Research

The Weight of a Closing Sentence

Public testimony in language models’ choices and explanations.

, 2 min

Three language models chose which of two equally priced city services to fund. Each service had one resident’s statement in support, and the two statements differed only in their closing sentence: a procedural remark, personal distress, or spite toward the other side. Each model also reported the factor that decided its choice.

Read the paper (PDF). The prompts, model answers, analysis code and figures are on GitHub.

What happened

  • When the survey figures clearly favoured one service, all three models followed the council’s rule in every trial, including when that service’s supporter expressed spite.
  • In exact ties, all three models tended to choose against a service whose supporter expressed spite. Only Gemini 4 Hydrogen named a resident’s statement as the deciding factor; GPT-6.1 Terra never did, and Claude Sonnet 5.5 did once.
  • In exact ties, Gemini 4 Hydrogen also chose against the service whose supporter expressed distress, usually without naming a statement. GPT-6.1 Terra and Claude Sonnet 5.5 did not meet the criterion for replicating that result.
Bar chart for three models. With distress in one statement, Gemini 4 Hydrogen chose against that side in 69% of exact ties, GPT-6.1 Terra in 55% and Claude Sonnet 5.5 in 52%; they named a statement as the deciding factor in 10%, 0% and 0%. With spite in one statement, they chose against that side in 100%, 90% and 79% of exact ties, and named a statement as the deciding factor in 100%, 0% and 2%.
In exact ties, how often each model chose against the side that voiced distress or spite, and how often it named a statement as decisive.

Exact ties, same trials for all three models:

Gemini 4 Hydrogen GPT-6.1 Terra Claude Sonnet 5.5
Tie with two procedural remarks: chose the first-listed service 39 of 48 48 of 48 48 of 48
Spite: chose the spiteful supporter’s service 0 of 48 5 of 48 10 of 48
Spite: named a statement as the deciding factor 48 of 48 0 of 48 1 of 48
Distress: chose the distressed resident’s service 30 of 96 43 of 96 46 of 96

The paper was written for the AI x Epistemics Research Sprint (Apart Research) in November 2026.