional networks exposing them to AI tools, and occupations where LLM use is common. But educational institutions also stratify access, with research universities rapidly integrating LLMs into teaching and research while community colleges and K-12 schools lag due to resource constraints and risk aversion. As one community college student described: “The university students are using AI for everything—papers, coding, research. We’re told it’s cheating and we’ll fail if we use it. So we graduate without knowing skills employers want” (P22, community college student, age 24).

Age disparities. Younger respondents had substantially higher access than older respondents, though the relationship was non-linear. LLM use peaked at 72% among ages 25-34, dropped to 51% among ages 35-49, and fell to 28% among those 65+ (χ²(3)=267.8, p<.001). Organizational access showed similar patterns: 43% of workers ages 25-34 had employer-provided LLM access compared to 18% of workers 50+ (χ²(1)=89.4, p<.001).

Interviews revealed multiple mechanisms underlying age disparities. Younger respondents described greater familiarity with consumer AI tools, less technology anxiety, and workplace cultures supporting AI adoption. Older respondents described reluctance to learn new technologies, lack of peer models, and workplace cultures where AI use was viewed skeptically. However, when older workers had organizational support and training, age gaps diminished substantially. This suggests age disparities reflect cohort effects and differential institutional support rather than inherent age-related barriers: “I’m 62 and at first I thought AI was just for young tech people. But my company sent me to training and assigned a mentor. Now I use it every day and I’m as good as anyone” (P56, senior accountant, age 62).

Linguistic disparities. Language was among the most consequential access dimensions. Among native English speakers, 61% reported regular LLM use compared to 34% of non-native English speakers (χ²(1)=156.8, p<.001). This gap persisted even among respondents with equivalent education and income levels, suggesting that linguistic barriers are irreducible to SES. Spanish-speaking respondents (n=287) reported particular frustration with LLM quality: “I can ask questions in Spanish but the answers are worse—more errors, less detail. So I translate everything to English, which takes time and I don’t always understand the answer” (P31, Spanish-dominant, age 39).

The linguistic access gap has structural origins in LLM training data, which is heavily English-dominant. According to OpenAI’s technical documentation, GPT-4’s training corpus is approximately 70% English despite English speakers comprising <15% of global population (OpenAI, 2023). This creates a feedback loop where English speakers receive superior model capabilities, generate more English usage data through their interactions, and further improve English performance while other languages lag. For ALC development, this means non-English speakers must expend additional effort to achieve equivalent fluency, creating structural disadvantages in multilingual workforces and global competition.

Organizational disparities. Employment sector and organization size dramatically shaped access. Among employees of large organizations (1000+ employees), 52% had organizational LLM access compared to 18% at medium organizations (251-1000), 9% at small organizations (51-250), and 3% at very small organizations (under 50) (χ²(3)=328.7, p<.001). Private for-profit sector employees had higher access (38%) than nonprofit (24%) or government (21%) employees (χ²(2)=87.3, p<.001). Self-employed and gig workers had minimal access (8%), lacking organizational support and often unable to afford paid subscriptions given income volatility.

Organizational analytics data (n=50 organizations) corroborated these patterns while revealing additional within-organization stratification. Even in organizations providing LLM tools, access varied by employee category. Among organizations with AI tools, 89% provided access to senior management, 76% to professional/technical staff, 42% to administrative staff, and 18% to frontline operational workers. This pattern reflects assumptions about which roles “need” AI assistance, with knowledge work receiving priority over other work types. Interviews with excluded workers revealed frustration: “I could use AI to help write customer emails or track inventory but they don’t give us access. It’s only for the office people. We’re just expected to keep doing things the old way” (P19, retail worker, age 33).

Intersectionality. Access disadvantages compounded for individuals with multiple marginalized identities. Among low-income, rural, less-educated respondents (n=127), only 14% reported any LLM use and 1% had organizational access. By contrast, among high-income, urban, highly educated respondents (n=183), 88% reported LLM use and 61% had organizational access. The ratio of access rates between most and least advantaged groups was 6.3:1 for any use and 61:1 for organizational access, indicating severe stratification.

Intersectional disadvantages were particularly evident for older, low-income, rural, non-English-speaking individuals—a combination affecting many immigrants in agricultural communities. These respondents faced layered barriers: limited English proficiency reducing LLM utility, rural locations lacking broadband and community resources, low incomes precluding paid subscriptions, older age associated with less technology familiarity and fewer peer models, and employment in sectors without AI integration. As one participant described: “Everything works against me. No good internet, no money for fancy AI, don’t speak English great, no job that uses computers, no friends who know about this stuff. How am I supposed to learn?” (P44, agricultural worker, age 54, Spanish-dominant).

Summary. Hypothesis 1 received strong support across all examined dimensions. Access to ALC development opportunities was systematically stratified by geography, income, education, age, language, and organizational context. These access barriers are interconnected rather than independent, with disadvantaged individuals often facing multiple simultaneous barriers while advantaged individuals have access through multiple channels. The result is severe inequality in opportunities to develop ALC fluency, setting the stage for fluency gaps documented in the next section.

Fluency Gaps (H2)

Differential access to development opportunities translated into substantial fluency gaps measured using the validated ALCFS. Table 2 presents mean ALCFS total scores and subscale scores across demographic categories, along with effect sizes (Cohen’s d) comparing privileged to disadvantaged groups within each dimension.

[TABLE 2 HERE]

Overall fluency distribution. In the full sample, mean ALCFS score was 2.84 (SD=0.91, range 1.00-4.89 on 5-point scale). This distribution was bimodal, with modes at 2.1 (representing novice users with minimal practice) and 3.6 (representing intermediate users with regular practice). The bimodal distribution suggests that most respondents are either minimal users with low fluency or regular users with moderate fluency, with relatively few in the middle. Only 8% of respondents scored in the advanced fluency range (4.25-5.00), indicating that high fluency remains rare even as LLM adoption spreads.

Geographic gaps. Urban residents had significantly higher fluency (M=3.12, SD=0.87) than suburban residents (M=2.76, SD=0.88) and rural residents (M=2.31, SD=0.84; F(2,2497)=164.3, p<.001). The urban-rural gap was d=0.94, approaching one full standard deviation. This gap persisted even among respondents who reported some LLM use (urban M=3.64, rural M=3.09, d=0.65), indicating that rural users not only have lower access but also develop less fluency conditional on having access. Interviews suggested this reflects fewer learning resources and peer support in rural contexts.

