The Algorithm Autopsy.

Tinder: a score is not chemistry

Tinder once described an internal Elo-style score. In its simplest form, a swipe from someone rated highly could raise your score more than a swipe from someone rated lower. It was a ranking system for desirability, not a psychological model of compatibility. Tinder says it stopped relying on Elo years ago; its public account now emphasises activity, proximity, preferences and signals such as likes and passes.

The distinction matters. A Tinder algorithm can learn who tends to approve of your profile. It cannot observe the ease of a conversation, the generosity of an interruption, or how two nervous people alter each other's pace in a room. Probability of a right swipe is a useful ranking target because it creates visible activity. It is not the same target as a relationship. The screen collapses both into one gesture, then invites us to confuse the gesture with knowledge.

Read the record: Tinder's earlier Elo score and Tinder's account of its current matching signals.

Hinge: designed to be measured

The Hinge algorithm uses stated preferences, behaviour and feedback to order recommendations. Its Most Compatible feature has been described through the Gale-Shapley matching problem: a way to suggest a pair whose preferences appear mutually promising. “We Met” feedback can add a post-date signal. This is more considered than a pure popularity score, and the stated ambition, “designed to be deleted”, is unusually clear.

But the app still needs measurable actions. Likes sent, matches made, replies and return visits are legible. The quality of a silence on a first date is not. Publicly available evidence does not prove Hinge simply optimises time in app, and it would be careless to say it does. The structural tension is enough: product teams can test engagement quickly, while a good relationship happens slowly and mostly off-screen. The promise points outward. The dashboard points back at the product.

Product record: Hinge's maintained help centre. Current research: decision-making on dating apps.

Bumble: the anxious clock

Bumble's heterosexual matching design originally required women to open the conversation, with a match expiring after 24 hours if no message arrived. Opening Moves has since softened that rule by allowing a pre-set question that a match can answer. The expiry clock remains part of the product's grammar: make a choice now or lose the possibility.

A deadline can prevent a match list becoming a museum of abandoned intentions. It can also turn ordinary life into a signal of rejection. Work ran late. A notification disappeared. Someone needed a day to decide what to say. The Bumble algorithm does not need to understand any of that; the interface simply converts elapsed time into closure. Urgency produces action, but action under threat of expiry is not necessarily attention. The clock makes hesitation look like failure when hesitation may be the most human thing in the exchange.

Read the record: Bumble's explanation of messaging and expiry.

Breeze: move the uncertainty

The Breeze dating app removes the chat phase. If two people like each other, they arrange and pay toward a real date. This is an intelligent answer to a familiar failure: lively matches that become flat message threads, then nothing. Breeze makes intention costly enough to be credible and moves the meeting into the world.

Removing chat solves the problem of endless chat. It does not design the date itself. Two strangers still arrive with little shared context, carrying the evaluative pressure that the interface was meant to escape. The venue, the prompt, the exit and the expectation now do the work the message thread used to do. That may be better. It is not neutral. Breeze demonstrates an important principle for IRL dating London experiments: when a product removes one stage, its emotional labour does not vanish. It moves downstream.

Product record: Breeze's maintained product site. The removed BBC Worklife article is no longer used as evidence.

Haystack: a good shortlist enters a bad room

Haystack dating events use a pre-event questionnaire to identify potentially compatible people before they meet. It is a sensible correction to the arbitrary guest list. In principle, the algorithm narrows a room without asking attendees to perform the narrowing themselves. BODA dating and Someone's Type London also sit within a broader return to curated or in-person alternatives, each making a different wager on selection and context.

But matching can only decide who enters. It cannot decide what the room rewards once they arrive. If everyone is pushed through brief, interview-shaped encounters, potentially compatible people may still meet as compressed versions of themselves. The questionnaire has done its job; the environment can undo it. Compatibility is not a parcel the algorithm delivers. It is a possibility that the format can reveal, distort or leave untouched.

Thursday: half a revolution

Thursday dating London events turn app attention into a time and a place. The app's original constraint was temporal, operating one day a week; its events make the escape from messaging literal. People who are tired of managing a parallel romantic inbox can at least encounter one another in full scale.

Removing the app from the decisive moment solves half the problem. Keeping the mixer format retains the other half. A loud open room still rewards rapid assessment, easy confidence and the ability to interrupt. It gives little structure to the person who becomes interesting through action, disagreement or a second encounter. Thursday proves demand for meeting offline. It does not, by itself, prove that the familiar singles mixer is the best container for that demand. Leaving the screen matters. What people enter next matters just as much. The project is looking for the other half.

Product record: Thursday's maintained product and events site.