How does AI calculate compatibility? Where the numbers come from

August 11, 2026 · 17 min read

In LoveScore the AI does not calculate compatibility. A deterministic algorithm does: 47 answers from each partner become nine dimension scores per person, those are combined into pair scores, and the pair scores are summed into one Love Score. Only when every number is final does a language model get involved, and its entire job is to write the prose that explains those numbers. It cannot raise a score, cannot invent one, and never sees a single line of your answers. That split is what makes the result reproducible: the same answers always produce the same number, down to the point.

Whether an AI compatibility test can be trusted depends almost entirely on which job the AI has been given — and three quite different products ship under those two words. Below, first the three designs and how to tell them apart, then the whole path of ours: from a tapped answer to a sentence in the report, where the formula works, where the model takes over, what it is forbidden to write, and what enforces the ban.

“AI compatibility test” describes three different things

The phrase is doing a lot of work on this market, and it hides a distinction that decides everything about whether the result means anything.

1. AI as a label on top of something else

The most common arrangement. The interface promises a neural network; the calculation underneath runs on names, birth dates or star signs. The giveaway is that the input never contains a single fact about your relationship — how you argue and make up, who carries the household, whether you have agreed about children — because it cannot: those facts did not exist on the day either of you was born. We took that substitution apart in detail in a piece on AI compatibility by birth date, and the wider comparison sits in compatibility test versus astrology.

2. AI as the thing that produces the number

Newer, and more interesting to argue with. Here a model really is involved: you describe your relationship to a chatbot, or a service feeds both partners' answers into a model and asks it for a percentage. The output looks like a measurement. It is not one, and the reason is structural rather than a matter of the model being good or bad.

A language model produces text by sampling, token by token, from a distribution over what could plausibly come next. A number generated that way is the most plausible-looking string in context — not the result of a rule applied to your data. You can check this in about two minutes without taking anyone's word for it. Open two fresh sessions, paste the same description of the same couple, ask for a compatibility percentage. The two numbers will differ. Then rewrite the description in warmer words without changing a single fact, and watch the number climb. An instrument that responds to adjectives is measuring the adjectives.

There is a quieter problem underneath. A model asked for a score will essentially never refuse, because refusing is not what the next-token objective rewards. It will produce a confident number for a couple it knows nothing about, for contradictory input, for three sentences typed in a hurry. Nothing in the mechanism distinguishes “I have enough to go on” from “I have almost nothing”.

3. AI as the writer on top of a fixed measurement

The third design keeps the two jobs apart on purpose. Scoring is done by a rule written in advance, published, and identical for everyone; the model is handed the finished numbers and asked only to turn them into something a person will actually read. Text is exactly the job a language model is good at, and scoring is exactly the job it cannot do. Everything below describes how that is built here.

Four steps from a tapped answer to a score

The whole calculation is four transformations in a row. No neural network takes part in any of them, none of them contains randomness, and none of them depends on who opened the report or when.

Step 1. The 47 screens are 45 scored statements and 2 checks

Nine dimensions, five statements each, plus two attention checks that carry a known correct action and never enter any score. Each statement is rated on a five-point agreement scale: the ends are labeled “strongly disagree” and “strongly agree”, the points in between are the numbers, and the whole questionnaire can be answered from the 1–5 keys without touching the screen.

Of the 45 scored statements, 27 are worded positively and 18 are reversed: agreeing with a reversed item lowers the score rather than raising it. That is two safeguards at once. If someone answers without reading, the positive and reversed items of the same dimension start contradicting each other, and the contradiction is visible in the calculation. And no one can lift a score by agreeing with everything, because a habit of agreeing pushes the two kinds of item in opposite directions.

Step 2. Five items become one score from 0 to 100

First the answers are brought into one direction: positive items are taken as given, reversed ones are mirrored, since agreeing with a reversed statement means the same thing as disagreeing with a positive one. Then the five values of a dimension are averaged and mapped onto a 0–100 scale. The mapping is linear and contains no adjustments inside the scale: all five items count equally, none is treated as more important than the rest, and the same shift in answers moves the score by the same amount wherever on the scale it happens. Each partner ends up with nine personal scores.

Missing answers have their own rule. While most of a dimension's items are answered, it is computed from what is there and marked in the calculation as partial. If too little is left, the dimension is not estimated “approximately” — it is dropped entirely, and the weights of the remaining scales are renormalized so that the total stays whole. And if several sides of the profile are left without data, no archetype is issued at all: instead of a portrait stretched over holes, the person is told plainly that there is nothing here to compute from.

Step 3. Two personal scores become one pair score

This is where the algorithm has something no single-person quiz can have: the other half of the data. The rule for combining depends on what is being measured.

