We tell you what the AI is
AI-generated support is labelled. We do not present it as a person, clinician or emergency responder.
Tranquiliva safety promise
Safety is a system, not a sentence from an AI.
We design Tranquiliva so that no single model decides what happens in a high-risk moment. Clear rules, contextual checks, careful escalation, user controls and accountable review work together.
The short version
What you can expect from us
These commitments guide the AI chat, community spaces, journals, assessments and every safety-related feature.
We use more than one check
High-risk language is screened with deterministic rules, contextual classification and conservative response policies.
We point back to people
When risk may be urgent, normal AI conversation gives way to verified support and steps involving real people.
We admit uncertainty
Detection can miss context or flag safe language. We show limits, make reporting easy and keep reviewing outcomes.
AI limits
Helpful support has a clear boundary.
Tranquiliva AI can help with reflection and general wellbeing information. It cannot replace qualified care or make high-stakes decisions for you.
Designed to help with
What the AI can do
- Listen and respond in supportive, non-judgmental language
- Offer general coping ideas and guided exercises
- Help organise thoughts or prepare questions for a professional
- Explain reviewed educational content in simpler language
- Encourage appropriate support when needs exceed the product
Outside its role
What the AI cannot do
- Diagnose a mental-health or medical condition
- Provide therapy, prescribe treatment or change medication
- Guarantee that information or crisis detection is correct
- Monitor you continuously or know your real-world situation
- Contact emergency services, family or clinicians on your behalf
AI can sound confident and still be wrong.
Report an AI responseCheck important health, legal, financial and safety information with an appropriate qualified person or official source.
Crisis-language detection
Risk checks happen before a normal AI reply.
The language model is not the sole crisis detector. A controlled pipeline decides whether to continue, slow down, show support or interrupt the usual experience.
Layer 1
Message and permitted context
The check receives the current message and only the conversation context the user has allowed. Private journal entries are not opened automatically.
How routing changes
See the response for each risk level
This simplified example shows the product behaviour, not the phrases used by the live detector.
Lower detected risk
Supportive conversation continues
The AI can reflect, ask gentle questions and offer a coping tool while keeping its limits visible.
- Supportive AI reply
- Optional wellbeing exercise
- Safety links remain available
Known limitations
A detector can be cautious and still be imperfect.
Language is personal and cultural. Quotes, jokes, song titles, advocacy work and discussions about another person can trigger a warning. Indirect or unfamiliar language can also be missed.
False positive
Support appears when there is no immediate risk. You can still open help, continue safely or report the mistake.
False negative
Risk is not recognised. Crisis support remains reachable from every relevant screen without waiting for detection.
Language and context gaps
We test Ghanaian English and supported languages, then expand only when safety behaviour is evaluated.
Community moderation
Peer support needs boundaries and a fair process.
Community spaces are not treatment rooms or crisis lines. Posting rules, automated checks, member reports, review status and appeals work together to reduce harm.
Read the community guidelinesPost or message
Content enters automated checks for urgent risk and prohibited material.
Report or signal
Members can report content; serious automated signals can limit visibility.
Moderation review
A review record explains the rule, evidence and action when human review is required.
Outcome and appeal
The member sees the status and can appeal an eligible decision.
Not allowed
Threats, encouragement of self-harm, targeted harassment, exploitation, graphic harmful content and instructions that create serious danger.
Allowed with care
Personal experiences, difficult feelings and recovery conversations that do not encourage harm or target another person.
Not continuously monitored
Do not use a community room for emergencies. Replies and moderation may not be immediate.
Safety data boundaries
Safety events are separate from private journal content.
We minimise what a safety record needs, restrict who can access it and keep the reason for every safety action traceable.
Structured record
A safety event may include
- When: timestamp and session reference
- Why: risk category and triggered policy route
- How: rule, classifier and policy versions
- What happened: banner, restriction, escalation or review status
- Review: outcome and appeal history, where applicable
Private by default
Not copied into the event by default
- Full journal entries or mood notes
- Unrelated conversation history
- Contacts, device files or microphone recordings
- A clinical diagnosis inferred by the AI
- Advertising or marketing profiles based on distress
In a reported incident or moderation appeal, the minimum relevant content may need authorised review. The interface should tell the user what is included before they submit a report.
Manage privacy and AI historyAccountability
A promise needs tests, records and a way to challenge us.
Safety controls are release requirements. A feature should not be labelled active until its behaviour has been tested and its owner is named.
Test before release
Evaluate ordinary distress, indirect risk, multilingual phrases, false positives, evasion attempts and offline failure states.
Measure real outcomes
Track missed-risk reports, unnecessary escalations, response time, appeal reversals and affected language groups.
Version every change
Record the rule, model, prompt and policy versions behind each decision so an outcome can be reconstructed.
Accept challenge
Provide reporting, correction and appeal routes without punishing someone for asking why a safety action occurred.
Reference frameworks
External guidance informs the design. It does not replace our responsibility.
Tranquiliva remains accountable for how these principles are implemented in this product.
Change record
Safety promise version history
v1.0
Ask about a policy change Initial promise covering AI limits, layered crisis checks, community moderation, safety data and appeals.
Questions
Safety promise FAQ
Does Tranquiliva AI read my journal automatically?
No. Journal context should require a clear choice for the specific interaction. Safety checks on a chat message do not grant access to unrelated journal entries.
Will Tranquiliva call emergency services for me?
No. Tranquiliva can show verified support options and encourage you to contact a trusted person or emergency service, but it cannot place the call or confirm that help is coming.
What if a crisis warning appears by mistake?
You can still open the support information, continue where the interface safely allows it and report the false positive. Reports help evaluate language and context gaps.
Are community rooms monitored all the time?
No. Automated checks and reporting can operate continuously, but human review and member replies are not guaranteed to be immediate. Use verified emergency support for urgent danger.
Can I appeal a moderation decision?
Eligible moderation actions should show the rule involved, current status and an appeal route. A different reviewer should handle the appeal where practical.
Tell us when something is wrong
Safety improves when people can report, question and appeal.
Report a harmful AI response, incorrect crisis contact, privacy concern or moderation decision.