23  Computational Social Scientists Need to Start Caring About the Competitiveness of the AI Market

Patrick Y. Wu
American University

Abstract: The structure and competitiveness of the AI market have the potential to dramatically affect research using large language models (LLMs) in quantitative social science. I argue that market concentration presents a methodological problem for researchers. A less competitive market creates cross-study problems: researchers face a steepening trade-off between replicability and performance, making studies harder to replicate. And as the market consolidates around a smaller set of frontier models, the errors that these models make become correlated across studies, producing findings that appear to converge but actually share a common bias. Concentration also forecloses potential remedies for the convergence of model outputs, undermining the use of multiple LLMs as independent measurements of a latent construct. I make three recommendations: researchers should use architecturally diverse models, researchers should validate these models against held-out human-annotated data, and the discipline should treat the competitiveness of the AI market as an object of study in its own right.

AI usage statement: Claude Opus 4.7 and Opus 4.8 were used to refine writing quality and the flow of individual sentences, and to format BibTeX citations.

23.1 Introduction

In this chapter, I argue that computational social scientists need to start caring about the competitiveness of the AI market. Specifically, I contend that market concentration in AI presents a methodological issue for computational social science.

The AI market has expanded rapidly (Maslej et al. 2025). For example, AI capital expenditures added more to GDP growth than consumer spending in 2025 (Lichtenberg 2025). But tech firms are not entering this new market as equals. Incumbent tech giants, such as Google, Meta, Amazon, Microsoft, and Nvidia, use their existing resources to continue dominating the AI market. Even relatively newer firms are tied to tech giants: OpenAI is backed by Microsoft, while Anthropic is backed by both Amazon and Google. But emerging competitors are putting pressure on these incumbent tech firms. For example, the release of DeepSeek-R1 shook the market. DeepSeek provided performance on par with then-contemporaneous reasoning models; at the same time, its developers claimed to use a fraction of the training resources (Guo et al. 2025). Although the training resources claim is contested (Patel et al. 2025), the market still reacted strongly to the news. In a single day, Nvidia lost nearly $600 billion in market cap.

Naturally, the incumbent tech giants have responded to this threat. They have both lobbied lawmakers and regulators to further entrench their dominant positions in the market (Henshall 2024; Oprysko 2025), and suppressed emergents through either outright acquisitions or quasi-mergers through licensing deals and “acquihires”, or arrangements where an emergent’s top talent are all hired away (Corrigan et al. 2024; Kazimirov 2025). For example, Google signed an agreement with Character.ai that granted Google a non-exclusive license on Character.ai’s LLMs and saw its co-founders hired away to Google. Soon after this deal, Character.ai stopped developing new LLMs (Criddle 2024). In another example, Meta acquired 49% of Scale AI, a full-stack data platform, and hired away much of its top talent, including its co-founder, Alexandr Wang. Worried that Scale would expose their research and technical priorities to Meta, Scale’s largest customers—Google, OpenAI, Microsoft, and xAI—cut ties with Scale almost immediately (Tong et al. 2025). In both cases, these disruptors were neutralized despite not being outright acquired.

Some scholars have argued that, given the high fixed costs associated with training AI models, machine learning and AI constitute a natural monopoly (see, e.g., Narechania (2022)). I have argued elsewhere that this is not necessarily the case—assumptions about data and computational resources have frequently been challenged in the past few years (Wu 2026), such as the development of reinforcement learning with verifiable rewards (Lambert et al. 2025). That said, I do not seek to adjudicate this issue in this chapter. Rather, I argue that computational social scientists must start caring about the competitiveness of the AI market, especially as AI tools and LLMs become more central to social science methodology (Ziems et al. 2024).