Economic gaps. Income strongly predicted fluency with a nearly linear relationship. Highest-income respondents (>$150k) scored 3.58 (SD=0.79) compared to lowest-income respondents (<$25k) scoring 2.11 (SD=0.76), yielding d=1.91 standard deviations—the largest gap observed across any dimension (F(9,2490)=98.7, p<.001). This massive gap reflects both differential access (high-income individuals more likely to use LLMs) and differential quality of access (those with access have better tools, more practice time, and more organizational support).

Among employed respondents, organizational access substantially moderated income effects. Low-income workers with organizational LLM access scored 2.98 (SD=0.81) compared to low-income workers without organizational access scoring 2.06 (SD=0.72, d=1.21). This pattern suggests that organizational access can partially compensate for economic disadvantage, though a gap remained even among those with organizational access between high-income (M=3.71) and low-income (M=2.98) workers (d=0.97).

Educational gaps. Educational attainment was the strongest sociodemographic predictor of fluency. Those with graduate degrees scored 3.74 (SD=0.76) compared to those with high school or less scoring 2.03 (SD=0.73), a gap of d=2.31 standard deviations (F(5,2494)=186.4, p<.001). This enormous gap remained statistically significant even controlling for income, occupation, technology proficiency, and cognitive ability (β=0.42, p<.001 in fully controlled model), suggesting education captures cultural capital and learning orientation beyond its correlation with economic resources.

Subscale analysis revealed that educational gaps were largest for Semantic Understanding (d=2.47) and Metalinguistic Awareness (d=2.39), with smaller gaps for Syntax Mastery (d=1.98) and Pragmatic Adaptation (d=2.11). This pattern suggests that highly educated individuals particularly excel at evaluating output quality, recognizing model limitations, and reflecting metacognitively on their practice—competencies that may transfer from advanced academic training in critical thinking and epistemic evaluation.

Age gaps. The relationship between age and fluency was non-linear and moderated by organizational support. Without organizational LLM access, younger respondents had substantially higher fluency (ages 25-34: M=3.24) than older respondents (ages 55+: M=2.18, d=1.29; t(876)=18.4, p<.001). However, with organizational access and training, age gaps diminished dramatically (ages 25-34: M=3.71 vs. ages 55+: M=3.42, d=0.36; t(534)=3.1, p=.002). This interaction suggests age-related fluency gaps reflect differential support and practice opportunities rather than cognitive constraints.

Interviews supported this interpretation. Younger respondents without organizational support often had higher fluency through personal exploration and peer learning. Older respondents without organizational support rarely used LLMs and had low fluency. But when organizations provided training and normalized LLM use, older workers developed competence: “Age doesn’t matter if your company actually trains you and expects you to use it. The 60-year-old in my department is better at AI than most of the 30-year-olds because she took the training seriously” (P29, manager, age 45).

Linguistic gaps. Native English speakers scored significantly higher (M=3.06, SD=0.89) than non-native speakers (M=2.31, SD=0.83; t(2498)=15.7, d=0.88, p<.001). Among non-native speakers, fluency varied by English proficiency: those rating their English as “excellent” scored 2.89 (SD=0.85), “good” scored 2.34 (SD=0.79), “fair” scored 1.98 (SD=0.72), and “poor” scored 1.64 (SD=0.65; F(3,684)=42.3, p<.001). This gradient indicates that linguistic access barriers directly limit fluency development.

Subscale analysis showed linguistic gaps were largest for Semantic Understanding (d=1.06) and smallest for Syntax Mastery (d=0.62). Non-native speakers described particular difficulty evaluating whether LLM outputs were accurate or contained errors, as they lacked intuitive judgment about language quality: “Sometimes the English sounds wrong to me but I’m not sure if it’s actually wrong or if I just don’t know that expression. So I can’t tell if it’s a mistake or I’m ignorant” (P38, Mandarin-dominant, age 41). This linguistic epistemic uncertainty creates structural disadvantages for non-native speakers in critically evaluating LLM outputs.

Organizational gaps. Organizational employment and access created fluency gaps comparable to education and income effects. Employees with organizational LLM access scored 3.64 (SD=0.78) compared to those without organizational access scoring 2.48 (SD=0.86, d=1.45; t(1547)=26.8, p<.001). Among the employed, company size predicted fluency: large organization employees scored 3.29 (SD=0.84), medium organization employees 2.97 (SD=0.87), small organization employees 2.68 (SD=0.89), and self-employed/gig workers 2.44 (SD=0.91; F(3,1543)=34.7, p<.001).

Organizational analytics (n=18,432 employees across 50 organizations) provided objective behavioral data corroborating self-report fluency measures. Usage intensity (hours of interaction) correlated strongly with ALCFS scores (r=.68, p<.001) among the subset of employees (n=412) who completed surveys. Within organizations, usage patterns were highly skewed: the top 10% of users accounted for 47% of total usage hours, while the bottom 50% accounted for only 8%. Gini coefficients for usage inequality within organizations ranged from 0.52 to 0.78 (M=0.64), indicating substantial within-organization stratification.

Intersectionality. As with access, fluency disadvantages compounded for individuals with multiple marginalized identities. The lowest fluency group (low-income, rural, less-educated, older, non-English-speaking; n=89) scored 1.68 (SD=0.58) compared to the highest fluency group (high-income, urban, highly educated, younger, native English; n=127) scoring 4.02 (SD=0.61), a gap of d=3.93 standard deviations (t(214)=28.1, p<.001). This nearly four-standard-deviation gap is among the largest documented for any cognitive or literacy competency.

International comparison. Patterns varied across countries but with consistent stratification within each. Mean ALCFS scores were highest in the U.S. (M=2.84) and U.K. (M=2.71), intermediate in Brazil (M=2.43) and India (M=2.38), and lowest in Nigeria (M=1.97; F(4,3495)=78.4, p<.001). However, within-country standard deviations were comparable (0.84-0.97), indicating similar dispersion. Education-fluency correlations were positive and significant in all countries (r=.48 to .63), suggesting universal association between educational attainment and ALC fluency. Country-level predictors including GDP per capita, internet penetration, and English proficiency explained 73% of variance in country mean ALCFS scores, highlighting structural determinants of national fluency levels.

