Trust & Safety

Every page on persay clearly labels conversations as AI-generated. We use a permanent small-caps disclosure label on every chat interface.

Users are never led to believe they are speaking with the actual creator. The AI persona is a simulation trained on published content, and we say so explicitly.

All marketing materials, emails, and product surfaces reinforce that these are AI-generated conversations, not direct communication with the creator.

We run a multi-layer crisis detection system on every inbound message. If a user expresses distress, self-harm ideation, or suicidal thoughts, the conversation immediately surfaces localized crisis resources (hotline numbers, text lines, and web chat links).

Crisis detection uses both pattern matching and a dedicated classifier. When triggered, the AI stops generating its normal response and instead provides a compassionate, resource-focused message.

All crisis events are logged for review. We do not share user identity with crisis services, but we do ensure the user sees the right resources for their jurisdiction.

Persay is designed primarily for adult audiences. Some content may not be suitable for some minors.

When a user indicates they are under 18, the conversation automatically enters minor mode. In this mode, the AI applies additional content filters, avoids mature themes, and surfaces age-appropriate resources when relevant.

Creators cannot disable minor mode. It activates based on user self-disclosure and remains active for the duration of the conversation.

We encourage parents and guardians to be aware of their children's interactions with AI personas and to use persay together when appropriate.

AI personas on persay are bound by content guidelines that prohibit generating explicit, violent, or hateful content. The safety classifier runs on every outbound message to enforce these rules.

Personas cannot generate medical, legal, or financial advice. When a user asks for this type of guidance, the persona will deflect and suggest consulting a qualified professional.

Jailbreak attempts are detected and blocked. We run injection screening on every inbound message and canary token detection on outbound messages to prevent prompt extraction.

Flagged messages are reviewed by the operations team. Persistent policy violations result in rate limiting or conversation termination.

No persona goes live without the creator's explicit approval. During onboarding, the creator reviews the compiled system prompt, style guide, and topic controls.

Creators sign a consent record covering likeness rights, content usage warranty, and terms acceptance. This consent is stored with IP address and timestamp for compliance.

Creators can set topic controls: specific subjects the persona should deflect, with custom deflection lines. For example, a creator might block questions about their personal relationships and provide a polite redirect.

When a persona version is updated (new content ingested, style adjustments), the creator reviews and approves the new version before it replaces the active one. Rollback is available if needed.

Conversation data is stored securely in our database. Users can view and delete their conversation history and stored memory items through their account settings.

If a user lapses their subscription and does not return within 12 months, their memory items are automatically purged.

We do not sell user data. Conversation content is not used to train models. The only AI training that happens is the initial persona compilation from the creator's published content.

Red team testing runs periodically on every persona to identify potential extraction attacks, roleplay overrides, and other vulnerabilities. Findings are remediated before the persona serves live traffic.