AI Transparency Statement
A note before this page
Most platforms that use AI either say nothing or say a generic sentence. We are going to do more than that. The reason is simple: you are sharing real things about yourself with a system you cannot see. You deserve to know how it works, what it can do, what it cannot do, and what we have decided it will never do.
This page is plain language. The corresponding internal technical document is published in our research framework at wouch.app/research for anyone who wants the underlying detail.
What the engine actually does
Wouch holds an internal model of what we have observed in your responses. The model represents five things:
- How safe closeness has tended to feel to you (we call this internally core security).
- How loudly your system reacts when a connection is uncertain (relational reactivity).
- How much you have learned to manage by stepping back from closeness (avoidant distancing).
- How exposed you are emotionally right now (current vulnerability).
- How much noise from your environment is currently shaping your responses (environmental noise).
These are statistical estimates, not statements about who you are. The model says: what we have observed in this person, so far, tends to look like this. It is updated as you share more - through the assessment, through modules, through reassessment if you choose to do it.
The model is private. We do not show you these numbers because we believe scoring a person on their inner life is a category mistake, not a missing feature.
What the engine is used for
Three things.
Recommendation. Which module the engine thinks would be useful at this stage. You can choose another or none.
Pacing. When matching becomes available. The engine waits, sometimes longer than you would prefer, because rushed connection on top of an unsettled foundation tends to repeat the patterns it would have been trying to help.
Safety filtering. A focused set of rules that block specific kinds of pairing from happening - never to be punitive, only where a clinical concern is well-grounded. The number of these rules is small and each one is signed off by a clinician. The engine does not write anything you read. We say more about that below.
Where the engine is from
The engine is built on six bodies of research:
- Attachment theory (Bowlby, Ainsworth, and successors). The foundational map.
- Developmental psychology. How relational templates form, and how they update.
- Interpersonal neurobiology. The biological substrate of connection.
- Gottman observational research. The small moments that predict relationship outcomes.
- Trauma-informed care. Building anything responsibly when some people arrive carrying relational injury.
- Complexity science. Relationships as complex adaptive systems. This is also where our intellectual honesty about prediction comes from, we do not over-promise outcomes we cannot deliver.
Specific instruments embedded in our assessment include the PHQ-4, the UCLA-3 loneliness scale, and the ECR-R attachment instrument. Where we have authored items, the authorship is traceable to clinical literature and reviewed by clinicians.
What we refuse to do
These are public commitments. Each is bound by our governance: changing any of them requires unanimous founder sign-off, public announcement, and a 90-day notice period before any change takes effect.
We will not generate user-facing content with a large language model. Every reflection, insight, module, and message you receive from Wouch was written by a human, reviewed by a clinician, and selected by our engine to be shown to you. We are not anti-AI; we are clear about its current limits. A language model that has not lived through what you are working through is the wrong author for what you read here.
We will not show you a numeric score of yourself. The engine works with numbers internally because that is how Bayesian inference is shaped. We will not surface those numbers to you. Scoring people on their relational interior is the kind of thing that sounds rigorous and is, on closer inspection, dehumanising.
We will not put a clinical label on you. "Anxious-preoccupied," "avoidant-dismissing," "highly sensitive" - these are concepts the research uses, with care, in specific contexts. They are not identities to wear. You are not a type. We will describe patterns we have noticed; we will not tell you who you are.
We will not use urgency, scarcity, or countdown design. No timers. No "your match expires in." No "X people viewed your profile this week." These are the standard tools of engagement design; we have refused them by hand, deliberately, throughout the product.
We will not engineer dependence on the platform. We track our own success by users who graduate from Wouch with what they came for, not by users who spend more time inside. We are aware that this trade-off costs us, and we make it anyway.
We will not sell your data, share it for advertising, or share it with third parties for marketing - ever. Not as a feature limit. As a foundational refusal. If we ever needed to fund the platform in a way that compromised this, we would tell you, ask you, and stop the platform if you said no.
Where automated decisions affect you, and your right to human review
Three places where the engine's outputs change your experience:
- Whether matching is currently available to you. The engine's read of your state, plus your module-completion progress, determines whether matching surfaces or whether the platform asks you to keep working before opening it.
- What module is recommended. Algorithmic recommendation; you may choose another.
- Whether a specific potential pairing is blocked. Where our safety logic indicates a pairing carries a clinical concern, we block the surface. You will not see who or why; the other person's privacy and our duty of care both require this.
For each of these, you have the right to a human review. The path is wouch.app/your-data-rights or in-app Settings → Help → Request review. An approved clinician on our roster - a real person, not an automated system labelled as a person - will read your situation and respond. Average response time is 24–72 hours depending on the case.
How we know if the engine is working
We track outcomes. Each insight, module recommendation, and safety boundary has an expected effect, and we track whether the effect actually shows up. Our internal protocol for this is published in our research framework.
Where outcomes diverge from expectations, the rule is flagged for review. Where divergence is sustained, the rule is changed or retired. This is normal practice for any clinically-adjacent system that takes itself seriously.
We will publish summary outcome statistics annually once the platform has been operating long enough for them to be meaningful - likely in the second year after public launch.
How the engine is governed
Wouch's internal governance has three layers.
The Refusal Register. The public commitments above. The most stable layer; changes require unanimous founder sign-off and public notice.
The Decision Log. Every meaningful change to the engine, the safety rules, the content, or the platform is logged with date, decision-makers, rationale, and reversibility classification. This makes it possible to ask the question "how did we get here?" at any point.
The clinical review. Every change touching content or safety is reviewed by a clinician before deployment. Daily review of platform behaviour by the Clinical Governance Lead. Quarterly outcome aggregates reviewed at the senior level.
The engine is not the work of one engineer alone, and it is not the work of one moment. It is the cumulative work of many decisions made carefully and reviewable in retrospect.
If something feels off
You can write to us. safety@wouch.app for anything that touches safety, privacy@wouch.app for anything that touches your data, you@wouch.app for anything else.
You can also request a clinician review of any decision the engine has made about you. This is the human-in-the-loop that the law requires us to offer and that we believe in independently of the law.