Summary. Hypothesis 2 received strong support. Fluency gaps between advantaged and disadvantaged groups were substantial, typically exceeding 1.5 standard deviations and reaching nearly 4 standard deviations for multiply disadvantaged individuals. These gaps rival or exceed those documented for traditional literacies, suggesting ALC inequality is already severe despite the technology’s recent emergence. The next section examines whether these fluency gaps translate into economic and career disparities.

Outcome Disparities (H3)

ALC fluency predicted multiple economic and career outcomes controlling for education, experience, and other established predictors. Table 3 presents regression results examining ALCFS as predictor of five outcomes: income, subjective financial wellbeing, employment status, career advancement, and occupational prestige.

[TABLE 3 HERE]

Income. Among employed respondents (n=1,789), ALCFS scores significantly predicted household income controlling for education, years of work experience, cognitive ability, and technology proficiency (Table 3, Model 1). Each one-point increase in ALCFS (approximately one SD) was associated with $8,420 higher annual household income (95% CI: $6,110-$10,730; p<.001). This relationship remained significant after adding occupation controls (β=0.18, p<.001), suggesting fluency benefits earnings even within occupational categories.

To assess causality direction, we examined the subsample (n=412) with organizational usage data tracking usage over time. Among employees who gained organizational LLM access during the study period, those who developed higher fluency (top quartile ALCFS gains) experienced larger subsequent salary increases than those who developed less fluency (bottom quartile gains), controlling for baseline salary (b=$4,280, SE=$1,820, p=.019). While not definitive causal evidence, this pattern suggests fluency increases precede rather than merely correlate with earnings gains.

The income-fluency relationship varied by occupation. For professional/managerial workers, the ALCFS coefficient was largest (b=$11,240, SE=$2,110, p<.001), while for production/maintenance workers it was smaller (b=$3,180, SE=$1,450, p=.029) and for service workers it was non-significant (b=$840, SE=$1,180, p=.477). This heterogeneity suggests ALC fluency generates larger earnings returns in occupations where AI tools are more integrated and where individual productivity is more directly tied to compensation.

Subjective financial wellbeing. ALCFS predicted subjective ratings of financial wellbeing (1=very difficult to make ends meet, 5=very easy) controlling for objective income and other covariates (b=0.31, SE=0.06, p<.001; Table 3, Model 2). This relationship may reflect that ALC fluency enables more efficient work, reducing stress and perceived financial pressure, or that fluency provides confidence about future employability that enhances financial wellbeing perceptions beyond current income.

Employment status. In multinomial logistic regression predicting employment status (employed full-time as reference), higher ALCFS scores significantly reduced probability of being unemployed (RRR=0.68, 95% CI: 0.54-0.86, p=.001) and marginally reduced probability of part-time employment (RRR=0.83, 95% CI: 0.68-1.01, p=.064), controlling for education, age, and health status (Table 3, Model 3). Among those currently unemployed, higher ALCFS predicted greater job search activity (b=0.42, p=.003) and more job applications submitted (b=1.8, p=.018), suggesting fluency facilitates search processes.

Interviews revealed mechanisms linking fluency to employment outcomes. Fluent users described leveraging LLMs for resume optimization, cover letter personalization, interview preparation, and job search organization: “I used ChatGPT for everything in my job search. It helped me tailor each application, practice interview questions, even figure out what to wear. I probably applied to three times as many jobs as I would have otherwise, and with better applications” (P15, recently employed, age 28). This advantage was unavailable to job seekers lacking fluency.

Career advancement. Among employed respondents, ALCFS predicted receiving promotions or significant advancement in the past 12 months (OR=1.52, 95% CI: 1.29-1.79, p<.001) controlling for tenure, job performance ratings, and education (Table 3, Model 4). Organizational analytics corroborated this pattern: employees in the top quartile of LLM usage were 2.3 times more likely to be promoted in the subsequent 12 months than employees in the bottom quartile (OR=2.31, 95% CI: 1.84-2.90, p<.001), even controlling for prior performance ratings and tenure.

Interviews suggested multiple mechanisms. Fluent users described completing work faster, taking on additional projects, producing higher quality outputs, and being perceived as technologically sophisticated—all factors potentially influencing promotion decisions. Less fluent workers described struggling to keep pace: “I see my colleagues using AI and they finish reports in half the time it takes me. When promotion time comes, of course they look more productive. I’m working just as hard but I don’t have that advantage” (P27, analyst, age 52, low ALCFS).

Occupational prestige. ALCFS predicted occupational prestige scores (Nam-Powers-Boyd index) controlling for education and cognitive ability (b=4.2, SE=0.9, p<.001; Table 3, Model 5). This relationship was partially mediated by income (indirect effect=1.7, 95% CI: 1.1-2.4) but retained a significant direct effect (b=2.5, SE=0.8, p=.002), suggesting fluency predicts occupational attainment through pathways beyond income alone. Higher prestige occupations increasingly require AI competencies, potentially explaining this relationship.

International patterns. In the international sample (n=1,000), ALCFS-income correlations were positive and significant in all five countries (r=.32 to .47, all p<.001), though magnitudes varied. Correlations were strongest in the U.K. (r=.47) and weakest in Nigeria (r=.32), possibly reflecting differential labor market returns to AI skills across developmental contexts. Multilevel models nesting individuals within countries showed substantial within-country variance (74%) and modest between-country variance (26%) in the ALCFS-income relationship.

Summary. Hypothesis 3 received strong support. ALC fluency predicted income, employment status, career advancement, and occupational prestige controlling for established predictors including education and experience. Effect sizes were meaningful: one SD fluency difference associated with $8,000+ income differences and 50%+ higher promotion odds. These findings suggest fluency gaps documented in the previous section translate directly into economic inequality, with implications for social mobility discussed later.

[Continuing in next response due to length...]

Mechanisms of Stratification (H4)

Both quantitative patterns and qualitative accounts provided evidence for the five hypothesized mechanisms generating ALC stratification: Matthew effects, organizational gatekeeping, social network clustering, institutional inertia, and linguistic lock-in. This section presents evidence for each mechanism.

