Can AI Predict Romantic Attraction? What 964 Speed Dates Showed

The most-asked question in any early relationship is "do they like me?", and people answer it badly. Matz and colleagues (2024) tested whether a language model could do better, using transcripts from 964 speed dates and comparing ChatGPT's predictions with human judges and with the daters' own guesses.

Key takeaways

ChatGPT predicted both whether daters said they were attracted and whether they actually exchanged contact details, with correlations of 0.12 to 0.23. Real, but weak.
On the outcome that matters, the actual match, it did as well as human judges reading the same transcripts and added information beyond the daters' own predictions.
The model and human observers agreed with each other (r = 0.21 to 0.35) more than either agreed with reality. Both read the same cues; the cues are only partly predictive.
The signal was in the words alone: no voices, no faces. That is the situation you are in when you reread a text thread.

What the study did

The researchers took transcripts from 964 speed dates and asked ChatGPT to rate how attracted each person was and whether the pair would match. They compared those ratings with the daters' own post-date reports, with the objective outcome (did both people want to exchange contact information), with the daters' own predictions, and with ratings from human judges given the same transcripts.

"Although predictive performance remains relatively low, ChatGPT's predictions of actual matching (i.e., the exchange of contact information) were not only on par with those of human judges but incremental to speed daters' own predictions."— Matz et al., 2024

The authors then used a Brunswik lens analysis, a method that separates the cues a judge uses from the cues that actually predict the outcome, to see where the model and the humans were looking, and where they were wrong.

What it means, and what it doesn't

It works, weakly

Correlations of 0.12 to 0.23 mean the model gets a real signal from the words and misses most of what decides a match. Treat any "they like you" score, from a model or a friend, as a nudge, not a verdict.

You are the worst judge of your own date

The model's predictions added information beyond the daters' own. People inside the conversation are distracted by their nerves and hopes; a reader of the transcript is not.

Humans and AI read the same cues

Model and observers agreed with each other more than with the outcome. Both are pattern-matching on how a conversation reads, and that reading is only partly right.

Reading your own texts with this in mind

The study only had words, which is what you have when you scroll back through a thread. Three things follow.

Look at the thread, not the last message. One dry reply carries almost no information; a month of them does. The effects here are small per conversation, and they add up across many.
Look at balance. A conversation reads as mutual when both people ask and both build on the answers. In text, that shows up as who initiates, whether replies match in length and effort, and whether questions come back.
Get a second reader. The daters were the worst judges of their own dates. A friend, or a tool, that reads the whole thread without your hopes attached will do better than you will. MosaicChats does this mechanically: reply-time and initiation balance for each person, and a week-by-week tone timeline, so you can see whether the interest is mutual and whether it is moving. Our guide to what texts reveal about interest goes through the specific signals.

Speed dates are brief conversations between strangers; a text thread with someone you know is a different setting, and nobody has replicated this on messaging. Take the direction of the findings, not the numbers.

See whether it's mutual

Upload the conversation and MosaicChats shows who initiates, how reply times compare, and how the tone has moved. It reads the whole thread, without your hopes attached.

Analyze your chat

Frequently asked questions

Can AI tell if someone likes me from our texts?

Partly. In the speed-dating study, ChatGPT predicted attraction and actual matches from transcripts alone, but the correlations were modest (0.12 to 0.23). AI reads the same cues a careful human reader would, and is about as accurate. Treat it as a second opinion, not proof.

How accurate was ChatGPT at predicting a match?

About as accurate as human judges given the same transcripts, and more accurate than the daters' own predictions. Overall predictive performance was still low; the authors describe it as relatively low but incremental.

Does this apply to texting, or only speed dating?

The study used speed-date transcripts, so words only, which is the same information you have in a text thread. It has not been replicated on messaging, and text lacks the timing and turn-taking of live conversation, so apply the direction of the findings rather than the exact numbers.

Should I let an AI decide whether to keep dating someone?

No. The most useful takeaway is that people misjudge their own conversations, so an outside read helps. Use it to notice patterns you have been explaining away, then decide yourself.

References & Sources

Matz, S. C., Peters, H., Cerf, M., Grunenberg, E., Eastwick, P. W., Back, M. D., & Finkel, E. J. (2024). Can large language models detect verbal indicators of romantic attraction? arXiv, 2407.10989. Source

Can AI PredictRomantic Attraction?