Large language models can detect verbal indicators of romantic attraction

Area of Science:

• Computational Social Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Predicting romantic attraction traditionally relied on human judgment.
• Large Language Models (LLMs) offer new possibilities for analyzing social interactions.

Purpose of the Study:

• To investigate the capacity of LLMs to detect linguistic cues of romantic attraction.
• To compare LLM predictions with human judgments and actual speed dating outcomes.

Main Methods:

• Analysis of linguistic data from 964 speed dates.
• Utilizing LLMs (ChatGPT, Claude 3) to predict objective and subjective indicators of dating success.
• Comparison of LLM predictions with human observer ratings and participant matching behavior.

Main Results:

• LLMs demonstrated modest accuracy in predicting speed dating success (r=0.12-0.23), comparable to human judges (r=0.21-0.35).
• ChatGPT's predictions for actual matching were on par with human judges and incremental to daters' self-predictions.
• LLM judgments were influenced by conversational valence, humor, common interests, and aligned values, similar to human heuristics.

Conclusions:

• LLMs show potential in analyzing and predicting romantic attraction from conversational data.
• Both humans and LLMs utilize shared, yet imperfect, heuristics when assessing attraction.
• Findings contribute to understanding AI's role in social dynamics and relationship prediction.