Can AI Really Listen to Job Candidates?

Companies are testing artificial intelligence to find and contact the best candidates. This method, called "active sourcing," combines four steps: automated search, skill extraction, profile matching, and personalized messages. The idea is that AI scans millions of online profiles, identifies the most demanded skills, and offers tailored job proposals. But there is a crucial question: who controls these decisions? The concept of "contestability" means that the people involved can intervene, correct, or challenge the algorithm's choices. In practice, a candidate should be able to ask why they were selected or excluded, and the company should explain the AI's reasoning. This approach boosts transparency and reduces discrimination risks. Yet challenges remain: the data fed to AI can be incomplete or biased, and automatic decisions may appear opaque. To make the system more ethical, experts suggest adding human checkpoints, periodic audits, and interfaces that clearly display evaluation criteria. Future experimental studies will assess how these measures improve candidate trust and hiring quality. The goal is a fairer recruitment process where technology assists, but does not replace, human judgment.

The active sourcing framework relies on supervised and unsupervised machine learning algorithms. In the active search phase, web crawlers extract data from professional platforms, social networks, and open-source project repositories. This data is normalized and indexed for rapid retrieval. Skill extraction employs NLP (Natural Language Processing) techniques such as Named Entity Recognition (NER) and Topic Modeling to identify both hard and soft skills in profiles. For matching, semantic embedding models like BERT or Word2Vec transform job descriptions and candidate profiles into high‑dimensional vectors. Cosine similarity between these vectors determines the match score. Contestability demands that these models be interpretable: Explainable AI (XAI) methods such as LIME or SHAP can generate local explanations of how a specific skill influenced the final ranking. Moreover, a feedback loop is essential, allowing candidates to report errors or bias, with this input fed back into the training set to mitigate model drift. From an ethical standpoint, GDPR compliance is mandatory, requiring anonymization of sensitive data and informed consent for its use. Finally, periodic audits by interdisciplinary committees (ethics, law, data science) can verify that decisions align with fairness and non‑discrimination principles.