Delegation Asymmetry in Agentic Recommender Systems: Measuring Two-Sided Receptivity in Online Dating

Autonomous LLM agents that converse on a user’s behalf are an emerging design pattern in matching platforms, yet their viability depends on a condition rarely examined: users must accept not only delegating conversation to an agent, but also receiving agent-mediated communication from others. We study this condition using two large-scale surveys of active users of a major dating platform (N = 2,894 on generative profile features; N = 2,617 on autonomous conversational agents, fielded in two languages). We develop a latent-variable measurement model of agent receptivity based on graded response models with latent regression, and show via model comparison that willingness to send and willingness to receive agent communication are distinct constructs. They are highly correlated (ρ= 0.92) but separable (ΔBIC = 52), with partial measurement invariance across languages. The model quantifies a systematic delegation asymmetry: deploying one’s own agent requires far lower receptivity (threshold θ= −0.38) than engaging a counterpart’s agent (θ= +0.32; full engagement +1.39), and mean deployment propensity exceeds engagement propensity roughly threefold. Under a random-pairing counterfactual derived from stated receptivity, only 4–13% of directed dyads combine agent deployment with receiver engagement, with a pronounced genderdirectional imbalance. Design counterfactuals quantify the levers: a reciprocity requirement halves interaction volume by excluding two-thirds of would-be deployers, while routing agent contacts on receive receptivity triples per-contact engagement: a lift that survives out-of-sample validation in which the routing score is estimated without the target item (AUC 0.88, 3.1× quartile lift under respondent-level cross-validation). We discuss implications for agentic recommender design, including disclosure, opt-in mechanics, and receptivity-aware matchmaking.

Introduction. Somewhere tonight, a dating-app user will get a warm, well-crafted opening message from a match: composed, sent, and if things go well, followed up on by a language model while its principal sleeps. The scenario is no longer speculative. Recent survey evidence suggests that more than a quarter of U.S. singles have already used AI somewhere in their dating lives, and a majority of dating-app users believe they have, at some point, exchanged messages with an algorithm rather than a person [2]. A cottage industry of “AI wingman” tools writes openers and replies on demand; users who receive such messages report feeling deceived when they find out [2], and merely suspecting AI authorship is enough to depress trust [18, 19]. What began as Cyrano de Bergerac with a keyboard - one party quietly borrowing better words - is being industrialized. Platforms are now productizing the next step. The first wave of generative features was assistance: AI-written profile summaries, suggested openers, tone polish, with the user remaining the author of record. The emerging wave is delegation: autonomous LLM agents that hold conversations on the user’s behalf are initiating contact, exchanging messages, screening counterparts, even negotiating common ground with the other side’s agent before the humans ever speak. The research community is actively building this machinery [28, 39, 43, 44], and online dating is its natural frontier: communication is not a feature of the product but the product itself, interaction volume is high, and the stakes are as personal as software gets. There is, however, a condition for this market that the building has run ahead of. Agent-mediated matching is a two-sided proposition: every delegated message needs a willing receiver on the other end. A market in which many users deploy agents but few will engage with one is not a matching market at all; it is automated outreach landing on closed doors, at scale, in the most trust-sensitive consumer domain there is. Yet the two sides of this condition have never been measured together. A large literature measures acceptance of AI tools by their users [9, 38], and a growing one documents receiver-side penalties when AI involvement is suspected [14, 18]; what is missing is a joint, within-person measurement of willingness to send and willingness to receive agent communication on a common scale - the quantity that determines whether both sides of a delegation market exist, and the quantity this paper provides. We measure both sides. Working with two large surveys fielded to active users of a major dating platform (N = 2,894 on generative

