An analysis of dating-platform survey data suggests that receptivity to deploying or configuring one’s own AI agent—software designed to hold a conversation—and receptivity to encountering or engaging with another person’s agent are not the same thing, but are closely linked. The result concerns self-reported attitudes, not observed behavior.
Two attitudes, one strong link
The agent-receptivity survey drew 2,617 records from active Fledge.Love users through a voluntary, self-administered in-app questionnaire offered in Russian and English. Its seven-item battery put four items in the principal role—one’s own agent—and three in the counterpart role—another person’s agent; one 1-to-10 item was grouped into five levels.
Statistical comparisons favored the two-role model over a single dimension. The estimated dimensions had a correlation of 0.92, with a bootstrap 95% confidence interval from 0.90 to 0.94. That is a strong association, but it still leaves a measurable distinction between the two forms of receptivity.
A second survey, and a language caveat
A separate survey contained 2,894 records, was fielded in March and April 2026, and measured interest in three passive generative-AI dating features. It was unlinked from the agent survey, so the release cannot show whether interest in those features and receptivity to agents occurred in the same respondents.
Cross-language use is the main technical caution. Tests flagged two counterpart-role items, labeled Y4 and Y5, for differential functioning between the English and Russian forms—meaning the items may not behave the same way across forms. No principal-role item was flagged, while gender-based testing flagged only Y5. The English form had 232 records, so the authors recommend partial-invariance handling or restricting comparisons to invariant items.
The analysis also underwent internal checks. In five-fold respondent-level cross-validation, withholding Y4 produced AUCs—a measure of how well predictions separate responses—of 0.89 for any endorsement and 0.88 for full endorsement. Alternative response-category orderings and latent-class checks were stable, with all 20 random restarts recovering the same solution; these diagnostics were internal and did not validate behavior outside the survey.
What this cannot show
The boundaries are clear. The surveys record stated preferences, not behavioral logs, message content or actual interaction outcomes. They came from voluntary recruitment on a single platform, and engaged users were overrepresented: 80.1% of Instrument A respondents said they opened the app several times a week.
The samples were cross-sectional and unlinked, and free-text responses were not released. The study therefore cannot establish whether conversational agents change dating behavior, influence trust or matching, or affect conversations and relationship outcomes—and it cannot establish representative attitudes across dating-platform users or cultures.
A reusable dataset, with disclosed ties
The dataset and code are archived at Zenodo, with three data files, a bilingual codebook, an anonymization log, processing code, canonical outputs and SHA-256 checksums; the data use CC BY 4.0 and the code is MIT-licensed. The release is useful as a measurement resource, while its real-world predictive value remains unresolved.
Two coauthors were platform employees involved in survey design and fielding; the academic authors independently performed anonymization, analysis and the publication decision. No funding was reported.
Paper data and sources
Original title: Two-sided receptivity to conversational AI agents in online dating: Bilingual survey data from Fledge.Love
Authors: Daria Leshchikova, Valentina V. Kuskova, Dmitry Zaytsev, Valerii Klimov
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text