Franklin AI Research Explainer: Prompt Injection in Automated Résumé Screening with Large Language Models: Single and Multi-Injection Settings

As Large Language Models (LLMs) become standard tools for screening and ranking job applicants, they create new opportunities for candidates to manipulate the hiring process. This paper investigates "prompt injection" in the context of recruitment—specifically, the use of subtle, self-promotional text added to a résumé that does not add actual qualifications but is designed to trick the LLM into giving the candidate a higher ranking. The researchers conducted controlled experiments to determine how effective these manipulations are and how they impact the fairness of automated hiring systems.

Understanding Prompt Injection in Hiring

In this study, prompt injection is defined as a strategic attempt to influence an LLM’s evaluation of a candidate without providing new, verifiable skills or experience. The researchers sought to understand how these subtle textual additions affect the ranking process. By testing various scenarios, they examined whether these manipulations could successfully bypass the intended logic of an automated screening system.

The Impact of Widespread Manipulation

The study reveals that the effectiveness of prompt injection is highly dependent on how many candidates use it. When résumé quality is similar across a pool of applicants and only a few people use prompt injection, those individuals can reliably improve their rankings. However, this advantage is not sustainable; as more candidates begin to use these tactics, the effectiveness of the injection diminishes. When manipulation becomes widespread, the system’s ability to distinguish between candidates effectively collapses.

Fairness and Candidate Quality

The researchers also explored how these tactics function when candidate quality is heterogeneous—meaning there is a clear difference in the actual qualifications of the applicants. While prompt injection is generally less effective in these diverse pools, the study found that it can still occasionally allow lower-quality candidates to outrank those who are more qualified. This finding highlights significant fairness concerns regarding the use of LLMs in recruitment.

Key Takeaways for Automated Screening

The research concludes that LLM-based screening systems are most vulnerable to manipulation when two conditions are met: the number of people attempting to manipulate the system is low, and the differences in actual candidate quality are small. These findings suggest that while LLMs offer efficiency in hiring, they are susceptible to strategic gaming that can undermine the integrity of the selection process.