W
Loading
Loading
Responsible AI
Wrati uses automated systems to help keep people safe and to suggest better connections. Those systems are useful. They are also fallible — and this page is meant to be honest about both.
Automated screening reduces harm; it does not remove it. A verified profile is not a character reference, a passed content check is not a promise about a person, and no filter catches everything. Please keep taking the same precautions you would meeting anyone new.
Each of these runs in Wrati today. The limitation beside each one is not a disclaimer — it is the reason human review exists.
Images shared in chat and uploaded to profiles are screened automatically for nudity and sexual content before they reach anyone.
Limitation: Automated image classification produces both false positives and false negatives. It is a filter, not a judgement.
Optical character recognition extracts text from uploaded images, including screenshots, so that content rules apply to words in a picture as well as words in a message.
Limitation: OCR accuracy varies with image quality, language and handwriting.
Message text is checked against content rules — including profanity and attempts to move people off-platform through shared contact details — before delivery.
Limitation: Language is contextual. Rules calibrated to catch harm will sometimes flag something harmless.
Perceptual hashing recognises when the same or a near-identical image reappears, which is how bulk and recycled imagery is spotted.
Limitation: Similarity is not identity. A hash match is a signal for review, not proof of wrongdoing.
Members can verify that they are the person in their photos. Verification produces a signal on the profile; it does not make a claim about character.
Limitation: Verification confirms a likeness, not an identity document, and not a person's intentions.
Text embeddings and vector similarity are used to suggest people with genuinely related interests and intentions, rather than ranking purely on activity.
Limitation: Recommendation is a suggestion. It should never be read as an endorsement of anyone.
Translation support so people who do not share a first language can still talk to each other.
Limitation: Machine translation loses nuance, and tone is the first thing to go.
Reports and safety events are reviewed by people through internal trust-and-safety tooling. Automated systems flag and filter; they do not have the final word on enforcement against an account.
There is currently no self-service way to formally appeal a moderation decision. Until there is, please use the grievance route below — it reaches a person. Building a proper appeals flow is planned work, and we would rather say that than describe a process that does not exist.
Systematic bias auditing and formal, documented model evaluation are not yet in place. Models are updated deliberately rather than continuously, and this page will be updated as that programme develops.
Content is processed for the purpose of safety and matching, not to build advertising profiles. Media shared in conversations is stored privately and reachable only through short-lived, access-checked links. How that works is described on the security page.
If automated screening or a moderation decision has affected your account and you believe it was wrong, raise it through the grievance process. A person will look at it. A dedicated appeals flow is planned, and this page will be updated when it exists.
Related