Research Summary: Reasoning Enhances Robustness to Prompt Injection in LLM-Based Consensus

Recent research demonstrates that while large language models (LLMs) used for preference aggregation and consensus generation in digital democracy are vulnerable to prompt-injection attacks, incorporating reasoning mechanisms improves their robustness against adversarial manipulation.

Why it matters

As institutions explore automated tools for public consultation and collective decision-making, the integrity of aggregated outcomes becomes paramount. Adversarial prompt injection poses a direct threat to the legitimacy of algorithmic consensus by enabling minority interests to distort deliberative outcomes. Incorporating reasoning-capable architectures is therefore essential to prevent malicious interference in automated governance processes.

Key insights

• Large language models are increasingly deployed to aggregate preferences and draft consensus statements in digital democracy initiatives.
• Prompt-injection attacks can compromise consensus-generating LLMs by amplifying specific viewpoints, erasing opposing perspectives, or derailing discussions toward irrelevant subjects.
• Empirical testing using UK public policy experiment data indicates that model reasoning capabilities enhance resilience against adversarial prompt manipulation.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2508.04281