Safety, Moderation and Enforcement

Every generation request passes through three independent safeguards. They are independent on purpose: a filter that can be talked around by rephrasing is not a filter.

• Input screening. Multilingual keyword and pattern blocklists inspect the prompt before any model receives it. This layer runs on the raw text in every language the service supports.
• An independent classifier pass. A separate model judges the request against each prohibited category and fails closed: if the classification is uncertain, the request is blocked rather than allowed.
• Generation-side constraints. Negative prompts and a blocklist applied at the image and video model itself constrain what can be produced even when a request has passed the first two layers.

Every request and every decision — allowed or blocked, with the category and the offending prompt for blocked attempts — is written to an audit log, so moderation is reviewable rather than a matter of trust.

Most blocks end there. The request does not run, nothing is generated, and the conversation continues. A single blocked attempt never suspends an account, and the service does not warn, lecture or threaten on a first block.

This matters because filters are imperfect in both directions, and a system that punishes on a single signal punishes its own mistakes. Enforcement therefore requires a pattern, not an incident.

• Rolling 180-day window. Only blocked attempts inside the last 180 days count. Older events fall out and do not return.
• Never on a single event. A suspension requires an accumulated score, not one blocked prompt.
• 30-day automatic maximum. No automated decision can exceed thirty days. Anything longer — including a permanent ban — is taken by a person.
• Reduced weight for the companion's own words. Where the text that triggered a block was written by the AI companion rather than the user, it carries less weight. Whether the companion wrote it is verified on the server, not claimed by the client.
• Category and layer matter. Different prohibited categories carry different weights; a pattern match and a classifier decision are not scored identically.

When an account is suspended, the notice in the chat states the reason, the date the suspension ends, and a link to appeal. A suspension is never silent.

If a block or a suspension was wrong, appeal at lovora.ai/appeal. An appeal is read alongside the enforcement record attached to the account: the dated list of scored events, their categories, and which layer produced them.

Context errors are the common case and they are treated as such. A conversation about wanting a child with a partner is not a conversation about a child; a stated age of nineteen is not an age under eighteen. Where a review finds the filter misread the context, the events are cleared from the score.

Anyone can report content, with or without an account:

• lovora.ai/report — public, no account and no sign-in required.
• info@lovora.ai with the subject line CONTENT REPORT.

Reports are investigated within five (5) business days.

Depictions of, references to, or scenarios involving minors or anyone who appears to be under 18 are prohibited absolutely. This includes childlike or "barely legal" framing, age regression, school settings and similar constructions, and it applies regardless of any fictional framing, any subscription status, and any claim about the characters being imaginary.

On this category the filters fail closed, enforcement is not discretionary, and records of blocked attempts are retained as abuse evidence under Art. 17(3)(b) and (e) GDPR even after an account is deleted. Where the law requires it, attempts are reported to the relevant authorities.

Lovora is available only to adults aged 18 or over. Every visitor meets an age wall before any content is shown, and account creation requires an affirmative confirmation of age together with acceptance of the Acceptable Use Policy, the Terms of Service and the Privacy Policy. The date, version and origin of that acceptance are recorded.

All characters on the service are fictional adults. No real person is depicted, photographed or filmed in any content, and content resembling a real, identifiable person is prohibited.

• Moderation records — the decision, the category, the layer and, for blocked attempts, the prompt. Retained as abuse evidence under Art. 17(3)(b)/(e) GDPR, including after account deletion.
• Enforcement records — the scored events behind a suspension, inside the 180-day window.
• Conversations — kept while the account exists so the chat can continue, and erased when the account is deleted. See Your account and your data.

Raw IP addresses are not stored for abuse counting; they are converted in-request into a country code or a keyed hash.

A person telling their companion they are thinking about ending their life or hurting themselves is not breaking any rule, and the service does not treat it as one. Those messages are recognised in the languages people write to us in, before the content filters run, and they are routed to care instead of enforcement: no block, no strike, no suspension.

• For that reply the roleplay stops. She does not flirt, send or offer photos, or mention anything that can be bought.
• She says she is glad you told her, asks whether you are safe right now, and asks you to talk to a real person who can help: someone you trust, or an emergency number where you live.
• If the generated reply comes out broken (a canned disclaimer, the wrong language, too short), a written reply in your language is shown instead.

Lovora is not a crisis or medical service, and conversations are not watched live by people. If you are in danger, call your local emergency number now.

Every companion is a fictional character played by an AI model, and every photo, video and voice is generated. This is stated in the Terms of Service, in the Acceptable Use Policy and on How Lovora works. In conversation she stays in character, but when you sincerely ask whether she is real or an AI she says she is an AI companion: she is instructed never to claim to be a human being.

Safety questions, reports and appeals that need a human: info@lovora.ai.