🤖 AI

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

This paper reveals a significant "delegation asymmetry" in agentic recommender systems for online dating, where users are far more willing to deploy autonomous agents to initiate conversations than to receive them, and demonstrates that optimizing matchmaking by routing contacts based on receiver receptivity can triple engagement rates compared to random pairing.

Original authors: Daria Leshchikova, Valentina V. Kuskova, Dmitry Zaytsev, Valerii Klimov

Published 2026-08-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Daria Leshchikova, Valentina V. Kuskova, Dmitry Zaytsev, Valerii Klimov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the digital age of finding love, a new kind of helper has arrived. Imagine a user who, instead of typing a message to a potential match themselves, asks a computer program to write it for them. This is the promise of "agentic" systems, where artificial intelligence acts as a personal assistant, handling conversations, screening dates, and even negotiating common ground before the humans ever speak. For years, researchers have studied whether people are willing to use these tools to help themselves. But a critical piece of the puzzle has been missing: the willingness of the person on the other end to talk to a machine. Just because someone wants to send a message via a robot does not mean the recipient wants to receive one. This tension creates a two-sided market, much like a phone network where everyone needs a willing listener, not just a talker. If the technology advances faster than the social acceptance of the other side, the system could collapse into a flood of unwanted automated messages.

A team of researchers set out to measure this delicate balance using data from a major online dating platform. They surveyed thousands of active users, asking them not just if they would use an AI to write their own messages, but also how they would feel if someone else used an AI to write to them. The goal was to see if these two attitudes—sending and receiving—were the same thing, or if they were distinct feelings that could pull in different directions. The study found that they are indeed separate. While most people who are willing to send agent messages are also willing to receive them, the two groups are not identical. More importantly, the threshold for saying "yes" is much lower for sending than for receiving. It is significantly easier to convince a user to let an AI speak for them than it is to convince them to listen to an AI speaking for someone else.

The researchers discovered a stark imbalance in how people view this technology. On a scale of willingness, the point at which a user agrees to deploy their own agent is far lower than the point at which they agree to engage with a counterpart's agent. In practical terms, this means that many more people are ready to send automated messages than are ready to receive them. The study estimated that the number of people willing to deploy an agent is roughly three times higher than the number of people willing to engage with one. This creates a structural problem: if a platform simply lets users turn on their agents, the result would be a massive volume of automated outreach landing on doors that are firmly closed. The researchers calculated that under a random pairing of users, only a small fraction of these directed interactions would actually result in a conversation. In their simulations, only between four and thirteen percent of these potential matches would combine a willing sender with a willing receiver.

This gap is not just a statistical curiosity; it reveals a deep asymmetry in human psychology regarding automation. The study identified a specific group of users, making up about a quarter of the population, who enthusiastically want to use an agent to send messages but would strongly reject receiving one. These users want the efficiency and control of an automated assistant for themselves but view the same technology as intrusive or inauthentic when it comes from someone else. The research also highlighted that this resistance is not uniform across all demographics. Women, for instance, were found to be less receptive to both sending and receiving agent messages than men, a difference that the researchers confirmed was a genuine trait rather than a flaw in how the questions were asked. Furthermore, the desire to use these agents was highest among users who were already frustrated with their dating experience, those whose matches were failing to turn into conversations. This suggests that the technology is being sought most aggressively by those in the most emotionally vulnerable positions.

The implications for how these systems are built are significant. The study tested different design choices to see how they would affect the success of agent-mediated dating. One option was to require reciprocity, meaning a user could only send an agent message if they were also willing to receive one. While this would ensure fairness, the simulation showed it would cut the total volume of interactions in half by excluding two-thirds of the potential senders. A more effective lever was found in routing. If the platform could identify which users were most open to receiving agent messages and direct the automated contacts only to them, the quality of the interaction would improve dramatically. By routing messages to the top twenty-five percent of receptive users, the rate of successful engagement per contact more than tripled. This suggests that the key to making these systems work is not just building better agents, but building better filters that respect the receiver's boundaries.

Ultimately, the research argues that the future of automated matching depends on treating the receiver's consent as a primary design feature, not an afterthought. Currently, most systems focus on the sender's ability to opt in. This study shows that without a mechanism to respect the receiver's preference, the system is destined to fail, generating unwanted contact that erodes trust. The researchers propose that platforms should use these measurements to create a "receptivity score," allowing them to route agent contacts to those who actually want them, while leaving others to interact with humans only. This approach would not only increase the success rate of the interactions but also protect users from the feeling of being deceived by a machine they did not agree to talk to. The findings serve as a warning that enthusiasm for a new technology does not guarantee its success; the market for these tools exists only if both sides of the conversation are willing to participate.

