New Study Reveals How AI Models Disagree on Public Comments
Researchers found that different large language models (LLMs) often categorize the same public comments in conflicting ways. This can shape what policymakers see, potentially biasing decisions. The study proposes a new audit pipeline to flag disagreements for human review.
A recent paper on arXiv (2605.29025) reveals that large language models (LLMs) frequently disagree when categorizing public comments submitted to federal agencies. These models are increasingly used to organize and summarize public feedback on proposed policies, but the study shows that different models can classify the same comment in materially different ways. This matters because a model's categorization directly influences which arguments policymakers notice and consider.
For example, a comment expressing concerns about climate change might be categorized as an "environmental objection" by one model and as "irrelevant noise" by another. The study argues that standard evaluation methods—which only measure accuracy against a small, hand-validated dataset—fail to detect these disagreements. As a result, important public input could be overlooked or misweighted in the policy process.
To address this, the researchers propose an Interpretive Audit Pipeline. The pipeline treats multi-model disagreement as a signal of interpretive complexity and automatically routes disputed comments to human reviewers for closer examination. This approach aims to make the use of AI in public feedback systems more transparent and accountable.
While the paper is technical, the core insight is straightforward: as agencies rely more on AI to process public input, ensuring that models interpret comments consistently is essential for democratic legitimacy. The full study is available on arXiv.