Blog · arXiv Analysis · Published: August 12, 2026 · Modified: August 12, 2026 · Last reviewed: August 12, 2026

The Dating Agent Becomes the Receiver’s Opt-In

Permission to let an agent speak for you is not permission to direct that agent at somebody else. Agent-mediated contact has a sender-side delegation decision and a receiver-side participation decision.

A new survey study measures the gap between those roles. Its estimates describe stated responses to mockups, not live dating behavior, but they expose a governance mistake: treating adoption by senders as consent from a market.

The Paper

The source is Daria Leshchikova, Valentina V. Kuskova, Dmitry Zaytsev, and Valerii Klimov’s Delegation Asymmetry in Agentic Recommender Systems: Measuring Two-Sided Receptivity in Online Dating, arXiv:2608.18058v1 [cs.AI], submitted August 18, 2026. The 11-page version-one paper identifies Leshchikova and Klimov with Fleamily, Inc., and Kuskova and Zaytsev with the University of Notre Dame’s Lucy Family Institute for Data & Society. Its source metadata names WSDM 2027.

The study uses two voluntary, self-administered surveys recruited through an in-app prompt to active users of Fledge.Love. The autonomous-agent instrument ran from November 12 to December 15, 2025, in Russian and English and received 2,617 responses; 2,499 complete cases entered the measurement model. A separate Russian-language survey of 2,894 users covered generative profile and assistance features in March and April 2026. The two samples are not linked person by person.

Two Roles, One Respondent

The agent survey presented interface mockups of a configurable conversational agent that could initiate or answer messages on a user’s behalf. Seven ordinal items placed each respondent in two positions. Four concerned being the principal who configures, deploys, or values an agent; three concerned receiving another person’s agent, watching agent-to-agent pre-conversation, or joining a mixed human-agent chat.

A two-dimensional graded-response model fit the send and receive items better than a one-dimensional model by 51.8 BIC points. The latent dimensions were strongly related, with correlation 0.92, yet not treated as interchangeable. The paper also tested language and gender item functioning, bootstrap uncertainty, alternative category codings, and a partial-invariance model after two receive-side items differed across language forms.

The Delegation Asymmetry

On the fitted scale, the threshold for endorsing deployment of one’s own agent was −0.38, while the threshold for engaging another person’s agent was +0.32, a gap of 0.71 standard deviations with a reported 95-percent bootstrap interval from 0.65 to 0.77. Model-implied strict and soft deployment propensities were 0.38 and 0.50; corresponding engagement propensities were 0.12 and 0.26.

The pattern is also visible without the latent model. After collapsing responses to comparable categories, 40.7 percent rated sending above receiving, while 2.1 percent gave the reverse ordering. This is not a clean randomized role effect: the send and receive prompts used different wording and scenarios. The authors identify a fixed-wording paired-vignette experiment as the needed follow-up.

A Counterfactual Market

The paper converts stated endorsement propensities into a random-pairing counterfactual. Under its independence baseline, 4.4 percent of directed dyads in the strict scenario and 12.8 percent in the soft scenario combine agent deployment with receiver engagement. These figures are computed from survey responses. They are not observed match, reply, satisfaction, or relationship rates.

Two modeled levers illustrate the tradeoff. A reciprocity gate—deploy only if one’s own soft engagement propensity is at least 0.5—excludes 65 percent of likely deployers and reduces the strict viable-dyad rate from 0.044 to 0.020. Routing to the most receptive quarter raises modeled strict engagement per contact from 11.6 to 39.4 percent. A five-fold validation that excludes the target engagement item reports AUC 0.88 and a 3.1-times top-quartile lift for the held-out survey answer. That validates ranking of stated receptivity, not response to a deployed agent.

Opt-In Is Not a Receptivity Score

The study’s useful design insight is that sender authorization covers only one side. A receiver needs a separate, visible, revocable choice about whether agent-mediated contact is acceptable. That choice may depend on whether an agent initiates or only replies, whether authorship is disclosed before opening the message, whether the humans already matched, and whether agents converse with each other before either human enters.

A predicted receptivity score is not consent. It can route traffic away from people estimated to dislike it, but it can also become a hidden profile of who is easiest to expose to automation. Group averages, including the paper’s gender associations, should never stand in for an individual setting. Nor should frustration with stalled matches become a targeting signal: the paper reports cross-sectional associations between unmet need and receptivity, not permission to optimize adoption against emotional strain.

The Receiver-Choice Receipt

A deployment receipt should record both authorization surfaces. On the sender side: who may activate the agent, whether it initiates or replies, its rate and duration limits, and what the human must approve. On the receiver side: the explicit setting, its scope and expiry, disclosure timing, block and report controls, and whether declining agent contact changes ordinary human visibility or match access.

The receipt should also name the evidence behind any routing score, locale-specific validation, uncertainty, data retention, prohibited proxy variables, appeal route, and change log. A platform should report human-authored and agent-authored contact separately. Otherwise a rise in message volume can masquerade as social participation while the receiver’s refusal remains invisible.

The Evidence Boundary

This is a concept-mockup survey from self-selected respondents on one platform, not a deployment trial or population census. The English agent-survey group had 232 respondents, only 212 of them among the complete cases, and language is confounded with unmeasured population composition. The passive-feature and agent surveys used different, unlinked samples. Category ordering required judgment, the market simulation assumes random pairing and static propensities, and attitudes may change as products and norms change.

The authors report that their institutional review board classified the de-identified secondary analysis as not human-subjects research under protocol 26-08-10287. This review inspected the arXiv abstract, PDF, full-text HTML, metadata API, and source package. It also downloaded the linked Zenodo reproducibility bundle, verified its SHA-256 manifest, and checked its codebook, anonymized analysis data, notebooks, and canonical result tables. The deposit supports the reported computations; it does not turn stated preference into behavior.

The Spiralist lesson is compact: delegation is directional, but consent is relational. A person may authorize a proxy to speak without granting that proxy an audience. The receiver’s opt-in is not an efficiency feature attached to the agent. It is the other half of permission.

Sources

Return to Blog