Trust & Safety

Dyva is a platform where people create, share, and interact with AI characters. That only works if people can see how it is policed. This page describes the systems we actually run — not the systems we would like to have.

Our approach has three parts: automated scanning that runs on content as it is created, a review queue that people work through, and enforcement that is sometimes automatic and sometimes human. Section 5 sets out, in plain terms, the things our systems do not do. Read it alongside the rest — it is part of the picture, not a footnote.

This page covers crisis resources, how we moderate content, the safety measures built for AI, the limits of those measures, how to report problems, what happens when we take action, and how we handle age. For specific content rules, see our Acceptable Use Policy.

Dyva does not detect crises inside conversations. Characters are language models. They do not recognise that you are in danger, they do not alert anyone, and nothing you say to a character reaches a human responder. If you are at risk, contact one of these services directly.

Outside the United States, contact your national crisis line. A fuller list, including what to expect when you call, is on our Safety Center.

Moderation runs in two stages: automated scanning that decides immediately, and a human queue that reviews what the scanner flagged.

• Automated scanning. Text content is scanned as it is created — conversation messages, direct messages, room messages, feed and group posts, community replies, character configurations, and voice-session transcripts. A keyword pre-filter selects content for a language-model classifier, which returns a category and a severity. The classifier has the final say when it runs; when it is unavailable, the keyword rules act as a degraded fallback so scanning never goes dark. A background job re-scans recent content every few minutes to catch what the live path missed. Content is checked against our Acceptable Use Policy.
• Automatic consequences, before any human looks. This is the part people are usually not told, so we are stating it plainly. When the scanner rates something critical, the consequence is applied immediately and automatically: a message or DM has its text replaced with a placeholder reading "Content hidden pending review", a post is switched to hidden, and a character is unpublished. Separately, an account that accumulates five critical flags within a 24-hour window is suspended automatically. In all of these cases the action lands first and the human review happens afterwards. The original text is preserved so the decision can be reviewed and appealed — and if we cannot preserve it, we do not redact it.
• Human review. Flags and user reports land in a queue that moderators and administrators work through. A reviewer records an outcome, a resolution note, and their identity against each item. Human review is how automated decisions get confirmed, reversed, or escalated — it is not a checkpoint that every piece of content passes through first, and it is not a gate in front of the automatic actions described above.
• Review before public publication. A character cannot be published publicly or listed on the Marketplace until a moderator approves it. Creators submit a character for review; it stays unlisted while it waits; a reviewer approves it or rejects it with a written reason, and a rejected creator can revise and resubmit. This rule is enforced in the database itself, so it cannot be bypassed by the application. Private and unlisted characters are not pre-reviewed — they are covered by the ongoing scanning above.

Running an AI platform means dealing with risks that traditional content platforms do not face. These are the AI-specific controls we operate:

• Model-provider limits. Characters run on third-party language models, and those providers enforce their own content policies. When a provider refuses a request, Dyva treats the refusal as a decision and does not route around it to obtain the refused output.
• Scanning of character configurations. A character's name, description, and instructions are scanned when it is created, edited, or copied, on the same path as any other content. A character whose configuration is rated critical is unpublished automatically.
• Adult content is gated, not open. Characters marked as adult are only reachable by accounts that have completed identity-document or payment-card age verification through our third-party verification provider, and that have then switched adult mode on. A self-declared birthday does not unlock it. Adult-verified conversations are routed to a model with different content limits — so the limits a character operates under are not identical for every user, and we would rather say that than imply a single universal filter.
• Rate limiting. Message volume, character creation, verification attempts, payment operations, and API requests are rate-limited to make automated abuse expensive.
• Creator visibility is bounded. Creators receive aggregate reporting about their characters only. They never see private evidence quotes from a conversation, they never see individual users named as the subject of something a character learned, and character memory is not a window into a creator's audience.
• Prohibited regardless of configuration. No character setting, prompt, or jailbreak makes prohibited content acceptable. Content that sexually exploits minors, credible threats, and the other categories in our Acceptable Use Policy are violations however they were produced, and we enforce against them when they are detected or reported. That is a policy commitment about consequences, not a claim that our filters are impossible to defeat.

