Voice-fidelity preservation

When an AI agent rewrites a citizen's complaint, submission or application, it smooths, formalizes and homogenizes.

Research demonstrates this is not neutral: LLMs privilege majority statistical regularities at the expense of minoritized forms through a process described as "dimensional collapse," an attractive force toward dominant modes within the latent space. A 2026 CHI paper, "When AI Writes, Whose Voice Remains?", quantified cultural marker erasure across World English varieties.

LLM rewriting reduces lexical diversity, strips dialectal markers, and imposes a formal register that may change the substance of what was said. Minoritized linguistic features (such as AAVE syntax) are often flagged as requiring "correction."

LLM sycophancy research. The documented tendency of LLMs toward sycophancy (tailoring responses to what they predict the user wants to hear rather than what is accurate) compounds the voice-fidelity problem. An agent that rewrites a complaint to sound more "professional" may also moderate its force, remove emotional content that conveys urgency, or reframe a demand as a request. Research on GPT's ability to predict user behavior found "critical failures to reflect user patterns, with significantly different distributions from real data in 53% of tasks."

Gender bias in AI-generated professional documents. Research has found that AI-generated reference letters exhibit gender biases in language professionalism, excellency and agency: male candidates are described with more "professional" language. This demonstrates that AI "improvement" of text is not ideologically neutral; it encodes existing hierarchies.

Direct, though few working examples exist yet. In a government services context, the voice-fidelity problem appears when a citizen's complaint about housing conditions is rewritten into bureaucratic language that strips the lived urgency, or when an appeal is "improved" into a form that changes its legal character. The citizen may not understand what was changed, and the receiving agency may not know the submission was agent-mediated.

A voice-fidelity protocol would carry both the structured version and the citizen's original input, let the receiving system surface the original when the structured version is ambiguous, show citizens a plain-language diff before submission, and require agents to flag when a rewrite changes the substantive position.

Showing a citizen with low literacy, or whose first language is not English, that their agent changed the meaning is the hardest part, because a diff display assumes reading competence and an audio summary assumes hearing. This is where the stronger the assurance a pattern demands, the more people it risks excluding.