I Generated My AI Soulmate 50 Times — What I Learned About My Own Type
One Enutrof team member ran fifty consecutive soulmate generations on himself and tracked the patterns. The exercise turned out to reveal less about a future partner and more about his own unspoken preferences.
Why I ran this experiment on myself The soulmate feature is one of the most popular tools on Enutrof, and also one I felt least comfortable defending. Users send us screenshots saying "this is uncanny, this is exactly my type" — and I, on the other side of the product, kept wondering: is this real signal, or is the system just generating attractive faces and users projecting onto them?
So I (I'll keep this in first person — this is one team member's exercise, not a company claim) did the obvious thing. I gave the system my own selfie and my own Saju, regenerated the soulmate fifty times, saved every output, and then looked at the strip at the end.
The setup - Same input selfie. Same birth date and time. - Each run regenerated from scratch — no seed reuse. - Fifty outputs over four sessions across a week. - I did not rate the faces while generating. The rating happened on day eight, looking at the whole grid at once.
The grid I ended up with is the hero image at the top of this article. Fifty faces in a tight mosaic — different ethnicities, ages, lighting, hair, expressions. At first glance: chaos.
The patterns I noticed only at the end Looking at the full grid, three patterns were too consistent to ignore.
**Pattern 1 — eyes.** Roughly forty of fifty faces had a slight downturn at the outer corner of the eye. It was not the most fashionable eye shape; it was a specific shape. I had never articulated this preference out loud, but my last three serious relationships, when I looked back at photographs, had it.
**Pattern 2 — mouth at rest.** The mouth was almost always closed, slightly upturned at one corner, not the other. A composed mouth, not a beaming one. Again, not something I had ever said I wanted; again, recognisably the people I had loved.
**Pattern 3 — the face was usually older than me by a year or two.** I am in my mid-thirties; the system kept producing faces that read as late thirties. I had to sit with that one. It was true of my partners. I had never thought of myself as someone with a "type" in that direction, but the evidence in the strip was clear.
What I think actually happened The technical answer is straightforward. The system takes my facial features, identifies what I tend to gravitate toward (this is what computer vision can detect from how my own features are arranged and from anonymous behavioural priors), reads my Saju for the elemental balance I am missing, and then constrains an image-generation prompt to produce a face that statistically completes both signals. With one run, you get one possible draw from that distribution. With fifty, you get the **shape of the distribution** — which is much more honest information.
The shape of the distribution is, near as I can tell, an honest portrait of my own preferences. Not of any real person. Not of a future spouse. Of the type my eye is already trained to find.
This is exactly what visualisation researchers like Markus and Nurius (1986) described as **possible selves** — people make clearer decisions about who they want when they can see a concrete image, not just words. "I think I prefer thoughtful people" is fog. A face with the specific eye shape I keep being drawn to is data.
What it did not tell me It did not tell me whom to marry. It did not predict whether the person I am currently dating fits — that is a separate, harder question that no image generator can answer for you. It did not even tell me my preferences were *right* or *healthy*; it just made them visible.
I want to underline that last point because it matters. Seeing your type in fifty variations is uncomfortable. Some of mine I would prefer to outgrow. The exercise does not bless your patterns. It just refuses to let you keep pretending they are not there.
What I would tell a user doing this for the first time If you only generate once and the face surprises you, **the most likely explanation is that one draw is noisy** — not that the system "doesn't know you" and not that the system has revealed a prophecy. Run it three or five times. Look at the strip. The signal lives in the repetition, the same way a Saju daily reading's signal lives in the month, not the day.
And then ask yourself the harder question. Not "is this my future spouse?" — which is unanswerable — but "what do these faces have in common, and is that a thing about me I am comfortable with?" That question, in my experience, is the actual gift the feature can give you.
What we have changed inside the product as a result This exercise pushed us to make two changes in the Enutrof soulmate flow: 1. Encourage users to regenerate at least three times before rating, with a soft tooltip. 2. Add a short reflection prompt under the result — *what do you notice about this face, and is that familiar?* — that nudges the reader toward self-examination rather than prediction.
Neither is a feature you would call exciting. Both came out of one team member spending a week looking at a wall of faces that were, in the end, a long letter to himself.
Important honest notes - This is one person's experience and one product team's reflection. It is not a study. - The system's distribution is shaped by the data it was trained on, which contains its own cultural biases. We are working on this; we are not done with it. - Soulmate generation is a self-reflection tool. It is not a prediction of a real person, and treating it as one will create more disappointment than insight.
References - Markus, H., & Nurius, P. (1986). Possible selves. *American Psychologist*, 41(9), 954–969. https://psycnet.apa.org/record/1987-09454-001 - Hassin, R., & Trope, Y. (2000). Facing faces: studies on the cognitive aspects of physiognomy. *Journal of Personality and Social Psychology*, 78(5), 837–852. https://pubmed.ncbi.nlm.nih.gov/10821193/ - Wang, Y., & Kosinski, M. (2018). Modern physiognomy: predicting personality and intelligence from the face. *Science China Information Sciences*, 61, 058105. https://link.springer.com/article/10.1007/s11432-016-9174-0 - Sprecher, S., & Felmlee, D. (2008). Insider perspectives on attraction. In Sprecher et al. (Eds.), *Handbook of Relationship Initiation*. Psychology Press.
> Personal experiment by one Enutrof team member. Drafted by the Enutrof editorial team with references above.