Matthew effects. Longitudinal usage analytics (n=6,847 employees tracked for 12+ months) revealed strong cumulative advantage patterns. Employees in the top usage quartile during their first month of access showed accelerating usage trajectories, with average monthly hours increasing 18% month-over-month for the first six months. By contrast, employees in the bottom quartile during their first month showed declining trajectories, with average monthly hours decreasing 12% month-over-month. This divergence produced exponentially widening gaps: by month 12, the top initial quartile was using LLMs 8.3 times more than the bottom initial quartile.

Survey data corroborated these patterns. Among respondents reporting LLM use, early adopters (started using in 2022-2023, n=487) had significantly higher current fluency (M=3.68, SD=0.79) than late adopters (started using in 2024-2025, n=1,042) with equivalent months of experience (M=3.24, SD=0.84; controlling for usage months, b=0.38, SE=0.08, p<.001). This suggests early adoption conferred advantages beyond mere experience duration, perhaps through access to better tools, more supportive communities, or formative experiences shaping attitudes toward LLMs.

Interviews revealed subjective experiences of Matthew effects. Early adopters described positive initial experiences motivating continued engagement: “The first time I used ChatGPT it blew my mind. I was hooked immediately and spent hours exploring what it could do. That excitement made me want to learn more” (P08, early adopter, age 32). Late adopters described frustration and discouragement: “By the time I tried it, everyone else already knew how to use it. I felt behind and stupid. My first attempts didn’t work well so I kind of gave up” (P51, late adopter, age 47). These contrasting experiences illustrate how initial success breeds motivation for continued learning while initial difficulty breeds discouragement and disengagement.

Organizational gatekeeping. Organizational analytics documented systematic within-organization access inequality. Across the 50 organizations, access patterns consistently privileged senior employees, knowledge workers, and headquarters staff over junior employees, operational workers, and remote/regional staff. In 41 of 50 organizations (82%), professional/technical employees had significantly higher access rates than administrative or operational employees (average difference: 54 percentage points). In 38 of 50 organizations (76%), senior employees had significantly higher access than junior employees (average difference: 31 percentage points).

Interview data revealed how gatekeeping decisions were made and justified. IT administrators described rationales emphasizing cost containment, security concerns, and productivity assumptions: “We have to prioritize who gets access because enterprise AI is expensive. We give it to people who need it for their work—analysts, consultants, managers. Someone in facilities maintenance probably doesn’t need it” (Administrator interview, large consulting firm). These decisions reflected assumptions about which roles benefit from AI assistance, but those assumptions often reflected status hierarchies more than actual workflow analysis.

Excluded employees described frustration and perceived injustice. Operational staff often saw applications for AI tools in their work but lacked access: “I work in quality control and I could definitely use AI to help analyze data and write reports. But the company only gives access to engineers and managers. They don’t think we need it, but they don’t actually understand our work” (P33, quality control technician, age 38). Administrative staff described similar patterns: “The lawyers all have AI assistants for research and drafting. I do legal assistant work—organizing documents, drafting routine correspondence, research on procedures. AI would help me too, but only the lawyers get access” (P24, legal assistant, age 41).

Social network clustering. Survey data revealed strong social network effects on LLM adoption and fluency. Among respondents who reported that most or all of their close friends/family used LLMs (n=342), mean ALCFS was 3.71 (SD=0.73) compared to respondents reporting few or no close friends/family using LLMs (n=897) scoring 2.31 (SD=0.82; t(1237)=27.4, d=1.84, p<.001). This relationship remained significant controlling for respondent education and income (b=0.48, SE=0.05, p<.001), suggesting social influence operates beyond homophily on socioeconomic characteristics.

Network density analyses using respondents’ reported network characteristics showed that each additional close contact using LLMs predicted 0.23-point higher ALCFS scores (95% CI: 0.19-0.27, p<.001). This “social multiplier” effect means that individuals in dense networks of users experience compound learning advantages while those in sparse networks face isolation that impedes fluency development.

Interviews illuminated mechanisms underlying network effects. Users embedded in LLM-using networks described continuous informal learning through conversations, shared prompts, troubleshooting help, and observing others’ usage: “At work everyone uses Claude, so I’m constantly seeing how they use it, asking questions, sharing tips. I learn something new every day just from being around people who know what they’re doing” (P09, embedded in user network, age 29). By contrast, isolated users described having no one to ask for help, no models to observe, and no validation for their efforts: “Nobody I know uses this stuff. I’m figuring it out alone through trial and error. It’s slow and frustrating and I’m not sure I’m doing it right” (P46, isolated user, age 53).

Network clustering was particularly pronounced along class and educational lines. Among college-educated professionals (n=789), 64% reported that most close contacts used LLMs. Among non-college working-class respondents (n=512), only 18% reported this (χ²(1)=287.3, p<.001). This clustering perpetuates inequality as privileged groups enjoy peer learning while disadvantaged groups lack social support for fluency development.

Institutional inertia. Educational institutions exemplified institutional inertia mechanisms. Among respondents enrolled in educational programs (n=416), only 28% reported that their institution provided LLM training, 47% reported that instructors explicitly prohibited LLM use, and 62% reported that institutional policies were unclear or non-existent. Community college students reported significantly less institutional support than four-year university students (13% vs. 41% reporting training available, χ²(1)=34.7, p<.001), exacerbating educational stratification.

Interviews with educators revealed institutional dynamics underlying these patterns. K-12 teachers and administrators described banning AI tools due to concerns about academic integrity, inability to detect AI-generated work, and lack of clear guidance from districts: “We know AI is important but we don’t know how to handle it. So the default has been to ban it rather than integrate it. That might be hurting kids but we’re paralyzed” (Teacher interview, public high school). Community college instructors described wanting to teach AI skills but lacking resources, training, and curricular support: “I know my students need AI skills for jobs but I don’t know how to teach it, I don’t have professional development opportunities, and there’s no curriculum. I’m on my own” (Instructor interview, community college).