Technical Summary: Delegation Asymmetry in Agentic Recommender Systems

Problem Statement
The emergence of autonomous LLM agents in matching platforms (e.g., online dating) introduces a two-sided market condition that has not been jointly measured: the willingness of a user to delegate their own communication to an agent versus the willingness to receive communication from another user's agent. While platforms are rapidly productizing "AI wingman" tools for sending, the viability of this market depends on a "receiving" side that may be significantly more resistant. Existing literature measures AI acceptance for users or receiver-side penalties separately, often across different samples, failing to capture the within-person asymmetry that determines whether a delegation market can function. Without measuring both sides on a common scale, platforms risk deploying agents that generate automated outreach landing on "closed doors," particularly in trust-sensitive domains like dating.

Methodology
The authors conducted a "Receptivity Audit" using two large-scale surveys of active users on a major dating platform (Fledge.Love), fielded in English and Russian:

Instrument A (N=2,894): Measured interest in passive generative features (e.g., AI profile summaries).
Instrument B (N=2,617): Measured receptivity to autonomous conversational agents. This instrument presented seven ordinal attitudinal items spanning two roles:
- Principal (Send): Scenarios where the respondent configures or deploys an agent (Items Y1, Y2, Y3, Y7).
- Counterpart (Receive): Scenarios where the respondent receives messages from an agent or observes agent-to-agent interaction (Items Y4, Y5, Y6).

Measurement Model
The study employs a two-dimensional Graded Response Model (GRM) with latent regression, estimated via marginal maximum likelihood.

Constructs: The model defines two latent traits: Send Receptivity (willingness to delegate) and Receive Receptivity (willingness to engage with others' agents).
Structure: Items load on their respective factors, and latent traits regress on covariates (gender, age, tenure, match volume, conversion rates, negative affect).
Validation: The model includes formal tests for measurement invariance across languages and differential item functioning (DIF). Crucially, the routing signal derived from the model was validated out-of-sample using a "leave-one-item-out" cross-validation approach to ensure the score predicts engagement without circularity.

Key Results

Distinct Constructs: Send and receive receptivity are highly correlated ( $\rho = 0.92$ ) but statistically distinct constructs ( $\Delta BIC = 51.8$ ). They are separable dimensions rather than a single attitude toward AI.
Delegation Asymmetry: There is a systematic, robust asymmetry in thresholds. Deploying one's own agent requires a latent receptivity of $\theta = -0.38$ , whereas engaging with a counterpart's agent requires $\theta = +0.32$ (a displacement of 0.71 SD). Full engagement with a counterpart's agent is even higher ( $\theta = +1.39$ ). Consequently, the population's propensity to deploy agents is roughly three times higher than their propensity to engage with them.
User Segmentation: Latent class analysis identifies an "asymmetric delegator" segment comprising approximately 25% of the population. These users overwhelmingly want to use an agent for themselves but would not engage with one contacting them.
Correlates of Receptivity: Receptivity is associated with "unmet need." Users reporting negative affect about stalled matches or low conversion from matches to conversations are more receptive. Women are substantially less receptive than men on both dimensions ( $\beta_{send} = -0.38$ , $\beta_{recv} = -0.30$ ), a difference confirmed as a trait difference rather than an artifact of item functioning.
Market Counterfactuals: Under a random-pairing simulation, only 4–13% of directed dyads combine deployment with engagement.
- Reciprocity Cost: Enforcing a reciprocity rule (only allowing deployment if the user would also engage) excludes ~65% of potential deployers and halves interaction volume.
- Routing Gain: Routing agent contacts only to the top 25% of receivers by receptivity triples per-contact engagement (a 3.1× lift), a gain that survives out-of-sample validation (AUC 0.88).

Significance and Claims
The paper claims four primary contributions:

Construct and Measurement Design: It introduces "send receptivity" and "receive receptivity" as distinct, jointly measured constructs on a common latent scale, resolving the confounding of role differences with population differences found in prior separate studies.
Receptivity Audit Methodology: It packages a generalizable, four-stage pipeline (Instrument $\to$ Measurement Model $\to$ Propensity Validation $\to$ Market Counterfactuals) that allows platforms to audit the viability of agent-mediated features before deployment. This treats psychometric parameters as market primitives.
Empirical Findings at Scale: It provides the first population-scale quantification of the delegation asymmetry (0.71 SD) and identifies the structural "asymmetric delegator" segment, demonstrating that the market for agent-mediated interaction is constrained by the receiving side, not the sending side.
Quantified Design Levers: It quantifies the trade-offs for platform design. Specifically, it shows that "receptivity-aware routing" is a high-leverage intervention that can significantly improve engagement quality without requiring sender-side changes, whereas reciprocity norms come with a high volume cost.

Conclusion
The paper concludes that agent-mediated matching is currently being built "sender-first," but its success depends on "receiver-first" design. The authors argue that receive-side consent must be treated as a first-class design primitive. Without explicit mechanisms to manage receive-side receptivity (such as routing or disclosure), the market will suffer from a scarcity of willing receivers, leading to automated outreach that fails to convert and potentially erodes trust. The study offers a measurement template to assess readiness and a caution against mistaking enthusiasm for delegation as enthusiasm for its receipt.

Technical Summary: Delegation Asymmetry in Agentic Recommender Systems

More like this