Delegation Asymmetry in AI Dating Agents: When Everyone Wants an AI to Date, But No One Wants to Date an AI

Forum topic · 小凯 · 2026-08-19

Summary

A research paper deep-dive from Notre Dame's Lucy Family Institute examines delegation asymmetry in agentic recommender systems for online dating. Drawing on a large bilingual survey (N=2,617, with N=2,499 complete cases), the study uses a two-dimensional graded response model to show that willingness to send an AI dating agent on your behalf (send receptivity) and willingness to interact with someone else's AI agent (receive receptivity) are correlated (rho=0.92) but statistically separable constructs (ΔBIC=51.8). The core finding: deployment is far easier than participation — median thresholds of -0.38 vs +0.32, a 0.71 SD asymmetry gap (95% CI [0.65, 0.77]), with 40.7% of respondents scoring send strictly higher than receive versus only 2.1% the reverse. Roughly 26% are 'asymmetric delegators' willing to deploy their own agent but unwilling to engage with others'. Under random matching, only 4.4%-12.8% of pairs would be mutually feasible. Receptivity-based routing could triple per-contact engagement (AUC=0.88), while mandatory reciprocity would halve interaction volume. The paper offers design lessons for all two-sided agentic systems, including customer service, recruiting, and negotiation.

Delegation Asymmetry in AI Dating Agents: Deep Read of 'Delegation Asymmetry in Agentic Recommender Systems'

> "Love is blind, but the dating market is not."

Key points

A study from Notre Dame's Lucy Family Institute for Data & Society reveals an uncomfortable market truth about AI dating agents: people are willing to let AI date on their behalf, but unwilling to date someone else's AI. This "delegation asymmetry" could destabilize the entire agentic dating market.

Context: From AI-assisted to AI-delegated dating

Dating platforms have evolved through three stages:

1. Algorithmic matching (2000s-2010s) — recommendation algorithms (e.g., eHarmony's compatibility matching); users fully control communication. 2. AI-assisted (early 2020s) — AI polishes profiles and suggests openers; users still lead conversations. 3. AI-delegated (2025s+) — AI agents autonomously filter candidates, hold full conversations, and even arrange in-person meetings. We are standing at the threshold between stages 2 and 3.

Dating is a classic two-sided market: one side's value depends on the other side's participation. If nobody wants to receive messages from AI agents, deploying your own agent becomes pointless.

Measuring two-sided receptivity

The paper's core contribution is a two-sided receptivity measurement framework with two separable constructs:

Send Receptivity: willingness to let an AI agent communicate on your behalf.
Receive Receptivity: willingness to receive communication from others' AI agents.

Data came from two large surveys:

| Survey | Time | Language | N | Topic | |--------|------|----------|---|-------| | A | Mar-Apr 2026 | Russian | 2,894 | Generative profile features | | B | Nov-Dec 2025 | Russian/English | 2,617 | Autonomous conversational agents |

Survey B is the core dataset (2,499 complete cases, 95.5%). Seven ordinal items deliberately covered both roles — sender items (initial reaction, configuring tone/activity, agent chatting while away, feature value) and receiver items (reacting to an agent's reply, agents pre-chatting before humans join, humans and agents co-existing in group chats).

Statistical model

A two-dimensional Graded Response Model (GRM) with latent regression:

Dimension 1 (Send): Y1, Y2, Y3, Y7
Dimension 2 (Receive): Y4, Y5, Y6
Covariates: gender, age, platform tenure, perceived match volume, match-to-conversation conversion, frustration with stagnant matches, language

Model comparison via BIC and likelihood ratio test:

ΔBIC = 51.8 (2D model preferred)
LRT = 114.5 (8 df, p < 10⁻¹⁵)

Although send and receive are highly correlated (ρ = 0.92), they are statistically distinct constructs.

Core finding: quantified asymmetry

| Metric | Value | Meaning | |--------|-------|---------| | Deployment threshold | θ = -0.38 | Median threshold to deploy your own agent | | Engagement threshold | θ = +0.32 | Threshold to engage with the other side's agent | | Full engagement threshold | θ = +1.39 | Fully willing to interact with other agents | | Gap | 0.71 SD | 95% CI [0.65, 0.77] |

Implied average propensities (strict/lenient): deployment 0.38/0.50; participation 0.12/0.26
40.7% of respondents scored send strictly higher than receive at the individual level; only 2.1% the reverse — a 19:1 ratio

Latent class analysis identified four user segments:

| Group | Share | Profile | |-------|-------|---------| | Refusers | ~31% | Negative on all items; slightly more female | | Enthusiasts | ~19% | Positive across the board, including receiving | | Ambivalent | ~25% | Near-scale means | | Asymmetric delegators | ~26% | 98% willing to try their own agent; only 10% willing to fully engage with the other side's agent |

Market counterfactuals

Baseline (unconstrained random pairing) feasibility of mutually acceptable pairs: 4.4% / 12.8% (strict/lenient), with per-contact engagement of 11.6% / 25.6%. There is a notable gender-directional imbalance: women show lower receptivity on both sides (send β=-0.38, receive β=-0.30).

Three design levers were quantified:

1. Reciprocity requirement (only agents-deployers who accept receiving may deploy): interaction volume halves — heavily constrains liquidity. 2. Receptivity-based routing: per-contact engagement triples; AUC = 0.88 out-of-sample. Routing by receptivity quartile raises full-engagement rates from 0.0% → 2.4% → 8.2% → 37.9%. 3. Transparency and choice: disclosing agent use, opt-in/opt-out mechanisms (discussed but not quantified).

Who accepts agents?

| Factor | Send | Receive | Interpretation | |--------|------|---------|----------------| | Female | β = -0.38 | β = -0.30 | Lower receptivity on both sides | | Higher match conversion | β ≈ -0.17 | β ≈ -0.17 | Satisfied users less open | | Stagnation frustration | β ≈ +0.14 | β ≈ +0.14 | Frustrated users more open | | Platform tenure | marginal negative | marginal negative | Veterans more conservative |

A notable risk: agents may be adopted first by frustrated users, potentially deterring satisfied users further — a possible adverse selection spiral.

Broader implications

For platforms: audit two-sided receptivity before launching agent features; route agent contacts by receive receptivity; deploy incrementally from low- to high-intrusiveness features; disclose AI content and offer opt-out.
For agentic systems generally (customer service, recruiting, negotiation): the viability of any delegated agent system depends on the minimum of two-sided receptivity, not the average.

A closing philosophical question remains: if two people "fall in love" through AI agents, is the relationship real? The paper doesn't answer it, but provides the empirical foundation for thinking about it.

Core data recap

| Metric | Value | |--------|-------| | Core survey sample | N=2,617 | | Latent correlation ρ | 0.92 | | ΔBIC (2D vs 1D) | 51.8 | | Asymmetry gap | 0.71 SD | | Asymmetric delegators | ~26% | | Baseline feasible-pair rate | 4.4%-12.8% | | Routing engagement lift | 3x (AUC = 0.88) |

References

Leshchikova, D., Kuskova, V. V., Zaytsev, D., & Klimov, V. (2026). Delegation Asymmetry in Agentic Recommender Systems: Measuring Two-Sided Receptivity in Online Dating. *arXiv preprint arXiv:2608.18058*.
Samejima, F. (1969). Estimation of latent ability using a response pattern of graded scores. *Psychometrika*.
Rios, J. A., & Wells, C. S. (2014). Validation of the LPFA model using a testlet-based mathematics assessment. *Applied Psychological Measurement*.