11 Fault Lines: Inequality and the Future of Quantitative Social Science in the Age of Artificial Intelligence
Alicia Eads, University of Toronto
Fedor Dokshin, University of Toronto
Abstract: TBD
AI usage statement: During the preparation of this paper, the authors used Anthropic’s Claude (Opus 4.7, via <claude.ai>) at several stages of the writing process: to stress-test the paper’s argument and framework, and for editorial feedback on organization, flow, and clarity. The authors undertook all substantive conceptual development and made all argumentative and interpretive decisions. The authors take full responsibility for the content of this publication, including its claims, citations, and conclusions.
11.1 Introduction
The current moment in AI deployment within the quantitative social sciences reflects science fiction writer William Gibson’s observation that “the future is already here — it’s just not very evenly distributed.” Large language models (LLMs) and related tools lower the cost of tasks that once required substantial specialized training or labor, including data cleaning, exploratory analysis, and writing and debugging code (Messing and Tucker 2026). Enthusiastic early adopters report dramatic transformations to their workflows, productivity gains, and high expectations for future research opportunities (Van Noorden and Perkel 2023). One reading of Gibson’s quote is that the future these early adopters are living is just around the corner for everyone else. The promise on offer is not just faster research but a more equal field. If AI can substitute for resources that the best-funded researchers have and others lack, such as research assistants, computing infrastructure, and methodological collaborators, then its broad deployment will narrow the gap between the well-resourced and the rest. This argument is common in discussions of AI deployment in science, but it misses the sociological insight of Gibson’s quote. The uneven distribution of the future is structural, not transitional. Inequality is a permanent state of the technological frontier. The challenge is to understand the dimensions and mechanisms of stratification. Among fields being reshaped by AI, the social sciences are uniquely positioned to theorize this transformation.
Here, we apply a sociological lens, developed through decades of research on earlier technological transformations, to argue that inequality in the age of AI must be understood as multi level. We conceptualize inequality in quantitative social science as operating along three analytically distinct but interrelated levels: (1) access to AI tools and infrastructure, (2) use of those tools in research practice, and (3) returns to that use in the form of recognition, status, and resources (P. DiMaggio and Hargittai 2001; Hargittai 2002; A. J. A. M. van Deursen and Helsper 2015). Dominant framings of AI in science tend to collapse these levels into one, treating inequality as either a temporary gap in access between early adopters and laggards or as a direct function of individual skill and effort. Against this view, we discuss how equalizing access does not equalize use, and equalizing use does not equalize returns. Instead, AI reproduces and reorganizes inequality through familiar sociological mechanisms, including cumulative (dis)advantage (Merton 1968), status‑structured evaluation of AI use (P. DiMaggio and Hargittai 2001; A. J. A. M. van Deursen and Helsper 2015), and the institutional organization of scientific labor (Joyce et al. 2021). The unevenness already visible in the adoption of AI does not simply reflect who encountered new tools first. It reflects how existing institutional positions shape who can adopt AI on favorable terms, whose use is seen as legitimate, and whose outputs are rewarded. The paper develops this argument through three fault lines. The first concerns access, adoption, and evaluation of AI tools, where uneven costs, reputational penalties, and double standards combine to produce what we call the equalizer paradox. The second concerns methodological stratification within the field, as AI simultaneously democratizes execution and pulls the methodological frontier further away. The third concerns cumulative advantage in an era of rising output, where productivity gains translate unequally into the recognition and resources that compound over a career. None of this is determined by the technology itself. Each fault line reflects choices the field is now making, often implicitly, about what counts as scholarship and who gets to do it. We conclude by naming two key choices the field now faces.
