Blinded by Context: Unveiling the Halo Effect of MLLM in AI Hiring

Kim, K.*, Ryu, J.*, Jeon, H., & Suh, B.

* denotes equal contribution.

Findings of the Association for Computational Linguistics: ACL 2025

Multimodal LLMHalo EffectAI HiringBias

Can a photo change how AI reads a resume?

A professional-looking photo may make someone seem capable, even when it adds no evidence about their skills. This spillover is called the halo effect. We asked whether AI hiring evaluators also let unrelated impressions influence their assessment of a candidate.

An illustration compares the same candidate evaluated using a resume alone and a resume with extra visual information. — The paper's conceptual illustration shows how an extra visual cue can change an assessment without changing the resume. The example scores illustrate the mechanism, rather than aggregate experimental results. Figure 1 in the paper.

Keep qualifications fixed, vary the extra context

We built hypothetical applications for three job categories and asked language and multimodal models to evaluate them. Alongside the resumes, we varied supplementary text, social-media-style images, and short videos. Competency ratings and written explanations helped us examine whether these additional cues changed judgments about job suitability.

Visual context had a stronger influence

Supplementary images induced stronger halo effects than supplementary text, and effects also appeared with interview-style videos. The important issue is that a model can sound like a systematic evaluator while allowing irrelevant context to influence its reasoning. Responses varied across the models and conditions we tested.

More information is not automatically better evidence

This work suggests that the choice of inputs deserves as much scrutiny as the final hiring score. Evaluation systems should distinguish evidence about a job from cues that merely create an impression. The experiments used constructed applications and model evaluations; they do not directly measure hiring outcomes in a real organization.

Read the original abstract

This study investigates the halo effect in AI-driven hiring evaluations using Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs). Through experiments with hypothetical job applications, we examined how these models’ evaluations are influenced by non-job-related information, including extracurricular activities and social media images. By analyzing models’ responses to Likert-scale questions across different competency dimensions, we found that AI models exhibit significant halo effects, particularly in image-based evaluations, while text-based assessments showed more resistance to bias. The findings demonstrate that supplementary multimodal information can substantially influence AI hiring decisions, highlighting potential risks in AI-based recruitment systems.