Related work. Online dating as a matching market. Online dating is now the dominant channel through which couples meet in several countries [31], and its market structure is well documented: users sort on observable attributes with strong revealed preferences [16]. As users direct most of their attention upward toward more desirable counterparts [7], most first messages go unanswered: reply behavior is scarce, skewed, and predictable from user and dyad features [40]. The domain’s relationship psychology is surveyed by Finkel et al. [12]. This attention-scarcity structure is the backdrop for our results: automated delegation promises to multiply first-message volume precisely where reply willingness is already the binding resource, and our receive-side measurements quantify how much colder the reception gets when the message is agent-authored. Reciprocal recommendation. Algorithmically, dating recommenders differ from item recommenders in requiring mutual interest. The reciprocal-recommendation line runs from RECON’s harmonic combination of directional preferences [29, 30] through behavioral models on large platforms [41], latent-factor formulations [24], and the survey of Palomares et al. [25]. Recent work treats matching platforms explicitly as two-sided markets with exposure trade-offs: fairness-aware reciprocal recommendation [37], welfare-based balancing of match rates and fairness, and off-policy evaluation for matching interventions [15]; agent-based models reproduce engagement cycles in swiping apps from simple behavioral rules [8]. Users, for their part, hold rich folk theories about dating algorithms and counter-strategize against perceived platform interests [1]. This literature optimizes who is shown to whom, taking the communication medium as given; agent mediation makes the medium itself a market variable, adding a second reciprocity layer: both sides must accept not only each other but the mode of contact. Our send/receive distinction formalizes this second layer, and our routing counterfactual (§7) is a reciprocal-recommendation intervention on it. AI-mediated communication. Hancock et al. [14] define AImediated communication and its research agenda. Receiver-side penalties are well documented: profiles suspected of AI authorship are trusted less [18], suspicion of LLM use falls unevenly across writers [19], and smart replies measurably alter language and relationships [17]. Closest to our setting, Ante [2] interviews both LLM-assisted senders and receivers in online dating, finding an “authenticity paradox” among senders and “digital betrayal” among receivers. These qualitative asymmetries motivate our contribution: measuring send- and receive-willingness jointly, within person, on one latent scale, at population size.

Method. 3 The Receptivity Audit Agent-mediated communication features pose a question that arrives before any system exists to log behavior from: do both sides of the market for this feature exist? Answering it post-launch is expensive: the failure mode is deployed agents generating unwanted contact at scale in a trust-sensitive domain. Answering it with off-the-shelf attitude scales is not possible, because existing instruments measure attitudes toward AI in general, score respondents as users only, and collapse responses into sums that cannot locate thresholds (§2). We therefore formalize the procedure this paper instantiates as a general, four-stage receptivity audit:

(1) Instrument. A short battery of ordinal items presenting the concrete feature to its intended population, with items deliberately spanning both roles: scenarios in which the respondent deploys the capability, and scenarios in which the respondent is on the receiving end of someone else’s deployment. Within-person coverage of both roles is the non-negotiable design element; it is what separates role effects from population effects.

(2) Measurement model. A confirmatory multidimensional graded response model with latent regression: items load on send and receive factors, latent traits regress on user covariates, and dimensionality itself is a model-comparison question rather than an assumption. The fitted model yields three things no sum score provides: threshold locations on a common scale (the asymmetry is a statement about thresholds), formal invariance tests licensing pooling across languages or segments, and per-user factor scores with principled uncertainty. (3) Propensities, validated out of sample. Model-implied endorsement propensities computed from the fitted item response functions at each user’s factor scores, such as the probability of endorsing deployment, and the probability of endorsing engagement when contacted, under transparent strict/soft operationalizations of intermediate response categories, treated as scenario bounds rather than calibrated behavior. Before the propensities feed any policy, the score behind them is validated predictively: the target item is held out, the score is re-estimated from the remaining items and covariates under respondent-level cross-validation, and outof-sample discrimination and lift are reported. This is what licenses using the score as a routing signal. (4) Market counterfactuals. A dyadic simulation over the empirical population that converts propensities into market quantities such as interaction volume, engagement per contact, directional imbalances, and prices design levers - eligibility rules (who may deploy), routing rules (who receives), and disclosure or consent mechanics. Because the levers are expressed in the model’s own parameters, each counterfactual is a closed-form computation, not a new experiment.

Two remarks on scope. First, the audit is deliberately parsimonious: every component we use is established machinery [3, 22, 32]; the contribution we claim is the construct and the bridge to market primitives, not a new estimator. Second, nothing above mentions dating. The audit applies wherever one side delegates communication to an agent and the other side must accept the medium: recruiting outreach, sales prospecting, agent-mediated customer contact, professional networking. Sections 4–7 instantiate the four stages on one such market; §8 returns to the general case.

4 Data and Instruments Two survey instruments were fielded to active users of a large dating platform, Fledge.Love as voluntary, self-administered online questionnaires. Respondents were recruited through an in-app prompt shown to active users. The agents instrument (B) was fielded first, from November 12 to December 15, 2025, as parallel Russianand English-language forms collected concurrently. The generativefeatures instrument (A) followed from March 22 to April 1, 2026, in Russian. The two samples are not linked at the respondent level. Instrument A: generative features (N = 2,894).

