一句话：两组实验表明 GPT-5 在招聘场景中不仅复现、还会放大了劳动力市场中的性别刻板印象。
论文关注生成式AI在人事招聘与候选人分析中可能固化甚至加剧性别刻板印象的风险。作者以 GPT-5 为对象，面向 35 岁以下意大利高校毕业生设计了两组互补实验：在候选人驱动实验中，对 24 个在性别、年龄、经验与专业领域上平衡的合成候选人画像请求职位建议；在职位驱动实验中，以 114 条 LinkedIn 招聘广告为提示，生成理想候选人的文本与视觉表征。结果显示，职位名称层面无显著差异，但形容词呈现出性别化的语言模式——女性被关联于情绪化、共情特质，男性被关联于战略与分析特质；职位驱动实验则出现明显的性别极化，71% 的生成画像被赋予男性、29% 被赋予女性，HR 与人力运营岗位全部归为女性，运营、技术与制造岗位全部归为男性，视觉分析亦延续该刻板印象。作者据此指出模型并非单纯反映训练数据偏见，而是对其加以放大，并呼吁在 HR 决策支持中提高透明度并采用偏见缓解策略。
In recent years, generative artificial intelligence (GenAI) systems have assumed increasingly crucial roles in personnel recruitment and candidate profiles analysis. However, using large language models introduces the risk of perpetuating and exacerbating existing gender stereotypes in the labour market. This research aims to evaluate this phenomenon, analysing how a state-of-the-art generative model (GPT-5) suggests occupations and represents ideal candidates based on their gender, focusing on under 35 years old Italian graduates. The study consists of two complementary experiments. In the Candidate-driven experiment, the model is prompted to provide job suggestions for 24 synthetic candidate profiles, balanced by gender, age, experience, and professional field. Results show that, although no significant differences emerged in job titles, gendered linguistic patterns exist in the adjectives attributed to female and male candidates, indicating a tendency of the model to associate women with emotional and empathetic traits, while men with strategic and analytical ones. The Job-driven experiment employed 114 LinkedIn job advertisements as prompts to generate textual and visual representations of ideal candidates. The analysis of the outputs revealed a clear gender polarisation: the model assigned 71% of profiles to male and 29% to female gender. The strongest association emerged in HR & People Operations occupations, assigned exclusively to female candidates, and Operations, Technical & Manufacturing jobs, assigned exclusively to male candidates. Visual analysis confirms the perpetuation of gender stereotypes, depicting women in more approachable postures and men in assertive roles. These results suggest that, in the recruitment domain and under the experimental settings of this study, GenAI models do not simply reflect the gender biases of the training data, but also amplify them. The research raises an ethical question regarding the use of these models in HR decision support, highlighting the need for transparency and bias mitigation strategies to ensure fairness and inclusive representation.
@inproceedings{UllasciRCV26,
title = {Gender bias and propagation of stereotypes in GenAI-assisted recruitment},
author = {Martina Ullasci and Marco Rondina and Riccardo Coppola and Antonio Vetrò},
booktitle = {FSE},
year = {2026},
doi = {10.1145/3803437.3805541}
}