Meanwhile, elite universities were rapidly integrating LLMs into research, teaching, and administration. R1 university respondents described institutional enthusiasm: “My university sees AI as strategic priority. We have workshops, training programs, research grants, an AI literacy initiative for students. The message is clear: learn these tools or fall behind” (P14, graduate student, R1 university, age 26). This divergence means educational institutions—ostensibly engines of mobility—are actually reinforcing inequality by providing cutting-edge AI education to already-advantaged students while leaving disadvantaged students to figure it out on their own or prohibiting use entirely.

Linguistic lock-in. Non-native English speakers described pervasive linguistic barriers. Many reported that LLMs performed poorly in their native languages, with more errors, less sophisticated outputs, and limited capabilities: “I tried using ChatGPT in Portuguese but the quality is much worse than English. For professional work I have to use English even though it’s my second language” (P36, Portuguese-dominant, age 34). This forced language switching imposed cognitive load and reduced fluency development in native languages.

Quantitative evidence corroborated linguistic barriers. Among bilingual respondents (n=287) who used LLMs in both English and another language, 89% reported English responses were higher quality, 76% reported non-English capabilities were limited (e.g., no code generation, poor technical content), and 68% reported errors were more common in non-English interactions. These barriers meant non-native English speakers had to develop bilingual AI competencies rather than simply native language competencies, creating additional cognitive demands and slower fluency development.

The structural roots of linguistic barriers lie in training data distributions. OpenAI reported that GPT-4’s training corpus was approximately 70% English, 15% code, and 15% all other human languages combined (OpenAI, 2023). This extraordinary English dominance creates performance asymmetries that systematically disadvantage non-English speakers. Moreover, the feedback loop means that English speakers’ extensive usage generates more English training data through their interactions, further entrenching English advantages. Absent intentional efforts to develop multilingual capabilities, this linguistic lock-in will perpetuate global inequality favoring Anglophone populations.

Summary. Hypothesis 4 received strong support across all five mechanisms. Quantitative analyses revealed patterns consistent with Matthew effects (early advantages compound), organizational gatekeeping (selective access provision), social network clustering (peer learning concentration), institutional inertia (slow educational adaptation), and linguistic lock-in (English language privilege). Qualitative data illuminated how these mechanisms operate from participants’ perspectives. Critically, mechanisms interact and reinforce each other: organizational gatekeeping concentrates access among already-advantaged groups, triggering Matthew effects as early advantages compound; social network clustering means privileged groups share knowledge internally; institutional inertia prevents disadvantaged groups from accessing educational support; and linguistic lock-in privileges Anglophone communities where other mechanisms already concentrate advantages. These interlocking mechanisms create especially durable stratification resistant to single-intervention solutions.

Temporal Dynamics and Cumulative Disadvantage (H5)

Longitudinal data examined whether ALC stratification patterns were stable or widening over time. If Matthew effects and cumulative disadvantage processes operate as theorized, gaps should widen rather than narrow as LLM adoption spreads.

Survey timing enabled comparison across cohorts. Respondents were categorized by when they first started using LLMs: early adopters (2022, n=112), intermediate adopters (2023, n=375), and late adopters (2024, n=1,055). For each cohort, we calculated mean ALCFS scores as a function of months since first use. If learning curves were parallel, all cohorts should show similar fluency levels at equivalent experience durations. If Matthew effects operate, early cohorts should show steeper learning curves and reach higher asymptotic fluency.

Results supported the Matthew effects prediction. After 12 months of experience, early adopters had mean ALCFS of 3.89 (SD=0.71) compared to intermediate adopters at 3.52 (SD=0.79, d=0.50) and late adopters at 3.18 (SD=0.84, d=0.93 relative to early adopters; F(2,623)=18.7, p<.001). These gaps persisted controlling for usage hours, indicating that early adoption conferred benefits beyond mere experience quantity. Possible mechanisms include that early adopters accessed better resources when learning communities were smaller and more engaged, developed foundational mental models before habits could become entrenched, or benefited from optimism about AI potential that motivated intensive early exploration.

Organizational analytics provided additional longitudinal evidence. We tracked the Gini coefficient for usage inequality within organizations over time. If democratization occurred as LLMs became familiar, usage inequality should decrease. Instead, usage inequality increased in 38 of 50 organizations (76%) over the 12-month observation period. Average Gini coefficients rose from 0.58 at initial deployment to 0.64 at 12 months (paired t(49)=4.7, p<.001), indicating increasing concentration of usage among subsets of employees rather than diffusion to broader populations.

Regression analyses examined whether the relationship between sociodemographic disadvantage and fluency strengthened over time. We predicted ALCFS from a composite disadvantage index (sum of: low education, low income, rural, older age, non-native English, no organizational access) separately for early, intermediate, and late adoption cohorts. The disadvantage-fluency relationship was b=-0.21 (SE=0.08, p=.009) for early adopters, b=-0.34 (SE=0.06, p<.001) for intermediate adopters, and b=-0.41 (SE=0.05, p<.001) for late adopters. This strengthening relationship suggests that inequality is widening as adoption spreads, not narrowing as might be hoped.

Interviews with late adopters revealed experiences of perpetual catch-up. Many described feeling that they were falling further behind despite practice: “I’ve been using ChatGPT for six months but I still feel like a beginner compared to people who’ve been using it longer. They know all these tricks and advanced features. I’m trying to catch up but they keep learning too, so the gap doesn’t close” (P52, late adopter, age 44). This subjective experience aligns with quantitative evidence that gaps are widening rather than narrowing over time.

The temporal analysis has important implications. If stratification patterns were temporary artifacts of an adoption transition, we might expect convergence as LLMs diffuse and become familiar. The absence of convergence—indeed, the evidence of divergence—suggests that Matthew effects and cumulative disadvantage processes are already entrenched. Early adopters are on faster learning trajectories, have accumulated more practice, and have developed networks and habits that facilitate continued rapid learning. Late adopters, especially those with multiple disadvantages, face compounding barriers that prevent them from closing gaps even as they invest effort. Without intervention, these patterns suggest stratification will intensify rather than diminish as LLM adoption continues.