I make this argument along two lines. First, a less competitive market creates cross-study problems, or harms to a body of literature rather than to any single study. It leaves fewer models—open and closed (proprietary) alike—for researchers to choose from, and risks concentrating development in a few frontier closed models. Because developers routinely update and deprecate these closed models, researchers face a steepening trade-off between replicability and performance, making studies more difficult to replicate. And as the AI market consolidates around a small set of frontier models, the errors that these models make become correlated across studies; this produces a body of literature whose findings appear to converge but actually share a common source of bias. Second, the remaining models are converging in their outputs and representations, which creates within-study problems with using LLMs. Researchers using multiple LLMs as an approximation of independent measurements of a latent construct may find that the models are not independent at all. While this convergence reflects a structural feature of how frontier models are currently developed rather than a result of market concentration, a concentrated market forecloses possible off-ramps. I close with three recommendations: researchers should use architecturally diverse models, researchers should validate these models against held-out human-annotated data, and the discipline should treat the competitiveness of the AI market as an object of study in its own right.

23.2 Fewer Models, Fewer Options

Computational social scientists increasingly use AI tools such as LLMs for tasks such as classification, free-form coding, content analysis, and the measurement of latent positions and attributes (see, e.g., Ziems et al. (2024); Wu et al. (2023); Le Mens and Gallego (2025); Licht et al. (2025)). But as noted in the Introduction, incumbent tech giants have suppressed emergents either through outright acquisitions or quasi-mergers such as licensing deals and “acquihires” (Kazimirov 2025). A less competitive AI market leaves researchers with fewer models to choose from, leading to two cross-study problems. First, computational social scientists will increasingly face a difficult choice between powerful but closed models with replicability problems and less powerful but open models with improved replicability (Spirling 2023; Barrie et al. 2025; Alizadeh et al. 2025). Market concentration tightens this bind along several dimensions. If improvements to closed models outpace those of open models, this trade-off becomes starker. When only a handful of firms produce frontier models, researchers have little leverage to demand archived snapshots or longer deprecation windows. Closed models are updated silently, retrained on new data, fine-tuned further using new proprietary techniques, and quietly retired (Ollion et al. 2024). And with fewer competitors, vendors can quickly raise API costs, making it unaffordable for researchers to replicate earlier work (Carammia et al. 2024). In short, computational social scientists who choose closed models effectively trade the replicability of their findings for performance. This trade-off becomes steeper as the market concentrates further and can render a wide range of studies irreproducible.

Second, computational social science may face an algorithmic monoculture problem as researchers reach for the same small set of frontier models. Kleinberg and Raghavan (2021) define algorithmic monoculture as the “notion that choices and preferences will become homogeneous in the face of algorithmic curation.” They find that a group of decision-making agents using the same algorithm can reduce the collective quality of decisions even when the shared algorithm is more accurate for any individual decision-maker in isolation. They attribute this to correlated failures across agents that use a shared algorithm. Bommasani et al. (2022) extend this concern from a shared algorithm to the shared data and models on which systems are built. They analyze outcome homogenization, or the tendency for the same individuals or cases to receive the same outcomes across otherwise independent systems. To explain outcome homogenization, they advance the component-sharing hypothesis: as systems are increasingly built on shared training data or shared models, the outcomes they generate become increasingly correlated. Although they test this hypothesis using conventional classifiers and discriminative foundation models rather than modern generative LLMs, they find that data sharing in particular reliably exacerbates outcome homogenization.

The findings of these papers map directly onto computational social science’s use of LLMs: if computational social science increasingly relies on a few models, the errors those models make are correlated across multiple studies, producing a body of literature whose findings may appear to converge but actually share a common source of bias. Studies that look like independent investigations of a phenomenon of interest may instead reflect a correlated blind spot among a handful of frontier models, and the same texts, speakers, or groups may be systematically mismeasured wherever those models are used. A study that attempts to diversify the generative LLMs used only has limited protection against this problem, since every study draws from the same shrinking pool of options. In short, this is the paradox Kleinberg and Raghavan (2021) identify at the scale of the discipline: a model more accurate for any single researcher can still degrade the reliability of the field’s collective findings.

