Organizations are increasingly routing employee feedback through large language model (LLM) summaries before delivering it to leaders. This unaudited layer can mute voices that have already been spoken. We introduce the Voice Retention / Representation Ratio (VRR) metric to quantify representational bias in summarization and apply it to a bilingual (English/German) corpus of 2,586 free‑text responses from a global professional‑services firm. The study reveals three key findings:
- Employees provide criticism far more reliably than praise; withholding praise occurs 82 times more often than withholding criticism.
- Across 45 leader‑generated summaries, the pipeline filters by popularity rather than sentiment: criticism tends to survive, while a concern voiced only once is dropped 86% of the time. Short entries and German‑only content suffer similarly (theme retention 0.14 vs 0.74; German directionality poorer).
- After controlling for frequency, sentiment has no independent effect on retention; the harm is driven by prevalence, a nuance missed by sentiment‑only audits.
A targeted prompt recovers only explicitly named themes and does not improve overall retention. We contribute the VRR metric, field evidence, and a disaggregated voice‑retention card for future audits.
Review