Preserving the person's own voice

As more submissions pass through an AI agent before they reach government, an agent's habit of smoothing, formalizing, and homogenizing what a user wrote becomes the default way it gets rendered.

Research demonstrates this is not neutral: LLM rewriting pulls text toward dominant forms at the expense of minoritized ones. A 2026 CHI extended abstract, "When AI Writes, Whose Voice Remains?", quantified that erasure of cultural markers 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 African American Vernacular English syntax) are often flagged as requiring "correction."

Hansard's substantially-verbatim rule. The UK Parliament's Official Report is edited, but only within a stated limit: it reports what Members said 'substantially verbatim', with repetitions and obvious mistakes removed. An institution that publishes speech on someone else's behalf has fixed the line, and put the obligation to hold it on the editor. Nobody is shown what changed before publication, and corrections run as a separate procedure afterward.

Sharma et al. on sycophancy in language models, ICLR. The documented tendency of language models toward sycophancy, tailoring responses to what they predict the user wants to hear rather than to what is accurate, bears directly on rewriting someone else's words. An agent that rewrites a complaint to sound more professional may also moderate its force, remove emotional content that conveys urgency, or turn a demand into a request.

Gender bias in AI-generated reference letters. The research found AI-generated reference letters carrying gender biases in language of professionalism, excellence, and agency, with male candidates described in more professional terms. Improvement of text by a model is not neutral, and it encodes existing hierarchies.

Direct, though few working examples exist yet. In a government services context, the voice-fidelity problem appears when a user'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 user may not understand what was changed, and the receiving agency may not know the submission was agent-mediated.

The precedents establish that the erasure happens and describe its mechanism; none of them show a preserve-and-flag protocol operating in practice, so that part of the response stays undesigned.

Showing a user 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. The more review the check asks of the user, the more people it risks excluding.