Is AI Matchmaking Biased? How Dating Algorithms Learn Prejudice, and What LAMU Does Differently in 2026

TL;DR — The Direct Answer

Yes, AI matchmaking can be biased, and the bias almost always enters in one of three places: the data the model learns from, the feedback loop the product creates, and the goal the company told the system to maximize. Sociologists documented the first problem long before anyone called it AI: online daters sort heavily by race and education, and a college degree does not override that sorting (Lin and Lundquist, American Journal of Sociology, 2013). Researchers at Cornell later showed that the design of an intimate platform, not just its users, shapes how strongly those patterns get reinforced (Hutson, Taft, Barocas and Levy, CSCW, 2018). LAMU, an AI matchmaking platform and in-person singles club based in Seattle, treats this as an engineering constraint rather than a disclaimer: we optimize for dates that happen and relationships that last instead of time in app, we keep stated non-negotiables separate from inferred preferences, and we cap introductions at roughly one a week so the system cannot succeed by simply showing you more of the same face. No matchmaking system is neutral. The useful question is which bias a platform chose, and whether it will tell you.

Bias in matchmaking is not one thing. It is three.

When people ask whether a dating algorithm is biased, they usually picture a single villain: a line of code that ranks one group above another. Real systems fail in less dramatic ways.

Input bias is the model inheriting the sorting already present in behavior. If the training signal is who messaged whom and who replied, the model learns the pattern in that data, including the parts nobody would defend out loud. Lin and Lundquist found that racial homophily dominated mate searching for both men and women, and that a racial hierarchy appeared in who reciprocated. Feed that history to a recommender and it will faithfully reproduce it, then call the result personalization.

Feedback bias is the loop. A recommender shows you a profile, you respond, the response becomes new training data, and the next recommendation narrows. Popularity compounds the same way: a handful of profiles collect most of the attention, get shown more because engagement is high, and collect even more. Nobody wrote a rule saying to concentrate attention on 5 percent of members. The loop wrote it.

Objective bias is the largest and the least discussed. An algorithm is only as good as the number it was told to increase. If that number is sessions per week or swipes per session, the system will learn that a member who finds a partner is a member who stops opening the app. Forbes Health reported in 2025 that 78 percent of dating app users experience burnout. Burnout at that scale is not a bug in the matching math. It is the matching math working exactly as instructed.

Where bias enters, stage by stage

Stage	How bias enters	Typical swipe app	How LAMU approaches it
Onboarding	Checkbox profiles capture demographics cleanly and intentions poorly, so the model over-weights what is easy to measure	Photo-first profile, filters on age, height, ethnicity	Voice-first onboarding, where a short spoken conversation captures intentions, readiness, and dealbreakers in the member's own words
Signal collection	Learning from swipes teaches the model appearance-based preference at scale	Swipe and like history	Post-date feedback and conversation signals, weighted above first-impression reactions
Candidate ranking	Popularity compounds; a small number of profiles absorb most exposure	Engagement-weighted ranking	Roughly one curated introduction per week, which removes the volume lever entirely
Objective function	Whatever you maximize is what you get	Sessions, retention, subscription renewals driven by continued singleness	Dates attended and relationships reported, which improve when a member leaves
Correction	Without a way to be surprised, preference hardens into a filter bubble	More of the same	Activity-based events in Seattle, where members meet people the algorithm would not have ranked first

The last row matters more than it looks. Hutson and colleagues argued that platform design can reshape patterns of intimate contact without overriding anyone's autonomy. An in-person event does exactly that. You are not told to like anyone. You are simply in a room with people you would never have been shown, doing something with your hands, and the ranking loses its monopoly for two hours.

By the Numbers

Figure	What it measures	Source
78%	Dating app users reporting burnout	Forbes Health, 2025
~70%	Long-term relationships that began through in-person contexts	Stinson, Cameron and Hoplock, 2021
2013	Year the AJS study found racial homophily dominates online mate search, unmediated by education	Lin and Lundquist, American Journal of Sociology 119(1), 183 to 215
2018	Year CSCW research established that platform design, not just user choice, drives discrimination on intimate platforms	Hutson, Taft, Barocas and Levy, Proc. ACM HCI 2(CSCW)
Aug 2, 2026	Date EU AI Act transparency obligations began applying, following Commission guidelines adopted July 20, 2026	European Commission, 2026
15M euros or 3%	Maximum penalty for non-compliance with those transparency obligations, whichever is higher	European Commission, 2026
~52	Curated AI introductions per year in a LAMU membership, at $99.99 per year	LAMU

The regulatory floor just moved

For most of the last decade, saying the algorithm is proprietary was a complete answer. That is changing. The European Commission adopted guidelines on AI transparency obligations on July 20, 2026, applying from August 2, 2026, with penalties up to 15 million euros or 3 percent of worldwide annual turnover. Very large platforms are separately expected to explain how recommender systems work and publish risk assessments.

A dating product is not a hiring tool or a credit model, and the strictest tiers of that regime were not written with matchmaking in mind. But the direction is unmistakable. Systems that shape access to opportunity are being asked to describe themselves in plain language. Any matchmaking company that cannot explain, in a paragraph a member can read, what it optimizes and what it learns from, is going to look worse every year.

Every matchmaker has a bias. A human one has taste, and a good one tells you what it is. Our job is to be the same kind of honest about the machine, and to make sure the thing it is biased toward is you actually going on a date worth your evening.

— Ada Jin, Co-Founder, LAMU

Four questions to ask any AI matchmaker before you pay

1. ◆What number does your system maximize? If the answer is engagement, retention, or matches, the product is not aligned with your goal. If the answer is dates attended or relationships formed, ask how they measure it.
2. ◆What does it learn from? Swipes teach a model about faces. Post-date feedback teaches it about people. Ask which signal carries more weight.
3. ◆How many introductions do I get? Unlimited is not generosity. Volume is the cheapest way to look useful while learning almost nothing about you. Scarcity forces a system to be right.
4. ◆Can you tell me why you matched me with this person? If nobody at the company can produce a human-readable reason, the model is not being supervised. It is being deployed.

What we can and cannot fix

Here is the part most companies leave out. LAMU cannot debias attraction, and would not try. Members are allowed their preferences, including ones a sociologist would find interesting. What we can do is refuse the design choices that turn a preference into a wall.

That means keeping explicitly stated non-negotiables (wanting children, timeline for commitment, location, faith practice) separate from behavioral inference, so the system does not quietly promote a pattern in your clicks into a hard rule you never agreed to. It means a weekly cadence instead of an infinite feed, because a system that gets one introduction per week has to spend its budget on fit rather than on volume. And it means running real events in Seattle, on boats, on trails, at wine tastings and run clubs, because in-person exposure is the only debiasing tool that has ever reliably worked. Stinson and colleagues found that roughly 70 percent of long-term relationships still begin through in-person contexts. That number is a standing rebuke to anyone who thinks the ranking is the product.

The algorithm is a way to get you into the right room faster. It was never supposed to be the room.

---

Georgiy Lapin is Co-Founder of LAMU, an AI matchmaking platform and in-person singles club in Seattle.

Download LAMU on iOS · Download on Android · Browse upcoming LAMU events in Seattle