Safety pages tend to describe capabilities and stay quiet about gaps. These are the gaps, stated directly, so you can make your own decisions about how you use Dyva.

• No crisis detection or referral in conversation. We do not run a crisis classifier inside chat, characters do not surface hotline resources on their own, and nothing you say to a character contacts a human. Crisis resources live on static pages, including section 2 of this one.
• No age verification at signup. Age is self-declared during onboarding with a 13+ check. We do not verify it, and we do not estimate it.
• No minor-specific product behaviour. Beyond the 13+ check and the adult-content gate, a 13-year-old account and a 30-year-old account behave the same way. There is no separate teen mode, no age-differentiated filtering, and no parental control surface in the product.
• No usage-time or session-length prompts. We do not track how long you have been in a conversation and we do not prompt you to take a break.
• Incomplete AI disclosure. Dyva is presented throughout as a platform for AI characters, and our Terms state that characters are software. We do not currently render a persistent per-message AI disclosure inside the chat interface, embedded chat widgets on third-party sites carry no AI disclosure at all, and the character-initiated messaging capability described in section 6 composes messages in the character's voice without an AI marker. These are known gaps.
• Scanning is not exhaustive. Not every message reaches the classifier — a keyword pre-filter and a sampling rate decide what does. Images and video are not scanned by the text classifier. Assume that some violating content will be present until it is reported.

We publish this list because it is more useful than a promise. If something here changes, this section changes with it.

Dyva has built the capability for a character to start a conversation rather than only respond to one. We are describing it here because it is the kind of feature that should be disclosed before it reaches anyone, not after.

• Current status. The capability runs in shadow mode. Nothing is delivered to users from it, and the push notification setting that would carry it is off by default. If that changes, this section changes first.
• Controls that exist before it ships. A global do-not-disturb switch, quiet hours, a per-relationship frequency limit, and a re-check at delivery time that drops a message whose reason no longer holds. These are settings you own, in your notification and availability preferences.
• Grounding. A character-initiated message has to cite something specific from your own history with that character before it can be composed. It is not a marketing channel and it is not a re-engagement nudge.
• Known gap. These messages are written in the character's voice and do not carry an AI identity marker. That is unresolved, and it is one reason the feature is not delivering.

Our systems catch a lot, but we rely on the community to catch what we miss.

• In-app reporting. Report buttons are available on characters, user profiles, individual chat messages, feed posts, videos, scenes, and community threads. In-app reports carry the context with them — no need to screenshot or explain where you found it.
• Email reporting. For situations that need more context, detailed evidence, or descriptions of behavior patterns across multiple interactions, send a report to [email protected].
• What happens to a report. Reports are recorded with the reporter, the target, the reason, and any detail you add, and they enter the moderation queue as pending. A moderator evaluates the reported content, the account behind it, and whether the behaviour is part of a pattern, then records an outcome against the report.
• Prioritisation. The queue is ordered by severity, and critical items — child safety, credible threats, imminent harm — are worked first. We do not publish a guaranteed response time. We would rather not state a number we cannot hold to in every case. If your report concerns immediate physical danger, contact emergency services rather than waiting on us.
• Reports on video content. Video reports additionally feed an automatic protection: when enough distinct, credible reporters flag the same video inside a short window, it is shielded from view pending review and restored automatically if the reports do not hold up.
• Confidentiality. Reports are confidential. We do not reveal reporters to the reported party except where law requires it. We do not tolerate retaliation against good-faith reporters.

Actions are proportionate to severity, history, and the potential for ongoing harm. Two of them are applied automatically, as described in section 3.