23.3 Converging Responses and Representations

A concentrating AI market can also worsen the use of LLMs within studies, although the problem is subtler. Computational social scientists frequently use multiple LLMs in their research because they typically aim to measure a latent construct. A latent construct is a theoretical and unobservable variable that cannot be measured directly (DeVellis and Thorpe 2022). Content analysis typically requires multiple independent coders because they provide approximately independent, noisy measurements of the underlying construct of interest. But if their labeling approaches are correlated due to shared biases or non-independent labeling, then inference fails (Krippendorff 2019). What the coders jointly measure, in that case, is the shared source of their correlation rather than the latent construct itself.

Researchers have extended this logic to LLMs and frequently use multiple models to label or measure the latent constructs of interest. Typically, we assume that models trained by different firms, on different data, and using different methods, behave as approximately independent noisy measurements of the construct (for an example of an application of this idea, see Verga et al. (2024)). But recent research has shown this assumption is increasingly untenable. The convergence is structural—a feature of how current generative LLMs are developed and trained, not of how many firms compete—but a concentrated market can foreclose alternative ways of doing so.

Frontier models are typically built from a largely shared bundle of components. They share architecture, as nearly all generative LLMs are decoder-only transformers and inherit the same inductive biases (Vaswani et al. 2017; Brown et al. 2020; Minaee et al. 2025). Their training corpora overlap a great deal, drawn from the same digital media data sources, such as the Common Crawl, along with a similar collection of high-quality data including books, code, and reference materials (Villalobos et al. 2024). Post-training follows similar techniques, such as instruction tuning and preference alignment methods (Ouyang et al. 2022). Developers often optimize against the same public benchmarks and leaderboards (Chiang et al. 2024; Phan et al. 2026). And models increasingly share lineage through distillation, as many smaller and newer models are trained on the outputs of frontier ones (Yang et al. 2025). Put another way, models assembled from the same parts tend to behave alike, regardless of which firm assembled them.

This convergence is visible empirically, even on subjective tasks where social scientists often reach for LLMs. Rozado (2024) finds that most major LLMs generate left-of-center responses to political orientation tests. Although these political orientation tests are not designed for LLMs and many contest their substantive interpretation, the consistent leftward signal across 24 models is the relevant finding here. Brown et al. (2026) examine whether LLMs systematically align with particular demographic groups on subjective annotation tasks. They find that demographic bias is not LLM-specific but dataset-specific: across four datasets, the demographic group that LLMs agree with varies by dataset. For example, LLMs sometimes agree more frequently with White annotators on certain datasets and with non-White annotators on others. But they find that, within a given dataset, multiple LLMs exhibit the same direction of bias. Wenger and Kenett (2026) find that LLMs tend to be homogeneously creative: LLM responses tend to mirror other LLM responses far more than humans mirror one another. Taken together, these works exemplify how LLMs, despite independent training, can produce similar outputs even on subjective tasks such as annotation and creativity.

These empirical findings have a representational counterpart. Huh et al. (2024) propose the Platonic Representation Hypothesis: as neural networks scale and train on increasingly similar data, their internal representations converge toward a shared model of the underlying concepts and knowledge that the data describes. Put another way, convergence is not an artifact of the present AI market but a tendency of the current training and development paradigm itself—larger models trained on more of the same data have representations that are more alike. When a researcher treats several LLMs as independent coders, that independence may be illusory. The ensemble measures the models’ shared bias, not the latent construct the researcher aimed to measure.

To be clear, convergence is a feature of the development and training paradigm of current generative LLMs rather than of monopoly. But competition is the mechanism through which alternatives to the dominant paradigm get discovered, funded, and built. While tech giants dominate AI, they seek to preserve their dominance, which naturally induces risk aversion (Lemley and Wansley 2025; Wu 2026). Disruptors and nascent competitors, on the other hand, are more willing to take on the risk, such as using a different architecture, training regime, or corpus assembled on different principles, in order to attract customers and enter new markets (Lemley and Wansley 2025). Although a competitive market does not guarantee that such an alternative paradigm exists, a concentrated market all but guarantees its absence. A concentrating AI market narrows the space of independent training pipelines: the remaining firms continue to train on the same data, license one another’s models, hire away top talent from would-be disruptors and nascent competitors, and benchmark against each other’s outputs. For example, as described previously, when Google licensed Character.ai’s models and absorbed its founders, the company stopped building its own models, and a distinct lineage of models disappeared. In totality, this has significant methodological consequences for computational social science: as independent model development pipelines disappear, so too do architecturally distinct models that a researcher could turn to in order to break the correlation; the more concentrated the market becomes, the harder the problem is to escape.