11.2 Three Levels of Inequality
In this section, we briefly lay out the three-level framework that structures our diagnosis of AI’s effects on quantitative social science. The framework identifies three levels at which inequality operates: access, use, and returns. It has been developed from sociological accounts of internet adoption that explain how pre-existing social hierarchies get reproduced and restructured through differentiated use of new transformative technologies (P. DiMaggio and Hargittai 2001; H. DiMaggio P. and Shafer 2004; Hargittai 2002; A. J. A. M. van Deursen and Helsper 2015; J. van Dijk 2005; J. A. G. M. van Dijk 2006). First-level inequality concerns access to the technology itself, including who can obtain it, who can afford to maintain it, and whether their access is autonomous and uninterrupted rather than shared, monitored, or constrained. Second-level inequality concerns what users do with the technology once they have access, encompassing both the skills and fluency they bring to it and the patterns of use that skill shapes. Third-level inequality concerns the differential returns and rewards users can translate access and use into, even among users whose access and skill profiles are broadly comparable. These include the economic, social, political, and educational benefits that the same nominal activity produces for differently positioned people. Keeping the levels analytically distinct is essential for understanding why the diffusion of AI does not, on its own, produce more equal outcomes. Even when first-level gaps narrow, second- and third-level inequalities can still reproduce pre-existing class, educational, and geographic stratification (Attewell 2001; A. van Deursen and Dijk 2019). Earlier general-purpose technologies such as the printing press and the personal computer show the same pattern. Surface-level broadening of access has repeatedly coexisted with widening inequality along pre-existing dimensions that the technology amplified rather than flattened (Dittmar 2011; David H. Autor, Katz, and Kearney 2008). The fault lines we discuss below show how all three levels are emerging in the use of AI for quantitative social science.
11.3 Fault Line 1: Equalizer Paradox
In an optimistic view, AI is an equalizer. Broad adoption of AI, in this view, will narrow the gap between those who currently have advantages and those who do not. One advantage is resources for research, which include funding to hire capable research assistants, computing infrastructure, and programming and methodological expertise. AI can substitute for many of these resources or significantly reduce their costs, so that the resource gap between elite and non-elite researchers, senior faculty and graduate students, and the Global North and Global South could decrease (Messing and Tucker 2026).
But there are reasons to be suspicious of this optimistic view. Although a limited-use monthly AI subscription is inexpensive, the kinds of AI use researchers are moving toward, including running AI models in premium tiers and sustaining agentic workflows, would already strain the budgets of many social scientists. The costs are also expected to rise, as AI companies consolidate their business models and exit the early stage of nearly unlimited support from investors (Jin 2025). AI research itself has already concentrated among large technology firms and elite universities (Ahmed and Wahed 2020; Bail 2024; Besiroglu et al. 2024).
These are first-level inequalities, concerning who can access the technology and who can afford it. Public policy initiatives such as the U.S. NAIRR, EuroHPC, and Canada’s SCIP (Parashar et al. 2023) aim to address such first-level disparities by providing shared infrastructure open to a wider community. These initiatives are laudable, but their impact is limited by deeper considerations. The tools themselves perform unevenly across languages, for example, working better in high-resource languages well represented in their training data (such as English or Spanish) than in low-resource languages where training data is scarce (such as Swahili or Tamil), even when the low-resource languages have comparable or larger speaker populations (Wang et al. 2025). A researcher with free NAIRR access is still using a weaker tool if they work in Swahili or Tamil. Equalizing access to a model does not equalize what the model can do.
First-level, access inequalities are only the beginning. As the three-level framework suggests, first-level equalization does not produce second- or third-level equalization. The equalizer argument implicitly assumes that disparities in research output and in career rewards are primarily caused by differences in skill and resources. By closing these gaps, AI adoption will produce more equal and more meritocratic outcomes. But AI-enhanced productivity is also shaped by reputational costs of adoption and by double standards in output evaluation. We consider each in turn. The reputational costs of AI adoption fall unevenly across the field. Today, researchers in many fields face a social penalty simply for using AI (Reif, Larrick, and Soll 2025), but this penalty falls harder on researchers whose competence is already viewed with suspicion (Gai, Hou, and Tu 2025). Such suspicions are well-documented in quantitative and technical domains, where women, scholars from underrepresented racial and ethnic minorities, and others outside the field’s demographic core are presumed less competent even when their records are comparable (Correll and Ridgeway 2006; Moss-Racusin et al. 2012; Eaton et al. 2020; Spencer, Steele, and Quinn 1999). In a field study of software engineers, the competence penalty for AI use was roughly twice as large for women as for men, and older workers were penalized more than younger ones (Gai, Hou, and Tu 2025). The same AI use can be understood as either a productivity tool in the hands of a rational, time-constrained expert or a crutch helping to mask incompetence of a dilettante. Broader cultural beliefs and stereotypes structure the attribution of meaning that others give to observed AI use and, consequently, who decides to use (and disclose their use of) AI. Those who anticipate harsher judgment will be more reluctant to adopt the new technology (Correll and Ridgeway 2006; Gai, Hou, and Tu 2025).