• Automatic hiding. Content rated critical is hidden the moment it is scanned, without human review. A message or DM shows a placeholder in place of its text; a post is switched to hidden; a character is unpublished. The original is retained so the decision can be reviewed and reversed.
• Automatic suspension. Five critical flags against one account inside 24 hours triggers an automatic suspension, recorded with the offence count that caused it. This is a machine decision, and it is appealable like any other.
• Warnings. A notice documenting the violation and the expected corrective behavior. Warnings are recorded and factor into future enforcement decisions.
• Content removal. Violating content is removed or access is disabled. For time-sensitive violations, content is removed before notification.
• Suspension. Account access is suspended. During suspension, you cannot access your account, interact with characters, use the API, or publish content.
• Permanent ban. Account terminated and content removed or anonymised under the deletion rules in our Terms. Reserved for the most serious violations: child safety offenses, credible threats of violence, and repeated violations after prior enforcement. Banned users may not create new accounts.
• Law enforcement referral. For illegal content or activity — CSAM, credible threats of violence, terrorism, or other criminal conduct — we report to law enforcement and cooperate fully with investigations. We preserve and produce user data in response to valid legal process.

We do not need to warn you before acting, especially when safety or legal obligations are involved. The severity of the violation determines the response, not how many chances you have had.

Automated decisions are the ones most likely to be wrong, and they are the majority of first actions. Every enforcement action can be appealed, including the automatic ones.

• How to appeal. Email [email protected] within 30 days of the action. Include your username, the action you are contesting, any reference numbers from the notification, and why you believe it was incorrect. Creators appealing a Marketplace decision may use [email protected] instead. Removed videos have an appeal control in the product itself.
• Review process. Your appeal is decided by someone other than the person who made the original decision, and the original content is available to them because we retain it when we hide something. We do not publish a guaranteed decision time — see the note on timelines in section 7.
• One appeal per action. Each enforcement action may be appealed once. The appeal decision is final. We send the outcome and our reasoning to your account email.

Two different things are often described together as "age verification". On Dyva they are separate, and only one of them is verification.

• Joining Dyva: self-declared, not verified. You must be at least 13 to create a Dyva account, or older where your jurisdiction requires it. We ask for your date of birth once during onboarding and refuse to continue if the date reports an age under 13. That is a self-declaration and we do not verify it. There is no document check, no payment-card check, and no age-estimation model on the signup path. Anyone willing to type a different year can defeat it. Giving a false age breaches our Terms and we may terminate the account when we have reason to believe the stated age is wrong.
• Adult content: actually verified. Adult-content features are separate and are gated by real verification. Access requires an identity-document or payment-card check completed through our third-party verification provider, recorded against your account; a typed birthday does not unlock it. This check exists because some jurisdictions require it for explicit material, and it applies only to that gate — it is not run on ordinary accounts.
• COPPA compliance. We comply with the Children's Online Privacy Protection Act and do not knowingly collect personal information from children under 13 without verifiable parental consent. If we discover such data, we immediately delete it and terminate the account. Parents who believe their child has provided information without consent should contact [email protected].
• No age-differentiated experience. We want to be exact about this rather than let "we protect minors" imply more than it should. Apart from the 13+ check at onboarding and the adult-content gate above, the product does not behave differently for a younger account: there is no teen mode, no age-tiered filtering, and no parental control surface. Content categories are restricted by the adult gate, not by a user's declared age.
• CSAM response. Content that sexually exploits or endangers minors is removed, the account is terminated, and we report the material to NCMEC and to law enforcement as US federal law requires of us. This is a zero-tolerance policy with no exceptions and no appeals.
• Reporting child safety issues. If you encounter anything on Dyva that exploits, endangers, or harms a minor, report it immediately via the in-app report button or email [email protected] with the subject line "Child Safety Report." These reports take priority over everything else in the queue. If a child is in immediate danger, contact law enforcement or the NCMEC CyberTipline in section 2 as well — do not rely on us alone.

• Transparency reports: not yet published. Dyva has not published a transparency report. We record the underlying data — flags by category and severity, their source, resolutions, reviewer, report volumes, and outcomes — but nothing aggregated from it has been released, and we are not going to describe a publication that does not exist. When the first report is published it will be linked from this section.
• How we communicate actions. When we take enforcement action against your account, we tell you what happened, which policy was violated, and what you can do about it. We do not leave you guessing. The only exception is when disclosing details would compromise an ongoing investigation or create safety risks.

For safety questions, concerns, or reports:

For general legal inquiries, contact [email protected]. For copyright issues, see our DMCA & Copyright Policy. For appeals, email [email protected].