23.4 What Can We Do?

At present, the AI market’s concentration depends on the layer examined in its technical stack. Compute, for example, is highly concentrated: in 2025, Nvidia controlled 92% of the GPU market (Carbon Credits 2026). On the other hand, the model layer appears to be more competitive. The Arena AI benchmark, for example, shows that both closed and open models are quite competitive with each other at the moment (Chiang et al. 2024). But this level of competition is not guaranteed: incumbent tech giants continue to acquire startups rapidly and aggressively through formal acquisitions and less formal quasi-mergers (Wu 2026).

Three approaches stand out. The first two are methodological, aimed at letting a researcher detect and break the correlated errors that Section 23.2 traced across studies and Section 23.3 traced within studies. The third is structural, calling on the discipline to take on a new policy role to help preserve the market conditions, such as diverse architectures and independent training pipelines, that the first recommendation depends on.

First, researchers should use architecturally diverse models. Vendor diversity alone does not purchase independence, as frontier models from different firms increasingly share corpora, benchmark targets, and post-training techniques. Instead, researchers can turn to models with different architectures, such as encoder-only or encoder-decoder models (Minaee et al. 2025). While the costs and expertise required to use encoder-only or encoder-decoder models are higher than those of calling a frontier model via an API, these architectures have fundamentally different inductive biases. Within a single study, architectural diversity is a more reliable source of independent measurement than an ensemble of frontier decoder-only generative LLMs. Using these additional models alongside generative LLMs may incentivize developers to continue developing them, even as commercial attention focuses on generative LLMs. But the continued existence of architecturally distinct, actively developed, and maintained models for researchers to choose from is, ultimately, a function of the AI market structure.

Second, researchers must continue to validate against held-out human-annotated data. As the quality of LLMs improves, the temptation to skip human annotation increases. Yet human-annotated data matters more than ever, even as it becomes more costly to collect and as annotation is increasingly delegated to LLMs themselves (Westwood 2025). Genuine human-annotated data remains one of the few ways to identify the collective blind spots of frontier models. Researchers using LLMs for measurement should plan from the start to collect a held-out human-annotated subsample under conditions that preclude annotator use of LLMs. They should then report calibration of model labels against that subsample, and, when possible, apply estimators that correct for residual labeling error rather than treating LLM annotations as ground truth (Egami et al. 2023).

The third approach involves directly engaging with antitrust and competition policy research. Computational social science has a stake in the competitiveness of the AI market: the diversity of available models—with different architectural designs, training regimes, and post-training techniques—bears squarely on the methodological soundness of the field, and a concentrated market erodes that diversity. If computational social science methods depend on the AI market, and the AI market is shaped by competition policy, then competition policy is no longer just something that exogenously affects the discipline. Rather, it becomes something that the discipline has both the standing and the obligation to study. While antitrust scholarship has generally been the purview of law and economics, the methodological consequences of market concentration fall within the expertise of computational social scientists.

In other words, the discipline should treat AI antitrust the same way that political scientists, for example, treat redistricting. Political scientists did not only study redistricting. Rather, they built methods such as the efficiency gap (Stephanopoulos and McGhee 2017) and ensemble simulations (Chen and Rodden 2015; Fifield et al. 2020) that courts now use to distinguish ordinary maps from gerrymanders, turning methodological approaches into evidentiary tools. The analogous question for AI antitrust, then, is what contributions computational social scientists may make. Possibilities include metrics that capture convergence of LLMs across subjective annotation tasks, monitoring infrastructures and observatories that track output diversity across deployed models, and technical quality frameworks that could serve as inputs to merger review.