Emerging research on gendered AI adoption is an illustrative case. Gender gaps in academic productivity, as measured by publications and citations, persist (Stockemer, Galassi, and Abou-El-Kheir 2026), which AI adoption could in principle narrow. However, women are about 25% less likely than men to use AI across regions, sectors, and occupations (Otis et al. 2024; Humlum and Vestergaard 2024). Among academic researchers, male productivity on SSRN grew 6.4% more than female productivity after ChatGPT’s release (Tang et al. 2025). Women in these studies report concerns about being judged as “cheating” or as lacking expertise, concerns that men do not report at comparable rates. The same presumption that has long made women’s quantitative competence suspect (Correll 2001, 2004; Moss-Racusin et al. 2012; Reuben, Sapienza, and Zingales 2014), makes their AI use suspect as well, and makes them more reluctant to take advantage of these powerful tools. The result is a second-level inequality, in which equal access to AI does not produce equal use.
If reputational costs shape who adopts AI, evaluation standards shape whose adoption counts. AI could be a force for reducing linguistic and regional inequality in quantitative social science. Non-native English-speaking authors face disadvantages in academic publishing, spending more time on manuscript preparation, facing higher rejection rates, and receiving lower scientific quality ratings on work with identical content but non-standard English prose (Politzer-Ahles, Girolamo, and Ghali 2020; Amano et al. 2023). Authors in the Global East, and non-native English speakers generally, have been adopting AI at similar or higher rates than English-speaking Western authors (Kusumegi et al. 2025; Khan et al. 2025). However, this has not translated into equal academic recognition. Western authors continue to benefit more from AI adoption than their counterparts (Khan et al. 2025). AI can substitute for English fluency, but not for the editorial and peer-review positions that determine whose writing gets the benefit of the doubt. The result is a third-level inequality, in which comparable AI use does not produce comparable returns and rewards.
An equalizer argument implicitly assumes that a disadvantaged group can access the technology on the same terms, adopt it on the same terms, and have its outputs evaluated on the same terms. All three assumptions fail, and they fail for a related reason. The pre-existing hierarchy that the technology was expected to flatten also governs access to the technology, the social conditions of its use, and the evaluation of its outputs.
11.4 Fault Line 2: Methodological Stratification Within the Field
Fault Line 1 concerned inequalities in access, use, and benefit among individual researchers. Another kind of second-level inequality concerns how AI is likely to restructure methodological skill within the field. The likely outcome is a polarization between researchers who can exercise methodological and analytical judgment and those positioned only to execute pre-specified procedures, or, in Lin and Sohail (2026) framing, between researchers who can direct AI tools and those who serve them.
This is not the first time a technology has reshaped a field’s skill structure. New technologies tend to bifurcate previously integrated domains of skilled practice, with a small conceptual layer retaining control over how the technology works and what it can do, while a much larger execution layer operates the technology without needing to understand it (Braverman 1974; Kellogg, Valentine, and Christin 2020). The 19th-century machinist, exercising craft judgment over how to cut a piece of metal, became the 20th-century CNC operator running a program written by someone else. The same pattern has been documented in the printing industry, where skilled typesetters were progressively replaced by keyboard operators running computerized systems designed by a much smaller elite (Wallace and Kalleberg 1982). However, new technologies do not have to result in widespread deskilling. When handheld calculators entered American classrooms, parents and educators feared that students would lose fundamental arithmetic skills and the capacity to evaluate whether an answer was even plausible. Yet there was not a systematic decline in computational ability, as long as calculators were introduced after foundational fluency had been established and were integrated into instruction that continued to value conceptual reasoning. The original worries were only vindicated when the technology was allowed to displace rather than supplement the formative stage (Hembree and Dessart 1986; Ellington 2003).
