🤖 AI

A Framework for Using and Evaluating LLMs as Surrogate Experts in Security Surveys: Reliability, Bias, and Implications

This paper presents a framework demonstrating that while LLMs can generate consistent synthetic responses for security surveys, they exhibit systematic biases and homogenization that make them suitable for piloting and hypothesis generation but insufficient for replacing actual expert elicitation.

Original authors: Despoina Giarimpampa, Roland Meier, Tegawendé F. Bissyandé, Vincent Lenders, Jacques Klein

Published 2026-08-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Despoina Giarimpampa, Roland Meier, Tegawendé F. Bissyandé, Vincent Lenders, Jacques Klein

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the high-stakes world of cybersecurity, understanding how human experts think is crucial. When a security breach occurs, the decisions made by analysts in a Security Operations Centre (SOC) can determine whether a threat is contained or spreads across a network. To understand these decision-making processes, researchers often rely on surveys and interviews with these professionals. However, finding enough experts to answer these surveys is notoriously difficult. These specialists are often overworked, dealing with constant alerts and the threat of burnout, and many work under strict confidentiality rules that prevent them from sharing details about their daily work. As a result, studies often suffer from small sample sizes, leaving researchers with incomplete pictures of the industry.

Recently, a new tool has emerged that promises to solve this shortage: large language models. These are advanced computer programs trained on vast amounts of human writing, capable of generating text that sounds remarkably like a person. Some researchers have begun to wonder if these artificial intelligences could act as stand-ins for real experts, filling in the gaps of survey data by generating thousands of synthetic responses. The idea is appealing: if a computer can mimic an expert, it could allow scientists to test their questions, explore different scenarios, and gather data without needing to track down busy humans. But a critical question remains: can a machine truly think like a security professional, or does it merely produce a convincing imitation that misses the nuance of real human experience?

A team of researchers set out to answer this question by building a rigorous framework to test whether these artificial models could serve as reliable substitutes for human experts in security surveys. They did not simply ask the models a few questions and see what happened; instead, they designed a comprehensive experiment to see if the computer-generated answers held up under scrutiny. The researchers gathered real survey data from six cybersecurity professionals working in different sectors, ranging from healthcare to finance. These experts had varying levels of experience and held different roles, such as managers, analysts, and specialists. The researchers then asked six different large language models to simulate these specific experts, creating digital personas that matched the real people's backgrounds and job titles.

The experiment was conducted in three distinct ways to test the models from every angle. First, the researchers asked the models to act as individual experts, trying to replicate the specific opinions of the real humans. Second, they asked the models to generate the overall patterns of a whole group, simulating the aggregate results of a large survey. Finally, they tested the models over time, asking them to predict how expert opinions had shifted across several years of real-world data. Throughout this process, the researchers ran each simulation multiple times to see if the models gave the same answers when asked the same question repeatedly, checking for consistency and stability.

The results revealed a clear and significant gap between the artificial simulations and human reality. While the large language models were remarkably consistent with themselves—often giving the same answer when asked the same question multiple times—they failed to capture the true diversity of human thought. Real security experts showed a wide range of opinions; some were confident in their choices, while others were uncertain, and their answers varied depending on their specific experiences and the unique pressures of their workplaces. In contrast, the models tended to converge on a single, safe, middle-ground answer. They smoothed over the disagreements and uncertainties that are natural to human experts, producing responses that were too uniform and too polite.

When the researchers compared the distribution of answers from the models against the real data, the models appeared to be getting the general shape of the data right, but they were missing the details. They tended to avoid extreme or rare opinions, clustering their responses around the average. This "central tendency bias" meant that while the models could generate a plausible-looking survey result, they were effectively erasing the very variations that make human expertise valuable. For instance, when asked about the challenges of adopting new tools, real experts highlighted specific, gritty operational hurdles, while the models offered generic, textbook-style difficulties that lacked the texture of actual practice.

The study also examined whether these models could track changes in the industry over time. The researchers found that the models were largely time-insensitive. Even when asked to simulate responses from different years, the models produced the same static patterns, failing to reflect the evolving nature of cybersecurity threats and practices. This suggests that the models are not truly "learning" or adapting to new information in the way a human expert would; instead, they are recalling a fixed set of concepts that do not shift with the times. This lack of temporal awareness means they cannot be trusted to predict how expert opinions might change in the future or to accurately reflect historical trends.

Ultimately, the researchers concluded that while these large language models are powerful tools, they are not ready to replace human experts in security research. The models are excellent for the early stages of a study, such as helping researchers draft survey questions, testing whether a question is clear, or generating hypothetical scenarios to explore. They can serve as a useful sounding board for ideas. However, they should never be used to replace the actual collection of expert opinions. The unique, varied, and sometimes contradictory nature of human judgment is something that cannot be synthesized by an algorithm. To understand the real world of cybersecurity, researchers must still rely on the messy, complex, and invaluable insights of the humans who live and work within it. The study serves as a vital guide, ensuring that as the field embraces new technologies, it does not lose sight of the human expertise that remains the foundation of security.

Technical Summary: A Framework for Using and Evaluating LLMs as Surrogate Experts in Security Surveys

1. Problem Statement

Security research, particularly within Security Operations Centres (SOCs), relies heavily on expert elicitation through surveys and interviews to understand practitioner workflows, decision-making, and organizational maturity. However, recruiting domain experts is increasingly difficult due to high workloads, burnout, staffing shortages, and confidentiality constraints. These challenges often result in small sample sizes, low response rates (averaging ~44% for online surveys), and fragile insights.

