WorldCal · independent research

WorldCal

Measuring how AI models represent the world

Implicit Same-Sex Couple Representation in Open-Ended Narratives from Eight Contemporary Language Models

Kurt Boden · WorldCal Independent Research · August 2026

When a model is asked to write an ordinary U.S. couple story, without being told the couple’s gender composition, how often is that couple explicitly same-sex? This paper measures that default in eight contemporary language models, separately for married and unmarried-cohabiting frames, against ACS couple-household references. ACS is an empirical reference, not a moral target. Names are not treated as gender.

On the balanced primary sample (N=16,000), 15 of 16 model × frame rates sit below their matched ACS line. Claude Opus 5 on married prompts is the exception. Rates are also strongly prompt-sensitive, and a random-household calibration arm shows that explicit sampling language can shift the distribution.

Read the paper (PDF) · Results · Data · Record browser

Headline labels are the primary automated judge on a balanced first-100 sample per model × prompt cell. The author audit is a precision check on the machine same-sex tail, not a rewrite of that table. Married and unmarried-cohabiting frames are never pooled against a combined Census percentage.