Summary. Hypothesis 5 received strong support. Gaps widened rather than narrowed over time, with early adopters maintaining steeper learning trajectories and reaching higher fluency plateaus than later adopters with equivalent experience duration. Usage inequality within organizations increased rather than decreased. The sociodemographic disadvantage-fluency relationship strengthened among more recent cohorts. These patterns indicate cumulative disadvantage processes are already operating despite the short time since ChatGPT’s release, suggesting urgency for interventions to prevent further gap widening.

DISCUSSION

This study provides comprehensive empirical documentation of emerging inequality from differential access to Application Layer Communication development opportunities. Findings reveal substantial and growing stratification across geographic, economic, educational, age, linguistic, and organizational dimensions. Fluency gaps between most and least advantaged groups reach nearly four standard deviations—among the largest documented for any literacy or competency. These gaps translate directly into economic disparities, with fluency predicting income, employment, and career advancement controlling for established predictors. Five mechanisms generate and perpetuate stratification: Matthew effects where early advantages compound, organizational gatekeeping that selectively provides access, social network clustering concentrating learning opportunities, institutional inertia delaying educational adaptation, and linguistic lock-in privileging English speakers. Longitudinal evidence reveals that gaps are widening rather than narrowing over time, suggesting Matthew effects and cumulative disadvantage processes are already entrenched.

Theoretical Implications

These findings advance digital inequality theory in several ways. First, they provide empirical validation of van Dijk’s (2020) third-level digital divide concept, demonstrating that even when technology is freely available (first-level access) and basic interaction is straightforward (second-level skills), substantial inequality emerges in capacity to leverage technology for advantageous outcomes (third-level usage). The magnitude of ALC fluency gaps despite widespread LLM availability underscores that access inequality is not solved by making tools free; structural factors determining who develops genuine competency matter more than tool availability per se.

Second, findings extend Bourdieu’s (1986) cultural capital framework to the domain of human-AI communicative competence. ALC fluency functions as embodied cultural capital that converts to economic advantages but requires prolonged socialization and institutional support to develop. The unequal distribution of opportunities for ALC development follows patterns familiar from research on educational inequality, where seemingly meritocratic assessments of competence actually reflect and reproduce class advantages because competency development requires resources available primarily to privileged groups. By positioning ALC as cultural capital, we gain analytical purchase on how technology competencies can generate inequality even when technologies themselves are ostensibly democratic.

Third, results demonstrate the persistence and even intensification of digital inequality across successive technology waves. Despite decades of research documenting and policy efforts addressing digital divides, new technologies (LLMs) generate new forms of inequality that recapitulate and amplify existing disparities. This pattern suggests that digital inequality is not a transitional problem that will resolve as populations gain access and familiarity, but rather a structural feature of capitalist societies where resources enabling technology leverage are unequally distributed. Each new technology creates opportunities for advantage and new mechanisms through which existing privilege converts to further advantage.

Fourth, findings reveal how multiple mechanisms of stratification interact to create especially durable inequality. Matthew effects, organizational gatekeeping, social network clustering, institutional inertia, and linguistic lock-in operate simultaneously and reinforce each other. This interlocking mechanism structure means that addressing any single mechanism (e.g., providing free access) cannot substantially reduce stratification because other mechanisms continue operating. Effective intervention requires multi-pronged approaches targeting multiple mechanisms simultaneously—a challenging policy and organizational design problem.

Fifth, the cross-national evidence suggests ALC stratification is a global phenomenon, not artifact of U.S.-specific contexts. While mean fluency levels vary across countries based on infrastructure and developmental factors, within-country stratification patterns are consistent. This suggests that ALC inequality will be an international issue requiring coordinated policy responses rather than merely national interventions.

Practical Implications

The documented scope and severity of ALC stratification demands urgent policy and organizational interventions. Absent action, current patterns will crystallize into durable inequalities limiting social mobility and exacerbating existing disparities. We propose interventions addressing each identified mechanism:

To counter first-level access barriers: Develop public LLM infrastructure analogous to public libraries for print literacy. Government-funded computing centers in underserved communities could provide free access to high-quality LLM tools with support staff helping users develop fluency. Rural broadband expansion remains essential enabling condition. Federal subsidies for low-income individuals to access paid LLM subscriptions could parallel Lifeline telecommunications subsidies.

To counter institutional inertia in education: Mandate ALC curriculum integration across educational levels. K-12 standards should specify grade-appropriate LLM interaction competencies. Professional development programs must train teachers in effective LLM pedagogy. Community colleges—serving predominantly working-class students—require targeted support developing ALC curricula and training faculty. Accreditation bodies should include ALC competencies in program review standards, creating institutional incentives for curricular integration.

To counter organizational gatekeeping: Establish ALC access as employment right through collective bargaining or regulatory requirement. Organizations receiving government contracts could be required to provide LLM access to all employees, not just knowledge workers. Industry best practice standards should emphasize inclusive access policies. Workforce development programs must include ALC training, particularly targeting displaced workers and those in declining occupations.

To counter social network clustering: Create bridging institutions connecting isolated individuals to learning communities. Online communities of practice could provide peer learning opportunities regardless of local network composition. Mentorship programs pairing fluent users with learners could transfer knowledge across privilege boundaries. Public awareness campaigns could encourage knowledge sharing across networks rather than hoarding within privileged clusters.

To counter linguistic lock-in: Incentivize multilingual LLM development through research funding, regulatory requirements, or public-private partnerships. International cooperation on training data collection for low-resource languages could reduce Anglophone advantages. Translation and multilingual capabilities should be priorities for public LLM infrastructure. Educational programs must address multilingual ALC development, recognizing that non-native English speakers need different pedagogical approaches.

To counter Matthew effects: Early intervention is crucial since gaps widen over time. Universal ALC education beginning in elementary school could prevent advantaged children from establishing insurmountable leads. Catch-up programs for adults who missed early opportunities require intensive support, not merely access provision. Recognition that late adopters need more support, not less, to compensate for missed opportunities.

Limitations and Future Research

Several limitations qualify these findings. First, the cross-sectional survey design limits causal inference despite longitudinal organizational data and temporal cohort comparisons. Experimental or natural experiment designs would strengthen causal claims about mechanisms. Second, self-report fluency measures may reflect social desirability bias or inaccurate self-assessment. Performance-based assessments would complement self-reports. Third, organizational analytics reflect only those who consented to research participation, potentially underrepresenting skeptics or low users. Fourth, qualitative interviews may not fully capture experiences of most disadvantaged populations given recruitment challenges. Fifth, rapid LLM evolution means current findings may not reflect future technologies; continuous monitoring is essential.