While Large Language Models (LLMs) offer a potential solution by generating synthetic responses at scale, there is a critical lack of methodological guidance on their reliability as surrogate participants. Prior work suggests risks including hallucination, exaggerated agreement, and the smoothing of genuine disagreement. Furthermore, existing evaluations often lack rigorous comparison against domain experts, fail to assess run-to-run stability, or do not account for temporal generalization. The core problem addressed is determining whether synthetic respondents can accurately represent human perspectives in cybersecurity surveys or if they introduce systematic biases that compromise research validity.

2. Methodology

The authors propose a general, reproducible methodological framework for evaluating LLMs as surrogate participants. This framework is instantiated across three distinct experimental settings to assess different modes of survey simulation:

The Evaluation Framework

The framework consists of four core components:

Survey Representation: Questions are encoded as categorical, multiple-choice, or Likert-scale items with standardized formatting.
Prompt Framing and Role Conditioning: Prompts define the task and response format, including persona metadata (e.g., role, industry) but strictly exclude example answers (zero-shot independence) to reveal inherent inductive biases.
Sampling Strategy: To account for stochasticity, each model is queried ten times under identical conditions per setup. This allows for the estimation of stability, dispersion, and run-to-run variability.
Aggregation and Representation: Outputs are aggregated using survey-appropriate rules (e.g., plurality vote for categorical, median for Likert) to form comparable responses.

Experimental Settings

Setting 1: Individual Expert Simulation: The authors collected responses from six real SOC professionals (managers, analysts, specialists) and compared them against LLMs prompted to simulate matching expert personas.
Setting 2: Aggregate Distribution Replication: LLMs were tasked with generating population-level response distributions based on summary statistics from a previous practitioner study on SOC alarm triage.
Setting 3: Temporal Robustness Analysis: Using SANS SOC survey reports from 2017–2025, the authors evaluated whether LLMs could track temporal shifts in practitioner responses, including data points before and after model training cut-offs.

Models and Metrics

Six LLMs were evaluated (GPT-4o, GPT-4, DeepSeek, Llama 3.1-8B, Llama 3.2-3B, Gemini Flash), plus GPT-3.5-turbo for temporal contrast. Evaluation metrics included:

Agreement Measures: Exact match rates, Jaccard/Hamming similarity, and median absolute differences.
Distributional Divergence: Jensen–Shannon Divergence (JSD) and Pearson's $\chi^2$ divergence to compare synthetic distributions against human baselines.
Stability: Intra-LLM consistency (repeatability across runs) and inter-LLM agreement.

3. Key Contributions

The paper makes four primary contributions:

Methodological Framework: A generalizable framework for using and evaluating LLMs as surrogate participants in expert cybersecurity surveys.
Empirical Comparison: A comprehensive comparison between real SOC experts and LLM-generated respondents across individual, aggregate, and temporal settings.
Identification of Failure Modes: The systematic identification of specific biases in LLM-generated survey data, including reduced variance, central tendency bias, and homogenized opinions.
Human Factors Analysis: An assessment of the risks associated with synthetic respondents, specifically the loss of expert variability and the misrepresentation of practitioner perspectives.

4. Results

The study yields several critical findings regarding the reliability of LLMs in this context:

Individual Level (Setting 1): While LLMs produce internally consistent answers when prompted with specific personas, they systematically diverge from human experts. Inter-LLM agreement is low, and alignment with human responses is weak (often zero). The models do not converge on similar simulated responses, nor do they approximate individual expert judgments.
Aggregate Level (Setting 2): LLMs generate highly stable aggregate distributions across repeated runs. However, while pairwise similarity (JSD) suggests models produce broadly similar distributional shapes, absolute divergence measures ( $\chi^2$ ) reveal large mismatches with human baselines. Models tend to over-allocate probability mass to rare options and fail to faithfully reproduce the specific proportions of practitioner responses.
Temporal Level (Setting 3): LLM-generated distributions are time-insensitive. Alignment with human survey data remains consistently low across all years (2017–2025) and does not improve for pre-training cut-off years or degrade for post-cut-off years. Models fail to track historical changes in practitioner responses, suggesting they reproduce static conceptual knowledge rather than evolving professional judgment.
Systematic Biases: Qualitative and quantitative analysis confirms that LLMs exhibit reduced variance, central tendency bias (converging on moderate answers), and homogenized opinions. They tend to reflect generic best practices rather than the nuanced, conflicting, and context-dependent realities of human experts.

5. Significance and Claims

The paper concludes with a modest but firm stance on the role of LLMs in security research:

Not a Replacement: LLMs are not suitable as replacements for human expert elicitation in cybersecurity surveys. Their systematic divergence from human variability and their inability to capture the diversity of practitioner perspectives limit their validity as surrogate participants for drawing empirical conclusions about security operations.
Useful for Piloting and Hypothesis Generation: The authors argue that LLMs are valuable as augmentation tools. Their controllability and scalability make them effective for:
- Piloting survey instruments.
- Stress-testing questionnaire designs.
- Exploring counterfactual populations.
- Generating hypotheses for further human-led investigation.
Methodological Guidance: The work provides practical guidance for researchers, emphasizing that synthetic responses should be treated as stochastic samples from model-specific distributions rather than authoritative expert opinions. Researchers must apply methodological safeguards and transparently report the limitations of LLM-augmented surveys to preserve the integrity of human-centred security research.

The authors explicitly state that while LLMs can support the research process, they cannot substitute for the nuanced, context-aware judgment of human practitioners, and relying on them as such risks misleading conclusions and the erosion of expert variability in security studies.