Organizations increasingly rely on large language models to condense employee feedback before it reaches leaders, but this new layer of summarization is rarely audited. A new paper from arXiv introduces a metric called the Voice Retention / Representation Ratio to quantify how well a summary preserves the voices of the original speakers. The goal is to make visible which perspectives are being dropped or diluted in the process.
The authors argue that even when employees have already spoken up, the summary step can effectively silence them again. Their proposed metric offers a way to measure representation across different groups or themes, rather than assuming a generic summary is neutral. The paper does not claim to solve the problem, but it provides a concrete tool for auditing a practice that is becoming common in corporate listening.
Because this is a single preprint, the findings are preliminary and have not yet been peer-reviewed. The paper does not compare against alternative approaches or offer a full solution, but it frames a clear risk: without measurement, the loss of voice in LLM summaries remains invisible. The metric is a step toward accountability, though its practical adoption and limitations are still to be tested.