Delegation Asymmetry in Agentic Recommender Systems: When AI Dates for You, But You Won't Date Their AI
> "Love may be blind, but the dating market is not."
A modern scenario: a dating platform rolls out an AI dating agent that can screen potential matches, initiate conversations on your behalf, keep chatting while you're offline, and even arrange in-person meetings. Sounds great—but a critical question lurks: if everyone deploys an AI agent, whose agent talks to whose?
Research from the Lucy Family Institute for Data & Society at the University of Notre Dame uncovers an awkward market truth: people are willing to let AI date for them, but unwilling to date other people's AI. This "delegation asymmetry" could undermine entire agentic dating markets.
From AI-Assisted to AI-Delegated
Dating platforms have evolved through three stages:
1. Algorithmic matching (2000s–2010s) — recommendation algorithms based on profiles and behavior; users fully control communication. 2. AI-assisted (early 2020s) — AI polishes profile copy and suggests openers; users still lead. 3. AI-delegated (2025+) — autonomous agents screen matches, hold full conversations, and arrange meetings; users intervene only at the end.
Dating platforms are a classic two-sided market with cross-side network effects: one side's value depends on the other side's adoption. If nobody wants to receive agent messages, deploying agents becomes pointless—like a party where every guest sends a social proxy but no one will talk to anyone else's proxy.
Measuring Two-Sided Receptivity
The paper introduces two constructs measured jointly within the same individuals:
- Send receptivity: willingness to let an AI agent communicate on your behalf.
- Receive receptivity: willingness to receive communication from others' agents.
- Dimension 1 (Send): Y1, Y2, Y3, Y7
- Dimension 2 (Receive): Y4, Y5, Y6
- Latent traits follow a bivariate normal with correlation ρ
- Covariates: gender, age, platform tenure, perceived match volume, match-to-conversation conversion, frustration with stalled matches, language
- Conduct two-sided receptivity audits before launching agent features, not just one-sided adoption metrics.
- Implement receptivity-aware routing to avoid pushing agent messages to agent-rejecting users.
- Deploy progressively, from low-intrusion features to autonomous conversation.
- Provide transparency and opt-out mechanisms.
- 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*.
Data comes from two large surveys:
| Survey | Period | Language | Sample | Topic | |---|---|---|---|---| | A | Mar–Apr 2026 | Russian | N=2,894 | Generative profile features | | B | Nov–Dec 2025 | Russian/English | N=2,617 (2,499 complete) | Autonomous conversation agents |
Survey B's seven ordinal items deliberately span both roles: sender items (initial emotional reaction, configuring agent tone, agent acting while absent, overall value) and receiver items (reacting to an agent's reply, agents chatting before humans join, human-agent co-presence in group chats).
Statistical Model
A two-dimensional graded response model (GRM) with latent regression was fitted:
Model comparison favors two dimensions: ΔBIC = 51.8; LRT = 114.5 (8 df, p < 10⁻¹⁵). Despite send and receive being highly correlated (ρ = 0.92), they are clearly not one construct.
Core Finding: Delegation Asymmetry
| Metric | Value | Meaning | |---|---|---| | Deployment threshold | θ = −0.38 | Median threshold to deploy your own agent | | Participation threshold | θ = +0.32 | Threshold to engage with the other side's agent | | Full-engagement threshold | θ = +1.39 | Fully willing to interact with the other's agent | | Asymmetry gap | 0.71 SD (95% CI [0.65, 0.77]) | Quantified delegation asymmetry |
Implied average propensities: deployment 0.38 (strict) / 0.50 (lenient); participation only 0.12 / 0.26. At the individual level, 40.7% of respondents rated send strictly above receive; only 2.1% the reverse—a 19:1 ratio.
Latent class analysis identifies 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 the mean, lukewarm | | Asymmetric delegators | ~26% | 98% willing to deploy their own agent; only 10% willing to fully engage with a counterpart's agent |
Market Counterfactuals
Under random pairing, baseline market viability is bleak:
| Scenario | Feasible pair rate (strict/lenient) | Engagement per contact | |---|---|---| | Baseline | 4.4% / 12.8% | 11.6% / 25.6% | | Male→female contact | 3.5% / 11.4% | 8.9% / 22.4% | | Female→male contact | 6.4% / 15.5% | 17.0% / 30.9% |
Women show lower receptivity on both sides (send β=−0.38, receive β=−0.30). Other predictors: higher match-to-conversation conversion lowers receptivity (β≈−0.17 both sides), while frustration with stalled matches raises it (β≈+0.14)—suggesting a potential adverse-selection spiral where desperate users adopt agents first, further deterring others.
Design Levers
1. Reciprocity requirement (only agents for those willing to receive): halves interaction volume—severe liquidity cost. 2. Receptivity-aware routing: triples per-contact engagement; AUC = 0.88 out-of-sample. Routing by receptivity quartile lifts full-engagement rates from 0.0% → 2.4% → 8.2% → 37.9%. 3. Transparency and choice: disclosing agent use, opt-in/opt-out (discussed, not quantified).
Broader Implications
The finding generalizes to all two-sided agentic systems: customers accept AI handling their service requests but resist being served by another AI; candidates accept AI resume help but resist AI interviewers. The principle: an agentic system's viability depends on the minimum—not the average—of two-sided acceptance.
Design recommendations for platforms:
The paper closes on a philosophical question: if two people "fall in love" entirely through their AI agents, is the relationship real? Consequentialism says the outcome matters; Kantian ethics asks whether treating someone as an algorithm output rather than an end-in-themselves is moral. The data can't answer that—but it does show that supply and demand in agentic dating are severely imbalanced, and until platforms solve this, even the best technology is built on sand.
Key Numbers at a Glance
| Metric | Value | |---|---| | Core sample | N=2,617 | | Latent correlation ρ | 0.92 | | ΔBIC (2D vs 1D) | 51.8 | | Deployment / participation thresholds | −0.38 / +0.32 | | Asymmetry gap | 0.71 SD | | Asymmetric delegators | ~26% | | Baseline feasible pair rate | 4.4%–12.8% | | Routing engagement lift | 3× | | Routing AUC | 0.88 |