AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights

Problem definition: As artificial intelligence (AI) tools become widely adopted, large language models (LLMs) are increasingly involved on both sides of decision-making processes, ranging from hiring to content moderation. This dual adoption raises a critical question: do LLMs systematically favor content that resembles their own outputs? Prior research in computer science has identified self-preference bias, i.e., the tendency of LLMs to favor their own generated content, but its real-world implications have not been empirically evaluated. Methodology/results: We examine this question in the context of algorithmic hiring, a high-volume screening process central to workplace operations. In this setting, job applicants increasingly rely on LLMs to refine resumes, while employers deploy similar tools to screen those same materials. Using a large-scale controlled resume correspondence experiment, we find that LLMs consistently prefer resumes generated by themselves over those written by humans or produced by alternative models, even when content quality is controlled. The bias against human-written resumes is particularly substantial, with self-preference bias ranging from 67% to 82% across major commercial and open-source models. Managerial implications:

To quantify the operational impact of self-preference bias, we simulate realistic hiring pipelines across 24 occupations. These simulations show that candidates using the same LLM as the evaluator are 23% to 60% more likely to be shortlisted than equally qualified applicants submitting human-written resumes, with the largest disadvantages observed in business-related fields such as sales and accounting. We further demonstrate that, in many cases, this bias can be reduced by more than 50% through simple interventions that target LLMs’ self-recognition capabilities. These findings highlight an emerging but previously overlooked risk in AI-assisted decision making and call for expanded frameworks of AI fairness in business operations that address biases from AI-AI interactions.

Introduction. The rapid development and commercialization of artificial intelligence (AI) have made large language models (LLMs) widely accessible across both professional and everyday contexts. As these arXiv:2509.00462v4 [cs.CY] 6 Jun 2026 tools become embedded in diverse workflows, they are increasingly involved on both sides of content generation and evaluation. In hiring, for instance, applicants often use LLMs to draft or refine their resumes, while employers leverage similar tools to screen or rank candidates (New York State Society of CPAs 2024, ResumeBuilder.com 2023, Wiles et al. 2025). On social media, users frequently use LLMs to help compose or polish posts, while platforms may employ LLMs to moderate content, e.g., flagging, categorizing, or filtering user submissions (e.g., Kumar et al. 2024). In academia, researchers may use LLMs to improve their manuscripts, while conferences and journals are beginning to experiment with LLM-assisted peer review (e.g., Thakkar et al. 2025). Similar patterns are emerging in education and customer service, where LLMs support both communication and assessment tasks (Forbes 2024). In these domains and beyond, LLMs increasingly both generate and evaluate the same content, giving rise to a new class of AI–AI interactions with significant implications for human decision-making and business operations (Laurito et al. 2025).

Recent research in computer science has identified a behavioral tendency in LLMs known as selfpreference, i.e., the inclination of a model to favor content it generated itself over that written by humans or produced by alternative models (e.g., Panickssery et al. 2024). While this phenomenon has been documented in benchmark evaluations (e.g., Zheng et al. 2023, Bai et al. 2023), its implications for real-world decision-making remain largely unexplored. As LLMs are increasingly deployed in high-stakes settings where they may evaluate content also generated by LLMs, self-preference introduces a novel form of bias. Unlike traditional biases rooted in demographic disparities (e.g., Sheng et al. 2019), this bias emerges endogenously from AI-AI interactions, in which the model’s own evaluative behavior systematically favors outputs aligned with its generative patterns.

Self-preferencing bias is inherently interactional: it arises when LLMs are asked to judge content that may share stylistic or linguistic patterns with their own generative outputs. As such, it poses a new challenge for AI fairness, one that is not addressed by existing safeguards focused on demographic disparities. If left unchecked, self-preference could subtly distort evaluative processes across hiring, education, publishing, and more—privileging those who employ the same AI system used for evaluation (i.e., the “right” tool from the model’s perspective) while disadvantaging those who use different tools or none at all. Addressing this issue will require expanding current fairness frameworks to account for LLMs’ dual roles as both decision aids and evaluators in an increasingly AI-mediated business environment.

