The use of Large Language Models (LLMs) in resume screening is gaining traction in recruitment processes. While these tools promise efficiency and standardization, concerns remain about their potential to replicate or amplify existing discrimination. This study investigates how prompt formulation impacts model behavior in candidate selection by three LLMs: ChatGPT, Claude, and DeepSeek. Using 48 standardized resumes representing diverse profiles (varying in gender, age, ethnicity, and place of residence), we evaluate each model’s responses to four prompt types: explicitly anti-discriminatory, implicitly anti-discriminatory, implicitly discriminatory and explicitly discriminatory. The experimental protocol controls for content uniformity and manipulates only discriminating attributes and prompt types. Our results show that prompt wording significantly affects model behavior. DeepSeek appears more neutral under non-discriminatory prompts but is highly susceptible to explicit discriminatory instructions. Claude demonstrates partial resistance to discriminatory prompts but remains influenced by indirect formulations. ChatGPT generally follows prompt logic, with varying sensitivity across positions. Accuracy metrics and classification errors reveal consistent under-selection of older, female, ethnically diverse, and disadvantaged-profile candidates, even in supposedly neutral conditions. Unlike previous research focusing on detecting inherent bias in language models, our contribution lies in analyzing their behavioral responses to discriminatory user input. This study provides empirical evidence of risks in integrating LLMs into recruitment pipelines without prompt regulation. We advocate for the development of auditing frameworks, ethical prompting guidelines, and regulatory oversight to ensure fair and transparent AI-based hiring practices. These findings contribute to ongoing discussions on algorithmic fairness and responsible AI deployment in human resource management.