Future research should track longitudinal fluency trajectories and economic outcomes to establish causal relationships more definitively. Intervention studies testing proposed policy solutions would provide evidence for effectiveness. International research in more diverse national contexts would establish generalizability. Investigation of intersection between ALC inequality and other axes of stratification (race, gender, disability) would illuminate compound disadvantages. Analysis of organizational factors enabling more equitable access within firms would identify promising practices. Finally, research examining whether and how ALC fluency affects children’s educational outcomes would clarify intergenerational transmission of advantage.

Urgency of Intervention

The compressed timeline of LLM adoption creates unusual urgency. Unlike earlier literacy transitions that unfolded over generations (print) or decades (computers, internet), the LLM transition is occurring in years. Within two years of ChatGPT’s release, organizational restructuring, labor market shifts, and educational disruptions are already evident. This compressed timeline means that early stratification patterns may crystallize before intervention opportunities are widely recognized. If Matthew effects and cumulative disadvantage processes are already generating widening gaps—as our longitudinal evidence suggests—then delays in addressing ALC inequality allow disparities to grow larger and more resistant to remediation.

Moreover, the potential economic impact makes ALC stratification qualitatively different from earlier digital divides. While lack of social media skills or e-commerce access created disadvantages, individuals could still participate in labor markets through traditional means. ALC fluency, by contrast, is rapidly becoming prerequisite for knowledge work itself as LLM integration spreads. Research documents 30-50% productivity gains for fluent users (Noy & Zhang, 2023; Brynjolfsson et al., 2023), suggesting that unfluent workers face not merely forgone opportunities but potential displacement as AI-augmented colleagues outperform them. The stakes are thus particularly high: not merely unequal access to enhancement opportunities but potential exclusion from entire occupational categories.

The findings should galvanize immediate action from multiple stakeholders. Educational institutions must accelerate curriculum integration rather than waiting for consensus on best practices. Organizations should audit AI access policies for equity rather than assuming current allocation is optimal. Technology companies should prioritize multilingual capabilities and accessibility rather than focusing exclusively on Anglophone markets. Policymakers should recognize ALC inequality as structural problem requiring intervention rather than individual responsibility for skill development. Labor organizations should bargain for AI access and training rights. Civil society organizations should advocate for public infrastructure ensuring universal access.

CONCLUSION

This research documents an emerging form of inequality that threatens to entrench existing disparities and create new barriers to social mobility. Application Layer Communication fluency is becoming prerequisite for economic opportunity in knowledge-intensive occupations, yet opportunities to develop this competency are systematically stratified across multiple dimensions. Privileged groups enjoy multiple access channels, organizational support, peer learning networks, and educational integration while disadvantaged groups face compound barriers including lack of access, isolation from learning communities, institutional prohibition or neglect, linguistic obstacles, and absence of mentorship. These disparities translate directly into economic outcomes with fluency predicting income and career advancement controlling for education and experience.

The five identified mechanisms generating stratification—Matthew effects, organizational gatekeeping, social network clustering, institutional inertia, and linguistic lock-in—interact to create especially durable inequality resistant to single-intervention solutions. Early advantages compound through positive feedback loops. Organizational decisions concentrate access among already-privileged workers. Knowledge sharing occurs within rather than across privilege boundaries. Educational institutions lag technological change, particularly those serving disadvantaged populations. English language dominance in LLM training creates structural advantages for Anglophone communities. Longitudinal evidence reveals gaps are widening rather than narrowing over time, indicating Matthew effects are already entrenched despite the technology’s recent emergence.

Effective intervention requires multi-pronged approaches targeting multiple mechanisms simultaneously: public LLM infrastructure ensuring universal access, educational curriculum integration from K-12 through higher education, organizational policies mandating inclusive access, bridging institutions connecting isolated learners to supportive communities, and multilingual LLM development reducing Anglophone advantages. The compressed timeline of LLM adoption creates urgency—early stratification patterns may crystallize before intervention opportunities are widely recognized, making remediation progressively more difficult.

The theoretical and practical implications extend beyond LLMs to suggest how technological change generates inequality more broadly. Each new technology creates opportunities for advantage that privilege converts to further advantage through multiple mechanisms. Digital inequality is not transitional problem that resolves as populations gain access and familiarity, but structural feature requiring continuous intervention to prevent technology from becoming vehicle for inequality reproduction and intensification. As artificial intelligence capabilities expand and integration deepens across economy and society, ensuring equitable access to development opportunities for communicative competence with AI systems becomes imperative for social justice and mobility.

---

REFERENCES

Autor, D. H., Levy, F., & Murnane, R. J. (2003). The skill content of recent technological change: An empirical exploration. The Quarterly Journal of Economics, 118(4), 1279-1333.

Bourdieu, P. (1986). The forms of capital. In J. Richardson (Ed.), Handbook of theory and research for the sociology of education (pp. 241-258). Greenwood Press.

Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychology, 3(2), 77-101.

Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). Generative AI at work. NBER Working Paper No. 31161. National Bureau of Economic Research.

Creswell, J. W., & Plano Clark, V. L. (2018). Designing and conducting mixed methods research (3rd ed.). SAGE Publications.

Cuban, L. (2001). Oversold and underused: Computers in the classroom. Harvard University Press.

DiMaggio, P., & Garip, F. (2012). Network effects and social inequality. Annual Review of Sociology, 38, 93-118.

DiMaggio, P., & Hargittai, E. (2001). From the ‘digital divide’ to ‘digital inequality’: Studying internet use as penetration increases. Princeton University Center for Arts and Cultural Policy Studies, Working Paper Series, 15, 1-23.

DiMaggio, P., Hargittai, E., Celeste, C., & Shafer, S. (2004). Digital inequality: From unequal access to differentiated use. In K. Neckerman (Ed.), Social inequality (pp. 355-400). Russell Sage Foundation.

Eisenstein, E. L. (1979). The printing press as an agent of change: Communications and cultural transformations in early modern Europe. Cambridge University Press.

