AI better than humans at spotting romantic sparks, researchers find

Researchers have fed speed-dating conversations into ChatGPT and discovered that large language models can detect who “clicked.” Can AI predict speed-dating success?

AI may soon displace Cupid, with researchers discovering that large language models (LLMs) can detect “romantic attraction during brief getting-to-know-you interactions” and predict whether two people will say “yes” after a speed date.

Researchers from Columbia Business School, University of California, and Northwestern University examined conversations from 964 speed dates. In a paper, they demonstrated that ChatGPT and Claude 3 LLMs can both predict objective and subjective indicators of speed dating success.

ChatGPT was on par with human judges in predicting whether the participants would exchange their contact information. It was also “incremental” to speed-daters’ own predictions.

Researchers recruited 187 undergraduate students via flyers and emails. Each engaged in approximately 12 speed dates, lasting four minutes. During that time, students exchanged an average of 864 words.

Privacy was not included, as the speed dates were evaluated by four independent sources: speed-daters themselves, ChatGPT, eighteen judges, who watched videos, and four judges who only read a random subset of transcripts without access to videos.

While human judges with access to the video were the most accurate, given text only, ChatGPT was on par with human judges in estimating if people would exchange contact information after the date.

“Notably, ChatGPT successfully predicted matching above and beyond participants’ self-reported intentions,” the paper reads.

Researchers also found that ChatGPT observed “unique variance that was neither accessible to the speed-dating participants nor the coders who evaluated the same transcripts.”

Human judges were better at reading participants’ reactions to their speed-date, but not their actual commitment to exchanging contact information, or actual matching.

The correlation between the ChatGPT predictions and actual “matching” was “modest,” although statistically significant.

“Our findings suggest that LLMs can “interpret” some of the conversational nuances that lead strangers to “click” during romantic getting-to-know-you interactions,” researchers said.

Research also found that positive emotions are not a good predictor of speed dating success. Both ChatGPT and human judges relied on them too much in their estimates.

“In contrast, references to money, health, and family were among the strongest predictors of actual matching (with money showing a negative relationship), but neither ChatGPT nor the human raters considered them important predictors,” the paper reads.

The researchers believe that LLMs might yield more accurate judgments when observing longer initial interactions and providing more metrics, such as smiles, eye contact, body language, and paralinguistic cues.