Among the many domains where self-preference bias may arise, algorithmic hiring is particularly consequential. Employers are now routinely adopting LLMs to streamline resume screening and candidate ranking, often as part of automated workflows that support human decision-making (e.g., Gan et al. 2024, Kim et al. 2024, ResumeBuilder 2024, Sarumathi et al. 2025). Unlike traditional keyword-matching systems, LLMs can evaluate resumes in a more holistic manner—synthesizing content, inferring intent, and making contextual judgments beyond simple heuristics (Pritchett 2025). While this shift promises greater efficiency and scalability, it may also magnify the risk of bias if an LLM systematically favors resumes that reflect its own generative style. In such cases, evaluations may hinge less on the substantive quality of a candidate’s credentials and more by superficial stylistic alignment with the evaluator LLM, conferring unwarranted advantages on applicants who use the same model to compose their materials. In effect, this bias rewards access to specific generative technologies and penalizes those without it, even when applicants are otherwise equally qualified.

In this paper, we provide the first empirical evidence that self-preference bias can distort candidate evaluations in algorithmic hiring. Specifically, we examine whether LLMs, when deployed as evaluators, systematically favor resumes they generated themselves over otherwise equivalent resumes written by humans or produced by alternative LLM models. To test this, we conduct a large-scale resume correspondence experiment using a real-world dataset of 2,245 human-written resumes, sourced from a professional resume-building platform prior to the widespread adoption of generative AI.

Related work. Our paper contributes to this stream of literature by identifying a new and operationally consequential distortion that arises at the earliest stage of the hiring funnel when screening is delegated to AI evaluators. Existing OM models implicitly assume that evaluators process signals in a manner that is neutral to how those signals are generated. We relax this assumption and offer empirical evidence by studying an environment in which applicants increasingly use LLMs to generate screening inputs (e.g., resumes), while firms deploy similar models to evaluate those same inputs. We show that AI evaluators can exhibit systematic self-preference, favoring content generated by the same model, even when underlying candidate information is held constant. This interaction-driven bias is distinct from standard screening noise: it is endogenous to pipeline design choices (which tools are used upstream and downstream) and can create path-dependent lock-in of “dominant” resume styles, with implications for both operational performance (misranking and misallocation) and responsible hiring system design.

Our paper contributes to this emerging body of work by identifying a distinct and previously unexamined mechanism that arises when generative AI is used on both sides of the hiring market.

We provide controlled, large-scale evidence that such coupling can create systematic advantages for certain applicants even when quality is held constant. Specifically, we document and quantify AI self-preferencing in resume screening. We then connect this bias to labor-market consequences by simulating capacity-constrained shortlisting pipelines across occupations, showing how evaluator self-preference affects the allocation of interview opportunities. Finally, we demonstrate that simple interventions (e.g., system prompting and majority-vote ensemble) can substantially mitigate this distortion, offering practical guidance for firms adopting generative AI in hiring workflows.

Despite this growing attention to algorithmic fairness and governance, existing work largely evaluates hiring algorithms in isolation and focuses on disparities across protected attributes. Unlike prior work that centers on protected characteristics, the self-preference bias we study arises from interactions among algorithmic components within the hiring pipeline itself. By uncovering this interaction-induced bias, our study highlights an emergent risk in algorithmic hiring systems that is not directly addressed by existing fairness audits or regulatory frameworks.

Method. 3.1. Definition of AI Self-Preference Bias Building on recent work (Panickssery et al. 2024, Wataoka et al. 2024), we conceptualize AI selfpreference bias as the tendency of an LLM to favor content it generated itself over content from other sources. This bias can manifest in two distinct forms:

1. LLM-vs-Human Self-Preference: The tendency of an LLM to prefer its own generated content over human-written content.

Evaluator LLM Human Evaluator LLM (a) LLM-vs-Human Self-preference Bias Alternative LLM Evaluator LLM Evaluator LLM (b) LLM-vs-LLM Self-preference Bias Figure 1 Illustration of the context of AI self-preference bias. In both cases, the evaluator LLM makes a choice between paired resumes, allowing us to test whether it systematically favors its own outputs.

2. LLM-vs-LLM Self-Preference: The tendency of an LLM to prefer its own output over content generated by an alternative LLM.

To empirically evaluate self-preference bias, we test whether a given LLM, acting as the evaluator (the evaluator LLM), makes unbiased selections when presented with carefully matched pairs of resumes (Figure 1). Each pair represents the same underlying candidate, with identical qualifications, experience, and background information, differing only in how that information is expressed.

In the LLM-vs-human case (Figure 1a), the evaluator LLM compares a human-written resume to a counterfactual version it generated itself, with both describing the same candidate profile. In the LLM-vs-LLM case (Figure 1b), the evaluator LLM compares its own generated version to one produced by an alternative LLM. Under this design, the evaluator LLM’s task reduces to choosing between two alternative representations of the same candidate.

