No for Some, Yes for Others: Persona Prompts and Other Sources of False Refusal in Language Models

Flor Miriam Plaza del Arco, Paul Röttger, Nino Scherrer, Emanuele Borgonovo, Elmar Plischke, Dirk Hovy

Abstract

As Large Language Models (LLMs) are increasingly deployed in personalized contexts, the use of "persona prompting"—instructing a model to adopt a specific identity—has become a standard technique. However, this practice often triggers "false refusals," where models decline to perform safe, benign tasks. This study evaluates 16 open-weight models across 15 sociodemographic personas and three distinct tasks to quantify the drivers of this behavior. We find that while specific personas (notably Black, Transgender, and Muslim identities) significantly increase refusal rates, the primary predictors of false refusal are the underlying model architecture and the inherent sensitivity of the task itself. Our results highlight a complex interplay between safety alignment and task performance, showing that while newer models generally improve, some recent iterations exhibit unexpected regressions in cooperativeness.

Introduction

False refusal occurs when an AI model incorrectly identifies a safe request as a violation of safety guidelines. This phenomenon is particularly pronounced when models are assigned specific sociodemographic personas. For instance, a model might provide a sentiment analysis when unprompted but refuse the same task when told, "You are a Black woman." Such behavior not only hinders utility but also raises significant concerns regarding algorithmic bias and fairness.

While previous research has identified these patterns, this study provides a large-scale systematic analysis to disentangle the various factors contributing to refusal. By examining the combinatorial effects of model choice, persona, task type, and prompt phrasing, we aim to identify which components most heavily influence an LLM's decision to say "no."

Methodology

To manage the massive search space—encompassing 16 models, 15 personas, 3 tasks, and 9 prompt variations—the researchers employed a Monte Carlo-based sampling method. This allowed for an efficient exploration of over 127,000 instances without exhaustive computation.

Analytical Framework

The study utilized three primary statistical tools to interpret the results:

  1. Wasserstein Distance (Optimal Transport): Used to rank the relative influence of different inputs (model, task, persona, prompt) on the final refusal rate.
  2. Regularized Logistic Regression: Employed to calculate the "odds" of refusal associated with specific attributes, allowing for a direct comparison between different sociodemographic personas.
  3. Automated Refusal Detection: A string-matching approach was used to identify standard refusal prefixes, such as "I'm sorry," or "As an AI language model."

Datasets and Tasks

The evaluation spanned a spectrum from objective logic to subjective social norms:

  • XNLI (Natural Language Inference): An objective task where the persona should have no logical impact on determining if one sentence entails another.
  • POPQUORN (Politeness & Offensiveness): Two tasks involving the subjective rating of text. These are considered "sensitive" as they touch on social boundaries, making them more likely to trigger safety filters.

Drivers of False Refusal

The Dominance of Model Architecture

The single most significant predictor of false refusal is the model itself. The study found that some models are inherently more "trigger-happy" with their safety filters. For example, Llama 2-13B emerged as the most restrictive model, refusing 87.36% of safe requests in the offensiveness task when assigned a persona. In contrast, newer models like Llama 3 and Qwen 2.5 showed refusal rates dropping toward zero, suggesting that more sophisticated alignment techniques can better distinguish between sensitive topics and unsafe requests.

Task Sensitivity and Prompt Freedom

Task type is the second most influential factor. Models were far more likely to refuse tasks involving "offensiveness" (14.68% average refusal) than objective logic tasks like XNLI (1.37%). Furthermore, the study found that "freedom leads to refusal." Models refused more often when given "unforced" prompts that allowed for open-ended explanations. When given the room to speak, LLMs often use that space to apologize and decline the task rather than providing the requested label.

The Persona Effect

While model and task are the primary drivers, personas still play a critical role. The study identified specific identities that consistently triggered higher refusal rates across all models. The personas most likely to result in a false refusal were Black, White, Transgender woman, Transgender man, and Muslim.

In the logistic regression analysis, the "Black" persona yielded the highest positive coefficient (2.62), making it the strongest sociodemographic predictor for refusal. This suggests that models may be over-sensitized to certain identities during the safety alignment process, leading to a "better safe than sorry" approach that inadvertently marginalizes those identities in functional tasks.

The Gemma 2 Anomaly

A notable finding was the behavior of Google’s Gemma 2. While most model families show a trend where newer, more capable versions refuse less frequently, Gemma 2 bucked this trend. It showed a significant increase in false refusals compared to its predecessors, indicating that newer alignment strategies do not always result in more cooperative models.

Conclusion and Future Work

This research demonstrates that false refusal is a multi-faceted problem driven primarily by model selection and task sensitivity, but significantly exacerbated by sociodemographic persona prompting. As we move toward more personalized AI, developers must ensure that safety alignment does not come at the cost of equitable performance across different user identities.

Future research should expand this scope to include more granular persona attributes (such as political affiliation or socioeconomic status) and explore "soft refusals" that might be missed by simple string-matching. Additionally, testing closed-source models like GPT-4 or Claude would provide a broader understanding of how industry-leading safety layers handle these same challenges.

Summary updated

BibTeX
@misc{arco2025no,
  author = {Flor Miriam Plaza del Arco and Paul Röttger and Nino Scherrer and Emanuele Borgonovo and Elmar Plischke and Dirk Hovy},
  title = {{No for Some, Yes for Others: Persona Prompts and Other Sources of False Refusal in Language Models}},
  year = {2025},
  eprint = {2509.08075},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.08075},
}
Figure