For level dimensions — Trust, Empathy, Emotional Maturity, Peace of Mind and Partner Responsiveness — higher is better, and the pair score is built from two quantities at once: the couple average and the score of whichever partner sits lower. The average counts for more, but the lower side counts appreciably too.

The tilt toward the weaker side is deliberate. A plain average would equate a couple where the dimension sits mid-range for both with a couple where one is at the ceiling and the other at the floor — and those are different lives. In the second one, a person is living without that support entirely. One partner's shortfall cannot be covered by the other's strength; it can only be softened.

For match dimensions — Togetherness & Space and Future Vision — there is no such thing as “higher”. A couple who both want children and a couple who have both decided against them are equally aligned. So exactly one quantity enters the calculation: the distance between the partners' answers, and the smaller it is, the higher the score. The level plays no part at all, and when the gap stops meaning a difference in taste and starts meaning mutually exclusive plans, the score falls to zero.

Two more dimensions, Communication and Responsibility, are hybrids: most of the score comes from level, a smaller part from how closely the two pictures agree. All nine scales are walked through with examples in a separate piece on the nine relationship dimensions.

Step 4. Eight pair scores become a Love Score

The Love Score is a weighted sum of eight of the nine pair scores. The ninth, Partner Responsiveness, is left out on purpose: it is the only scale where you answer about your partner rather than yourself, and including it would count one person's view twice in a number that belongs to the couple. The eight weights are fixed in advance and differ: Trust and Communication carry the most, because nothing else works without them, and the two match scales carry the least, since a couple who agree already get the maximum there.

On top of the sum sits a ceiling that cannot be bought out. If one of the critical dimensions — Trust, Communication, Peace of Mind, and for couples past the first few months Future Vision as well — is deeply low, the final score is capped from above: the deeper the shortfall, the lower the ceiling, and strong results elsewhere do not lift it. That is what “non-compensatory” means in practice. Without such a rule a couple with broken trust and excellent household logistics would see a cheerful number and leave reassured.

The last step rescales the number to the two anchors of the instrument itself: a couple who answered “sometimes” to everything lands exactly on 50, and a couple who are at the maximum everywhere and agree everywhere lands on 100. This is a translation into a readable scale, not a comparison with other couples — no population enters the calculation anywhere, we hold no database of average couples, and the report never refers to one. The band label is assigned from the already-rounded number, so the words on screen never disagree with the digits next to them. What the bands mean is unpacked in how to read your Love Score.

Everything else in the report is computed the same way: the couple's supports and friction points, the asymmetry index, the six scenarios, the pair roles. Even the choice of which three scenarios are visible for free is made by a formula, not by a model. Each partner's archetype is a calculation too: it takes that person's own profile, finds which dimensions depart furthest from their own average — a fixed per-scale correction is applied to those departures, evening out the odds of each dimension leading — and picks one of twelve portraits from the leading dimension or from a stable pairing of two, while an even profile with no clear lead gets a twelfth portrait of its own. The partner has no influence on it: a solo report and a paired one have to produce the same archetype. The derivation is described step by step in relationship archetypes.

Where the model actually comes in

The language model receives control at the very end, when the numbers are computed, stored and no longer subject to change. What it gets is a compact block of data: the finished scores and their band labels, dimension names, both archetypes, scenario names and bands, the pair roles, a named list of the supports and friction points that fired, and a handful of observations assembled by the core. All of it is output of the calculation; there is nothing in the request beyond what has already been computed.

What the model does not get: either partner's item-by-item answers. It physically cannot read how someone responded to the question about money or jealousy, because that data is not in the request. This is why the report talks about dimensions and the way they combine rather than about your answer to a particular question — that answer was never in front of it.

Its work is exactly one thing: turn a set of numbers into connected text a person will read. To say that Trust and Future Vision are pulling in different directions in this couple, and to give a domestic scene in which that is visible. It can neither raise a score nor smooth over a divergence: both were settled before it was called.

What the model may not write — and what enforces it

Prohibitions would work badly if they rested on a polite request inside a prompt. So every generated text passes through a set of automatic checks, and text that fails them is never shown to a reader. The checks are these.