Subsequent work has generalized the lesson of these historical cases. Technology does not deskill uniformly. It polarizes, hollowing out the middle of the skill distribution while the frontier pulls further away and a broad lower-skill layer expands (D. H. Autor, Levy, and Murnane 2003; Acemoglu and Restrepo 2022). In quantitative social science, the polarization that matters is between researchers who exercise analytical judgment and produce methodological innovation and researchers who run pre-specified procedures without understanding them well enough to exercise judgment or innovate. Quantitative social science has already lived through a milder version of this transition. Successive generations of statistical software packaged complex procedures into menus and commands, thereby black-boxing many assumptions and analytical choices that required explication before. Broad deployment of statistical software shifted statistical pedagogy from interpretive judgment toward technique execution, producing applied users dependent on the software to “do statistics” and foregoing more fundamental knowledge about statistical analysis (Uprichard, Burrows, and Byrne 2008).
AI tools push this pattern further along the same axis. On one hand, generative AI dramatically reduces the cost of methods that once required substantial technical investment. An AI-assisted researcher can now complete tasks like text analysis, complex data cleaning, web scraping, and sophisticated regression analyses in an afternoon, without the specialized training or a methodological collaborator that would have been required in the past. On the other hand, the same forces raise the methodological frontier. If everyone can train a word embedding model, the marginal contribution of doing so is minimal, and the premium flows to researchers who can build, modify, or critically evaluate the underlying models. The likely outcome is the polarization pattern documented in earlier domains, now reproduced inside quantitative social science. The middle of the methodological skill distribution, i.e. the journeyman researcher executing pre-specified methods, is compressed, while frontier methodologists pull away. Democratization at the lower layer and concentration at the frontier are not contradictory. They are two sides of the same structural shift.
What determines which side of the polarization a researcher ends up on is not only the distribution of resources, but the foundational skills acquired before AI became capable of substituting for the activities that build them. Researchers who learned to clean data, write code, or estimate models without AI assistance can recognize when an output is suspicious and can ask the questions that would expose a default that should not have been used or a model that was misspecified. These skills are built precisely through the activities AI can now do for social scientists (Lin and Sohail 2026). Research shows that prior domain knowledge functions as a complement to AI rather than a substitute for it (Vendraminelli et al. 2025), and what looks like AI-driven equalization across users risks masking a growing dependence on AI among those without foundational knowledge and skill (Zhang 2026). The popular assumption that digital or AI “natives” should be better positioned to use these tools well is not supported by the evidence (Bennett, Maton, and Kervin 2008; Kirschner and De Bruyckere 2017).
Relatedly, discussions of AI in academia have rightly identified skill atrophy as an important concern (Messing and Tucker 2026), but the debate too often frames this risk at the level of individual researchers. Rather than established scholars losing core competencies through AI automation, field-level deskilling is more likely to occur through cohort replacement (Ryder 1965). Established researchers, having acquired foundational skills before the advent of AI substitutes, are well positioned to delegate routine tasks to AI while retaining the discernment required to evaluate and contextualize its outputs. New cohorts, by contrast, if trained in environments where AI displaces rather than supports foundational learning, may never acquire the skills necessary to recognize what AI is doing, or what it is doing badly. The historical lesson from calculators applies here. The technology supplements skill when foundational fluency is established before the tool arrives, but substitutes for skill when the technology is adopted first.
The result is a structural shift that the three levels of inequality framework helps us understand. Skill is not simply unevenly distributed across researchers. It is being polarized, with first-level inequalities in resources and access, and methodological foundations, together determining who can operate at the frontier, and third-level returns concentrating among those researchers. This is a second-level inequality, an inequality of use, but a distinctive one. The use itself is being polarized into directing and serving, and which side a researcher lands on is determined by first-level inequality advantages and by foundational training.
11.5 Fault Line 3: Cumulative Advantage in an Era of Rising Output
Fault Lines 1 and 2 identified inequalities in access, adoption, recognition, and methodological position. Fault Line 3 concerns how such inequalities compound over time. Returns to productivity gains do not distribute evenly, because productivity translates into recognition through institutional mechanisms that already sort scholars hierarchically.