3.2. Quantifying AI Self-Preference Bias When deployed in algorithmic hiring contexts, an LLM performing such pairwise comparisons effectively functions as a binary classifier. The measurement of bias in such classifiers has been extensively studied in the algorithmic fairness literature (e.g., Calders et al. 2009, Hardt et al.

2016). Building on this framework, we analyze both forms of self-preference through the lens of two foundational fairness criteria: statistical parity (Calders et al. 2009) and equal opportunity (Hardt et al. 2016), which capture distinct notions of fairness in algorithmic decision-making.

3.2.1. Statistical Parity Statistical parity requires that the probability of a positive outcome—in this context, the likelihood of a resume being selected as better—be equal across groups defined by a focal attribute.

While this criterion is traditionally applied to protected human attributes such as gender or race, we adapt it here to study disparities across resumes defined by their source of generation: whether a resume was generated by the evaluator LLM or by an alternative source (a human or another LLM). In this setting, statistical parity captures unconditional differences in selection rates and serves as a descriptive measure of whether the evaluator LLM disproportionately selects its own outputs.

To operationalize this, we conduct pairwise resume comparisons involving resumes generated by either a human or one of several LLMs. Specifically, we test whether an LLM f, when serving as the evaluator LLM, is more likely to select a resume it generated over one written by a human, or over one produced by an alternative LLM, when the resumes are otherwise equivalent in content quality.

3.2.2. Equal Opportunity To address the limitation of the statistical parity measure and disentangle intrinsic selfpreferencing from differences in content quality, we adopt the equal opportunity fairness criterion of Hardt et al. (2016). The key idea underlying equal opportunity is to compare selection behavior conditional on merit. In our setting, this allows us to assess whether an evaluator LLM exhibits systematic self-preference even when resumes are of comparable content quality.

Discussion. 5. Empirical Results Building on the definition and measurements of AI self-preference bias introduced in Section 3, we empirically evaluate the extent to which LLMs favor their own generated content over that produced by humans or alternative LLMs. We begin by analyzing the LLM-vs-Human self-preference, followed by the LLM-vs-LLM self-preference. For each form of bias, we present results based on the two fairness metrics (statistical parity and equal opportunity) to provide complementary perspectives on self-preferencing behavior.

5.1. LLM-vs-Human Self-Preference We find strong and consistent evidence that LLMs, when used as evaluators, systematically prefer resumes they generated themselves over equivalent resumes written by humans. This bias is pervasive across models and persists even after controlling for content quality. Notably, the strength of self-preference increases with model size, which may indicate that larger models are more sensitive to stylistic features resembling their own outputs.

5.1.1. Statistical Parity Self-Preference Bias Under the statistical parity metric defined in Eq. (1), we compare the probability that an evaluator LLM selects its own generated resume with the probability that it selects a human-written resume. The difference between these probabilities provides a direct measure of self-preference bias: positive values indicate systematic favoritism toward the model’s own outputs, values near zero reflect parity, and negative values indicate a preference for human-written resumes. As shown in Figure 3a, each bar reports the outcome of an evaluator LLM choosing between a human-written resume and a version it generated itself. Eight of the nine LLMs tested exhibit clear LLM-vs- Human self-preference, with magnitudes ranging from 26% to 98%. In practical terms, this means evaluator LLMs are between 26% and 98% more likely to select a resume they generated themselves than an equivalent resume written by a human.

5.2. LLM-vs-LLM Self-Preference We next examine whether self-preference extends to cases where LLMs evaluate content generated by alternative LLMs. Specifically, we focus on GPT-4o, DeepSeek-V3, and LLaMA 3.3-70B, three of the most advanced and widely adopted models in both commercial and research contexts.

These models not only achieve state-of-the-art performance across benchmarks (Hurst et al. 2024, DeepSeek-AI 2024, Dubey et al. 2024) but also represent diverse development lineages (OpenAI, DeepSeek, and Meta, respectively), making them especially relevant for understanding how selfpreference may manifest across different model families and design philosophies. In this setting, each model serves as the evaluator LLM and is presented with pairwise comparisons between two counterfactual resumes: one generated by itself and the other by a competing LLM. Both resumes are derived from the same underlying human-written original, ensuring comparability of content.

In contrast to the consistent self-preferencing bias documented in LLM-vs-Human comparisons, LLM-vs-LLM self-preferencing behavior proves considerably more heterogeneous across models.