Research Summary: Whose Voice Survives the Summary? A Voice-Retention Audit of LLM Employee Listening

Research indicates that Large Language Model (LLM) summarization of employee feedback, a practice increasingly adopted by organizations to route insights to leaders, introduces significant representational biases. This process disproportionately silences unique or less frequent voices, particularly criticisms voiced by a single employee, and exhibits a bias against shorter content and non-English submissions, irrespective of sentiment.

Why it matters

The widespread adoption of LLMs for internal communication, particularly in summarising employee feedback, poses a significant risk to organizational understanding of employee sentiment and concerns. Biases in these systems can lead to an inaccurate perception of the workforce, potentially hindering effective decision-making, fostering disengagement, and undermining trust in leadership.

Key insights

• Organizations are increasingly utilizing LLMs to summarize employee feedback for leadership, creating an unaudited layer in the communication process.
• A Voice Retention / Representation Ratio metric was introduced to quantify representational bias in summarization.
• Employees are significantly more reliable in providing criticism than praise; withholding praise is 82 times more common than withholding criticism.
• The LLM summarization pipeline filters content primarily by popularity rather than sentiment.
• A concern voiced only once is dropped 86% of the time in leader-summaries.
• Short content and German-only content are disproportionately lost, with theme retention for German content at 0.14 compared to 0.74 for English.
• The study was conducted on a bilingual (English/German) corpus of 2,586 free-text responses from a global professional service company.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.38818