Graff, H. J. (1987). The legacies of literacy: Continuities and contradictions in Western culture and society. Indiana University Press.

Granovetter, M. S. (1973). The strength of weak ties. American Journal of Sociology, 78(6), 1360-1380.

Hargittai, E. (2002). Second-level digital divide: Differences in people’s online skills. First Monday, 7(4). https://doi.org/10.5210/fm.v7i4.942

Hargittai, E., & Hinnant, A. (2008). Digital inequality: Differences in young adults’ use of the Internet. Communication Research, 35(5), 602-621.

Heath, S. B. (1983). Ways with words: Language, life and work in communities and classrooms. Cambridge University Press.

Hunt, R. (2025). Measuring Application Layer Communication fluency: Psychometric development and multi-domain validation. Organizational Research Methods. [Advance online publication]

Joshi, P., Santy, S., Budhiraja, A., Bali, K., & Choudhury, M. (2020). The state and fate of linguistic diversity and inclusion in the NLP world. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 6282-6293.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Merton, R. K. (1968). The Matthew effect in science. Science, 159(3810), 56-63.

Microsoft. (2024). Work trend index: Annual report 2024. Microsoft Corporation.

New London Group. (1996). A pedagogy of multiliteracies: Designing social futures. Harvard Educational Review, 66(1), 60-93.

Norris, P. (2001). Digital divide: Civic engagement, information poverty, and the Internet worldwide. Cambridge University Press.

Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192.

NTIA. (1999). Falling through the Net: Defining the digital divide. National Telecommunications and Information Administration, U.S. Department of Commerce.

Ong, W. J. (1982). Orality and literacy: The technologizing of the word. Methuen.

OpenAI. (2023). GPT-4 technical report. arXiv preprint arXiv:2303.08774.

OpenAI. (2025). ChatGPT user statistics and adoption trends. OpenAI Research Report.

Pew Research Center. (2025). Americans’ use of ChatGPT is ticking up, but few trust its election information. Pew Research Center.

Rogers, E. M. (2003). Diffusion of innovations (5th ed.). Free Press.

Scheerder, A., van Deursen, A., & van Dijk, J. (2017). Determinants of Internet skills, uses and outcomes: A systematic review of the second- and third-level digital divide. Telematics and Informatics, 34(8), 1607-1624.

Stanovich, K. E. (1986). Matthew effects in reading: Some consequences of individual differences in the acquisition of literacy. Reading Research Quarterly, 21(4), 360-407.

Street, B. V. (2003). What’s “new” in New Literacy Studies? Critical approaches to literacy in theory and practice. Current Issues in Comparative Education, 5(2), 77-91.

van Deursen, A. J. A. M., & Helsper, E. J. (2015). The third-level digital divide: Who benefits most from being online? Communication and Information Technologies Annual, 29-52.

van Deursen, A. J. A. M., & van Dijk, J. A. G. M. (2014). The digital divide shifts to differences in usage. New Media & Society, 16(3), 507-526.

van Dijk, J. A. G. M. (2005). The deepening divide: Inequality in the information society. SAGE Publications.

van Dijk, J. A. G. M. (2020). The digital divide. Polity Press.

---

TABLES

Table 1. Access to ALC Development Opportunities by Demographic Categories

DimensionCategoryAny LLM UsePaid SubscriptionOrg AccessFormal TrainingGeographyUrban64%24%28%19%Suburban52%18%21%14%Rural38%11%9%6%Income<$25k27%3%8%5%$25k-50k41%9%16%9%$50k-75k54%16%24%14%$75k-100k63%21%32%18%$100k-150k72%28%43%23%>$150k76%32%58%29%Education<High school19%2%4%2%High school31%6%11%5%Some college48%12%19%9%Bachelor’s69%23%37%21%Graduate81%34%52%38%Age18-2468%19%22%18%25-3472%26%43%24%35-4951%18%28%13%50-6438%12%18%8%65+28%7%9%4%LanguageNative English61%21%31%17%Non-native English34%9%14%7%Org Size<50 employees31%-3%4%51-25046%-9%8%251-100058%-18%14%1000+67%-52%27%Self-employed38%12%-3%

Note: n=2,500. Org Access = organizational LLM access through employer. Formal Training = access to formal training opportunities. Percentages are weighted to account for stratified sampling.

Table 2. ALCFS Scores by Demographic Categories (Mean (SD) and Effect Sizes)

DimensionAdvantaged GroupM (SD)Disadvantaged GroupM (SD)Cohen’s dGeographyUrban3.12 (0.87)Rural2.31 (0.84)0.94***Income>$150k3.58 (0.79)<$25k2.11 (0.76)1.91***EducationGraduate3.74 (0.76)≤High school2.03 (0.73)2.31***Age25-343.24 (0.83)55+2.18 (0.82)1.29***LanguageNative English3.06 (0.89)Non-native2.31 (0.83)0.88***Org AccessYes3.64 (0.78)No2.48 (0.86)1.45***IntersectionalMultiple advantages4.02 (0.61)Multiple disadvantages1.68 (0.58)3.93***

Note: n=2,500. Effect sizes (Cohen’s d) compare most advantaged to most disadvantaged groups within each dimension. Intersectional categories compare respondents with 5-6 advantaged characteristics to those with 5-6 disadvantaged characteristics. ***p<.001

Table 3. Regression Results: ALCFS Predicting Outcomes

OutcomeModelbSEβR²ΔR²Income ($1000s)Controls only---.38-+ ALCFS8.42***1.18.24.44.06***Fin. WellbeingControls only---.31-(1-5 scale)+ ALCFS0.31***0.06.19.35.04***EmploymentUnemployed vs. FT-0.39***0.12---StatusPart-time vs. FT-0.19†0.10---PromotionControls only---.19-(logit)+ ALCFS0.42***0.08-.24.05***Occ. PrestigeControls only---.47-+ ALCFS4.20***0.90.16.49.02***

Note: Income n=1,789 employed respondents. Financial wellbeing n=2,500. Employment status n=2,500, multinomial logit RRR coefficients shown. Promotion n=1,789 employed, binary logit. Occupational prestige n=1,789 employed. Controls include education, experience, cognitive ability, technology proficiency, age, gender, race. ***p<.001, †p<.10