• Stop list. Predictions of outcome (“you will break up”, “odds of divorce”), clinical labels and diagnoses — including the fashionable attachment styles — false normativity (“better than most”, “above average”) and ratings of a person as a good or bad partner. The list runs in both languages.
• The number rule. Every significant numeral found in the text is checked against the list of numbers handed to the model as input. Not on the list — the whole text is rejected. The model cannot invent a score or a percentage, and it cannot invent a quantity of any other kind either: “in 15 minutes” fails the check exactly as a made-up score would.
• No assessing the partner. In the personal report, which only one person reads, the partner may appear solely as someone affected by the reader's behavior. Any claim about their traits, deficits, answers or scores is rejected. Separately rejected is any sentence where the partner is mentioned next to a number: even a number that legitimately belongs to the reader reads, in that position, as the partner's score.
• The restatement check. The least obvious one. The reader has already seen their own numbers for free, and text that simply lists them again (“Trust is 84 out of 100, which means…”) adds nothing. So the share of such sentences is measured, along with the count of sentences that connect two different dimensions. A recital of percentages does not pass.
• The specificity check. The mirror image of the previous one: a model boxed in by prohibitions drifts toward smooth prose with no name and no number in it — which is to say, toward a horoscope. If the text names neither the archetype nor the leading strength nor the growth area, and contains no number from the data it was given, the report counts as not done.
• The name-substitution check. Partner names are inserted into the text after generation, mechanically. A language-specific check catches the constructions where that substitution would come out broken. A small thing — and the first thing a reader sees.

A text that fails gets another attempt, and if needed at a different model. If none pass, a deterministic template is printed instead, assembled from the same numbers: drier, but correct. Unverified text is never shown, and that is cheaper than any elegant sentence.

What “reproducible” does and does not cover

In practice it is three concrete properties, and one honest limit.

The same answers give the same score. The calculation contains no call to a model, no random numbers and nothing that depends on the time of day or on how many couples have gone through before you. Recomputing the same two questionnaires yields the same figure.

Who answered first does not matter. The calculation is symmetric by construction: swap the partners and not one pair-level quantity changes. Only the personal sections change places.

The version is pinned. The algorithm version and the questionnaire version are stored alongside the report. If the formulas ever change, it will be visible which version your report was computed under — rather than a vague “we improved something”.

The honest limit: the numbers are reproducible, the wording is not. The text is generated once and stored, so you see the same text every time you open the report. But had it been generated again, the phrasing would have come out different — with the same numbers and the same conclusions. The model is language, not measurement, and no amount of engineering makes sampling deterministic in the way arithmetic is.

The other half of trust: whether the answers are any good

A formula is honest exactly as far as its input is. Alongside the scoring, independent checks run on the quality of the responses themselves: the two attention items with a known correct action; pairs of statements asking the same thing in different words; a comparison of positive and reversed items that catches an agreeing style; response pace; and long runs of identical taps. The specific thresholds are not published — they exist for the system, not for people trying to get around it.

The outcome is one of three states. When the picture is calm, the report is built as usual. Under moderate doubt it is still built but flagged: individual scores are shown as ranges instead of exact figures, and a note appears in the text naming honestly what looked off — the pace, the sameness of the answers, or a divergence between answers to questions that mean nearly the same thing. The note says “this couple”, without pointing at a person: who exactly was rushing is not disclosed. Under clear unreliability the report is not built at all, and the service offers a retake — a report assembled from random taps would be a lie with good typography.

Three questions worth asking any AI compatibility test

The full checklist for telling an honest test from a good-looking label is a separate article — seven signs of an honest compatibility test. Here are only the three points that turn on the word “AI”, and they are worth asking of us as much as of anyone.

• What exactly does the AI produce — the number or the text? If a service cannot say where that line falls, the numbers most likely do not come from your answers. Ask it to name what the model receives as input; a product with the boundary drawn properly can answer that in one sentence.
• Where does the second person's data come from? Compatibility is a quantity about two people. If only one of them answers, what is being measured is one person's opinion, not a couple. And if both answer from the same phone with the other watching, what gets measured is agreement out loud, which is a different thing entirely.
• What is promised as the result? “100% accuracy”, “probability of divorce” and “compatibility with anyone in under a minute” are promises no questionnaire can keep, and why they cannot is worked through in can a test predict a breakup. Accuracy of 100% is unavailable in relationship psychology in principle: what is being measured is the self-report of two people, not the length of a table.

What this method does not do

The limits deserve the same plain words as the mechanism. The test measures self-report: how you describe your relationship, not the relationship itself. It has not been through clinical validation, it does not diagnose, and it does not replace a specialist. It does not predict the future: no score here has been checked against what actually happened to real couples, so the Love Score is an index of how far your answers agree with the test's model, not a forecast. And it does not compare you with other couples, because there is nothing to compare against: we hold no normative sample, and inventing one would be precisely the genre we are distancing ourselves from.

Inside those limits the method does what it promises: it turns 94 answers from two people into a structured map — with numbers that do not depend on the mood of a generator, and text that explains those numbers in human language. The calculation is written out in full, from questionnaire to the rules for combining scores, on the methodology page.

Frequently asked questions

Keep reading

LoveScore is an entertainment and educational service for adults 18+. Articles and reports are not a psychological consultation or a diagnosis, and they do not replace working with a specialist.