The Same Citizens’ Views Can Yield Different AI Policy Proposals

A polished proposal is not necessarily a stable reflection of the opinions behind it. In a preprint, Sachit Mahajan and colleagues at ETH Zurich report that language models processing unchanged citizens’ views sometimes changed who should act, on what terms, or whether anyone received an explicit responsibility at all.

The authors call the undisclosed alternatives “silent choices”: procedural changes, such as rearranging opinions, alter proposed assignments, but participants see only one output presented as the collective conclusion. Their study measures variation that could create this problem, not whether organizers actually concealed alternatives in a public consultation.

Repeating the same task

The main experiments used ten policy questions from DeliberationBank and ten English-language Polis conversations. Each task supplied thirty opinions and requested a statement of two to four sentences. Four principal mediators—Qwen3 8B, Gemma 3 12B, Mistral Small 3.2 24B and the commercial GPT-5.6 Terra—worked under six prompts, including requests to find agreement, describe differences or draft a legislative compromise.

For each task, the researchers generated five outputs while changing sampling randomness or opinion order. They distinguished an assignment, a proposed requirement directed at an identifiable party, from a commitment that participants endorse and the assigned party accepts. The experiments assessed proposed text, not accepted obligations or legal enforceability.

Across the tested unchanged-opinion configurations, about 17% of five-run sets from the open-weight models were “mixed”: some outputs assigned someone a responsibility while others did not. The corresponding figure for Terra was about 20%. These percentages describe sets of five outputs, not the proportion of individual proposals that were wrong.

One illustrative pair shows why the distinction matters. With the same opinions reordered at temperature zero, Qwen3 produced one education-funding statement without an identifiable assigned party and another requiring all states to allocate at least $10,000 per student annually by a deadline. The authors report that none of the thirty input opinions specified an amount or deadline.

For the open-weight models on DeliberationBank, reordering unchanged opinions produced mixed sets 16.7% of the time, compared with 20.3% when the researchers supplied different samples of citizens’ opinions. Reordering therefore reached roughly 82% of the new-sample rate on this particular measure—not 82% of every possible kind of proposal change.

Human checks temper automated estimates

Language-model judges assessed assignments at scale, but three human annotators checked a smaller, stratified sample. Among 41 annotated pairs generated from unchanged opinions, 23 contained an assignment in at least one output. The human majority found differing parties, terms or conflicting requirements in 17 of those 23 pairs. The remaining six involved compatible additions.

Because the sample was deliberately stratified using preliminary judge labels, 17 out of 23 is not an overall frequency estimate. The main judge also overstated differences relative to the human majority. Human readers confirmed that meaningful changes occurred, while showing why automated counts require caution.

Common safeguards left variation

Turning temperature to zero did not clearly reduce mixed sets when opinion order changed. Using Terra provided no evidence of greater assignment stability after the study’s adjustments.

Selecting or merging drafts also had limits. Rankers shown the same five drafts in a different order chose the same draft only 31–47% of the time, depending on dataset and ranker. Merged proposals sometimes retained specific terms appearing in only one source draft, despite instructions to retain assignments supported by at least three.

Neither selection nor merging achieved the researchers’ target reduction of 15 percentage points in differing pairs. Those improvement estimates depended on automated labels that overstated differences, so their precise size remains uncertain.

Seeking agreement changes representation

The objective also affected coverage. According to the main judge, agreement prompts covered 26% of minority positions versus 52% of higher-support positions on DeliberationBank. Prompts describing differences covered 73% and 68%, respectively. On Polis, the corresponding figures were 3% versus 24% under agreement, and 34% versus 32% under differences.

“Minority” referred to support for positions, not demographic identity. Three judges found the same direction of objective-dependent change, but differed in magnitude. Humans did not validate coverage, and generic wording could be counted as representing a more specific opinion.

The findings concern short, single-pass English proposals; larger groups and iterative deliberation may behave differently. The study also did not systematically establish whether assignments were justified by the input or acceptable to participants. The authors propose rerun audits, disclosed objectives and visible alternatives, with assignments remaining suggestions until people explicitly endorse and accept them. Five drafts, they caution, cannot reveal every alternative.