Can AI tell if your date likes you?

AI tools increasingly promise to decode messages, score romantic interest and tell you whether a date “really likes you.” A peer-reviewed study published in Scientific Reports on May 11, 2026 gives those claims a serious test. It also gives them a serious limit.

The researchers found that ChatGPT and Claude 3 could detect some verbal indicators of attraction from complete speed-date transcripts. Their predictions were associated with later ratings and matching decisions. The associations were modest, and some of the cues used by both AI and human observers were not valid indicators of the actual outcome.

What the 2026 study tested

The research reused recordings from eight speed-dating events held at a large Midwestern university in 2007. The original sample included 187 undergraduates, 93 women and 94 men, with a mean age of 19.6. Each four-minute date involved a woman and a man.

Professional transcribers produced the conversations. The research team matched 964 transcripts to the participants’ later reports and decisions. Each transcript averaged 864 words.

This matters because the model was not handed a list of isolated signs. It saw the full back-and-forth of a short conversation. It was then compared with a private, later decision rather than a reviewer’s opinion about whether the date looked successful.

The result, in proportion

In mixed-effects models that accounted for participants going on several dates, ChatGPT’s prediction remained associated with the outcomes. The fixed effect explained a modest share of variance, with marginal R² values of approximately 0.01 to 0.04.

AI and people shared some bad guesses

The researchers examined which conversational features influenced ChatGPT’s predictions and whether those features predicted actual matching. The overlap was imperfect. ChatGPT and human observers often relied on similar cues, but some cues were weakly related, unrelated or pointed in the opposite direction from the later decision.

Positive tone illustrates the distinction. It correlated r = 0.28 with ChatGPT’s prediction, but only r = 0.05 with actual matching. A conversation sounding positive helped the model form a confident impression. It was far less informative about whether both people exchanged contact information.

That is the central risk in an AI “interest detector.” A polished explanation can reveal which cues the model associates with attraction without showing that those cues are valid for the person being discussed.

What newer speech research adds

A July 25, 2026 preprint examined a different dataset: Japanese speed-dating conversations involving 147 adults aged 19 to 60. The researchers combined a transcript-only language model with a supervised model using speech features.

The combined system improved pairwise ranking accuracy over the transcript-only model in all four tested settings. Gains in participant-level Pearson correlations varied, however, and none remained statistically significant after correction. The authors’ conclusion was conditional: speech can add information in some settings, not that audio creates a universal attraction detector.

The preprint is useful because it changes the question. Conversation text may not contain every relevant feature. Tone and speech patterns can add signal. They still do not turn prediction into access to another person’s private state.

What the study does not establish

• It does not show that an AI tool can diagnose attraction from a few screenshots or selected messages.
• It does not establish accuracy across ages, cultures, sexual orientations, relationship stages or ordinary dates.
• It does not show that the studied model versions behave like every current or future system.
• It does not convert mutual contact exchange into a complete measure of attraction.
• It does not measure consent. Attraction, a second date and consent to a particular act are different questions.

The participants consented to research recording and analysis. That does not create permission to upload another person’s private messages or recordings to an AI service for a secret assessment. Privacy remains part of the decision.

A better use of AI when you are uncertain

AI can be more useful as a check on your own reasoning than as a machine that assigns someone else a hidden feeling. Ask it to separate observation from inference, remove pressure from an invitation or identify where you are treating ambiguity as certainty.

The bottom line

The 2026 study is evidence against two extreme claims. AI was not reading random noise. It detected a small amount of real signal in complete speed-date conversations. It also did not come close to reliably knowing what one person wanted.

Use the result the way the size permits. A model’s interpretation can be one hypothesis. The other person’s voluntary response to a clear, low-pressure invitation is better information.