Discussion. Direction matters. Because women are less receptive on both dimensions, the market is directionally imbalanced: male-deployed agents contacting women achieve 8.4% engagement per contact (strict) against 13.4% for female-deployed agents contacting men. The heavier prospective traffic (male-to-female) faces the colder reception: the same structural pattern that plagues human-authored first messages, reproduced and potentially amplified by automation. Reciprocity gate. Requiring symmetric willingness, when a user may deploy an agent only if their own soft engagement propensity is at least 0.5, excludes 65% of likely deployers and halves interaction volume (0.044 →0.020 strict). This prices the free-rider structure directly: about a quarter of users want to send agents but will not receive them, and a reciprocity norm removes exactly this traffic. Receptivity-aware routing. Routing agent contacts only to receivers in the top receptivity quartile raises engagement per contact from 11.6% to 39.4% (strict; 25.6% →64.0% soft), or a 3.4× quality gain, at the cost of restricting coverage to 25% of receivers. This is not an artifact of scoring receivers with the engagement item itself: with Y4 held out and receptivity re-estimated under cross-validation, the top-quartile lift is 3.1× (§6.7). Figure 2 traces the full quality–coverage frontier, which a platform can tune like any exposure policy. Routing on receive receptivity is a matchinglayer intervention invisible to senders; it converts the measurement model into a pre-deployment policy input. Two caveats bound these numbers. The simulation assumes random dyad formation, ignoring matching structure and homophily in receptivity. It treats propensities as static, whereas receivers plausibly update after good or bad agent encounters, reflecting that an equilibrium version is future work. The qualitative conclusions: receive-side scarcity, directional imbalance, and the volume/quality lever structure, do not depend on these simplifications.

8 Design Implications for Agentic Recommenders Receive-side consent is a first-class primitive. Current agent designs treat deployment as the consent event: the sender opts in. Our results locate the scarce resource on the other side. A receivingconsent surface, or an explicit setting for whether, and from whom, agent-mediated contact is acceptable, is not a compliance nicety but the mechanism that determines market quality; without it, three quarters of agent contacts land on users who did not want them (Table 7). Disclosure follows directly: undisclosed agents convert unwilling receivers into deceived ones, the “digital betrayal” documented qualitatively by [2], and our receive-side estimates should be read as an upper bound on tolerance for undisclosed contact. Receptivity is a routable signal. Because receive receptivity is measurable from a handful of items and predicts held-out engagement responses with AUC 0.88 even when the engagement item itself is excluded from scoring (§6.7), it can enter the matching layer like any exposure feature (in production, behavioral proxies would play the items’ role). The quality–coverage frontier (Figure 2) is the tuning knob: routing agent traffic to the top receptivity quartile more than triples per-contact engagement while leaving human-authored contact untouched for everyone else. This is the cheapest intervention in our counterfactual set and requires no visible product change for senders. Reciprocity is a values choice with a known price. A symmetricwillingness rule (deploy only if you would engage) halves agentinteraction volume and excludes two thirds of would-be deployers - almost exactly the asymmetric-delegator segment.

Conclusion. Agent-mediated matching is being built sender-first, but it will succeed or fail receiver-first. Measuring both sides of delegation within person, we find distinct constructs separated by a robust 0.7 SD asymmetry, a quarter of users wanting to send agent traffic they would not fully accept, and, under a random-pairing counterfactual, only 4–13% of directed dyads combining deployment with engagement unless receive-side consent becomes a design primitive. The measurement model that produces these numbers is cheap with only seven ordinal items, and its pre-deployment outputs are directly usable, from a validated routing signal to the quantified price of a reciprocity norm. As LLM agents enter interpersonal products, we offer this as a template for measuring readiness and a caution against mistaking enthusiasm for delegation as enthusiasm for its receipt.

Limitations. Our estimates are stated preferences elicited from concept mockups, not behavior under deployed agents; the technology-acceptance literature suggests stated and revealed adoption correlate but diverge, and the direction of divergence for receiving agents is unknown. Both samples are self-selected respondents from a single platform, and the EN subsample is small (n = 232) and self-selected, so language is confounded with population composition. Our crosslingual claims are therefore limited to the partial-invariance form we report, and the cultural interpretation of the receive-side DIF remains a hypothesis. Ordinal codings required judgment calls for response options without a canonical order; the sensitivity analyses show the headline quantities are robust to the defensible alternatives, but cannot rule out orderings we did not consider. The two instruments were fielded to non-linked respondents, so the passiveversus-active contrast is between-sample. The market simulation assumes random dyad formation and static propensities, abstracting from matching structure, homophily in receptivity, and the plausible dynamics in which receivers update after agent encounters. Finally, attitudes toward agent mediation are likely nonstationary as the technology normalizes; our estimates are a 2026 snapshot of a moving object.