The Science Behind
AI-Powered Compatibility Matching
Compatibility matching has moved from questionnaires to models that read how people actually communicate. Here is how vector embeddings, personality inference, and similarity scoring fit together, and what they still cannot see.
Dating apps spent a decade matching on self-reported preferences, and research kept finding that what people say they want in a partner is a poor predictor of who they are happy with. The newer approach reads behavior instead: how someone writes, how they respond to conflict, what they pay attention to. This post walks through the machinery behind that shift and is honest about its limits.
Key takeaways
- Modern matching converts communication and personality signals into vector embeddings, then measures how close two people sit in that space.
- Deep learning models can infer personality traits from text with accuracy that now rivals or exceeds human raters.
- Multiple conversation samples matter: one chat shows how you talk to one person, not how you communicate in general.
- Chemistry, culture, and personal growth remain outside what any model can score, so treat compatibility numbers as an input.
- MosaicChats scores compatibility within a conversation you upload; it does not match you with strangers.
Why matching moved from preferences to behavior
Traditional apps relied on demographics, stated preferences, and short questionnaires. Machine-learning systems instead analyze real communication and the personality traits it reveals. A 2024 paper in Scientific Reports showed deep learning models predicting agreeableness from word embeddings of written text, one of several results establishing that language is a usable personality signal.
What vector embeddings do
An embedding converts something complex, such as a person's writing style, into a long list of numbers. Two people with similar styles end up with similar lists, so compatibility becomes a distance calculation instead of a rulebook.
How embeddings map compatibility
Personality traits, communication habits, and behavioral signals are encoded as high-dimensional vectors. OpenAI's embeddings documentation describes the property that makes this work: semantically similar inputs cluster together in the vector space.
Semantic similarity
People with compatible communication styles sit closer together than people who merely share interests.
Many dimensions
Current embedding models use up to 3,072 dimensions, enough to encode nuance that a ten-question quiz cannot.
The payoff is capturing relationships people do not consciously recognize. Two people can look nothing alike on a profile and still share deep similarities in emotional processing or conflict style that only appear once their language is compared mathematically.
How well AI reads personality from text
Personality is the foundation of most compatibility models, and the evidence that machines can read it from text is now strong. A survey in Artificial Intelligence Review reports models predicting MBTI traits from conversation with up to 90% accuracy, and a 2025 study in Communications Psychology found AI models outperforming human experts, including personality researchers, at predicting correlations between personality items.
Communication style
How each person gives and receives information, makes decisions, and processes emotion. An introvert paired with a patient extravert often reports more satisfaction than two people competing for processing time.
Cognitive complementarity
Whether thinking, feeling, sensing, and intuition preferences balance each other, so one person's strengths cover the other's blind spots.
Conflict style
How each person behaves under disagreement and stress. Matching styles that are compatible, or productively different, matters more than matching hobbies.
Growth trajectory
Some systems try to model how traits shift over time, so that two people stay compatible through life changes. This is the least mature part of the field.
Why one conversation is not enough
A single chat shows how you talk to one person. Systems that match strangers need several samples, across different relationships and periods, before they can separate your core style from the way you adapt to a particular friend or partner. Multiple conversations let a model:
- Capture variability: how you write to family, friends, and partners reveals different facets of the same personality.
- Find the constant: traits that persist across every thread are the ones worth matching on.
- Sample emotional range: stress, joy, conflict, and intimacy rarely all appear in one chat.
- Check authenticity: consistency across contexts separates genuine traits from performance.
MosaicChats' compatibility score takes the narrower, more grounded path: it scores the one relationship in the conversation you upload, reporting where your communication styles align and where they collide. It is a mirror for an existing relationship, not a search for a new one.
What a matching pipeline looks like
Five stages
1. Language processing: NLP models extract meaning, emotional context, and communication patterns from the text.
2. Feature extraction: hundreds of signals are pulled out, from response timing to word choice and how emotion is expressed.
3. Embedding: those signals are encoded as vectors with a model such as OpenAI's text-embedding-3-large.
4. Similarity search: nearest-neighbor algorithms find people whose vectors indicate compatible styles.
5. Scoring and explanation: a model turns distances into a score and, in better systems, a reason you can read.
What works and what does not
Where it helps
- Measures behavior instead of self-reported preferences
- Finds pattern similarities people do not notice in themselves
- Assesses many dimensions at once rather than a checklist
- Improves as outcome data accumulates
Where it fails
- Needs substantial text before results are reliable
- Struggles with personality change over time
- Under-represents cultural context and non-Western norms
- Inherits bias from whatever data it was trained on
- Cannot score chemistry or physical attraction
Research is actively working on the bias and cultural problems; our piece on bias in relationship technology covers what has been found so far.
What ethical matching requires
- Transparency: you can see why two people were scored as compatible.
- Representative data: training sets that include diverse populations, orientations, and cultures.
- Control and consent: you decide what is analyzed and can withdraw at any time.
- Privacy: encryption, minimal retention, and no raw conversation exposure.
- Bias monitoring: ongoing checks for discriminatory patterns, with corrections when they appear.
The right expectation
AI matching can identify people who communicate in compatible ways with more precision than any questionnaire. It cannot tell you whether you will like someone across a table. Used as one input alongside your own judgment, it removes some of the guesswork; used as an oracle, it will disappoint you.
Score the relationship you already have
Upload a conversation and MosaicChats reports a compatibility score for the two of you, with the communication-style and personality signals behind it.
Analyze your conversationFrequently asked questions
Can AI predict whether two people are compatible?
Partly. Models can measure communication-style similarity, personality signals, and how two people handle disagreement, and those correlate with satisfaction. They cannot measure chemistry, timing, or shared values that never come up in text. Treat a compatibility score as one input, not a forecast.
What is a vector embedding, in plain terms?
It is a list of numbers that summarizes something, such as a person's writing style, so that similar things end up with similar numbers. Comparing two people's embeddings gives a similarity score without anyone reading the raw text.
Does MosaicChats match me with other people?
No. MosaicChats scores compatibility inside a conversation you already have: you upload a chat between you and one other person, and the analysis reports how well your communication styles fit and where they clash.
Related reading
References & Sources
- "Using deep learning and word embeddings for predicting human agreeableness behavior." Scientific Reports, 2024.Source
- "Embeddings." OpenAI Platform Documentation.Source
- "Machine and deep learning for personality traits detection: a comprehensive survey." Artificial Intelligence Review, 2025.Source
- "AI can outperform humans in predicting correlations between personality items." Communications Psychology, 2025.Source