Early evidence suggests AI raises research output substantially. Analyzing 2.1 million preprints across arXiv, bioRxiv, and SSRN, Kusumegi et al. (2025) found that researchers who adopted LLMs increased their manuscript output by 23.7 to 89.3 percent depending on field and author background, with the largest gains going to non-native English speakers (see also Noy and Zhang 2023). On its face, this looks like democratization. More output, faster, with disproportionate benefits for those historically disadvantaged by language.
But every technology that raises average productivity also raises the bar against which productivity is evaluated. And here the sociology of science offers a mechanism that helps us anticipate future productivity trajectories. Merton (1968) observed that eminent scientists receive disproportionate credit for work of comparable quality, and that early recognition translates into tangible resources, such as grant funding and publication success, which in turn enable further recognition. Small initial differences amplify into large eventual inequalities (Dannefer 2003; DiPrete and Eirich 2006). Bol, Vaan, and Rijt (2018) showed the magnitude: among Dutch early-career applicants, grant winners just above the funding threshold accumulated more than twice as much research funding over the next eight years as near-identical non-winners just below it.
AI is likely to accelerate these dynamics rather than disrupt them. Researchers bring different resource bases to AI. Ample research funding, research assistants who can build AI-enabled pipelines, institutional access to proprietary data, and collaborators at firms with private models all compound into faster output, which in turn compounds into citations, grants, graduate students, and editorial positions at rates unavailable to less advantaged scholars. The researcher best positioned to exploit a new productivity technology is, as a rule, the researcher who least needed it.
Meanwhile, evaluation norms adjust upward and apply uniformly. Scholars with time constraints (caregiving, heavy teaching loads) and scholars with normative concerns about AI integration are all evaluated against a productivity bar shaped by scholars facing none of these constraints. The bar rises fastest in AI-adopting corners of the field where demographic disparities were already largest, so the evaluation standard inflates most where the evaluated population is least positioned to meet it. In our framework’s terms, this is a third-level inequality: access and use may converge, but the returns to productivity gains accrue unequally.
11.6 The Choices Facing the Field
A pre-existing hierarchy shapes who can use AI, how their use is read, and what their output is worth. The distributional consequences of a new tool are shaped by the institutions surrounding it, not determined by the tool itself (Acemoglu and Restrepo 2022; Joyce et al. 2021; Scheidel 2024).
First, the direction of AI’s effects on inequality depends on choices we are making now. Journals have moved quickly on disclosure in principle, but implementation lags. A comprehensive study found that 70 percent of academic journals have adopted AI disclosure policies, although explicit disclosure rates are at a mere 0.1 percent (He and Bu 2026). No parallel infrastructure has developed for bias-assessment of disclosed AI use. Without it, stronger disclosure enforcement risks amplifying rather than reducing evaluation inequality, by giving reviewers more information to act on without changing the standards by which they act. Beyond journal-level policy, the field faces a choice about evaluation standards. Publication expectations have been rising for decades, and the rise has fallen hardest on scholars with caregiving responsibilities, heavy teaching loads, and limited collaborative resources (Tripodi et al. 2025). AI accelerates this trend. The default is for expectations to rise with tool-aided output, which locks in the cumulative-advantage dynamic described above. Countering this requires evaluation practices that recognize productivity as conditional on the resources available to produce it, not as a property of the individual scholar. This principle underlies existing reform efforts in research assessment, including the San Francisco Declaration on Research Assessment and the narrative-CV formats now used by several major funders (Hicks et al. 2015; Declaration on Research Assessment (DORA) 2012). These decisions, about disclosure infrastructure and evaluation standards, have distributional consequences, and they are being made right now, but too often by default.
Second, the inequalities AI produces are a research object the field is unusually well-equipped to study. We can measure how access to premium AI tools and computing infrastructure is distributed across institutions, regions, and career stages and test whether AI use produces equivalent citation and publication outcomes for different populations. A sociology of AI focused on inequality and structural change already exists (Joyce et al. 2021; Alegria 2023), though its empirical attention has so far been primarily on ML’s effects outside the academy. We must extend that attention inward, to how AI is reshaping quantitative social science itself. Few fields being reshaped by AI have the methodological tools to study their own reshaping. Ours does.