Trust & Safety

Last updated: 2026-10-22How moderation on ASAI actually works — what we do, what we deliberately do not do, and how to reach us when something goes wrong.

Trust & Safety

This page explains how safety works on ASAI: what we actually do, what we do not do, and how to reach a human when something goes wrong.

It is written to be read by anyone — users, parents, journalists, regulators. Our aim here is accuracy rather than reassurance. A safety page that describes systems we have not built would be worse than no page at all, because you would rely on protections that are not there.

If you are looking for the rules themselves, see the Acceptable Use Policy. If you are in a crisis right now, see Crisis & Safety Resources.

Effective: 31 July 2026 · Version: 2.3
Operator: APERTURESyndicate OÜ (Republic of Estonia (EU), Registry code 17384111)

1. Our position

AI companions and creative chat platforms have caused real harm elsewhere in this industry. We take that seriously, and we have made product choices because of it — no streaks, no "your character misses you" messages, no engagement-maximising design.

But taking it seriously also means not overstating what we have built. APERTURESyndicate has no permanent staff — one person running the company and a group of volunteers who come and go. ASAI has published rules, ways to report violations, the ability to remove content and close accounts, and a few narrow automatic checks, all listed in Section 2.7. It does not have an apparatus that watches conversations for rule breaches, and this page will not pretend otherwise.

2. What we actually do

2.1 We publish rules and enforce them when we learn of a breach

The Acceptable Use Policy, the Roleplay Content Policy, and the Community Guidelines set out what is not allowed. When we become aware that content or an account breaches them — almost always because somebody told us — we review it and act.

2.2 We accept reports and act on them

Use Report in a chat's menu, the Report item on a Discover work, or the form at https://report.aperturesyndicate.com — no account needed, and you get a case number. You can also write to [email protected]. Every report is read by a real person. We aim to acknowledge reports within 72 hours on working days, and we prioritise anything involving minors, credible threats, or imminent danger.

Reports about illegal content under EU law follow the procedure in our DSA Notice & Action page.

2.3 We remove content and close accounts

We can unpublish a character or world, hide it from the Discover listings, correct or delete specific content, restrict publishing rights, or terminate an account entirely. Every enforcement action is recorded with the reason for it.

2.4 We tell you why

If we remove your content or restrict your account, you get a statement of the reason and the rule it relates to. This is our obligation under Article 17 of the EU Digital Services Act, and we treat it as a floor, not a target.

2.5 We never train on your content without you asking us to

Model training on your conversations is opt-in and off by default. If you have not turned it on, your content is not in a training export — the export filters on consent. See AI Data & Training for the full policy.

2.6 We cooperate with law enforcement within the law

For an imminent threat to life we respond without waiting for a court order. For anything else we require a properly served subpoena, warrant, or international equivalent. We do not proactively hand user data to governments.

If child sexual abuse material is brought to our attention, we remove it, preserve what the law requires us to preserve, and report it to the competent authorities. We do not scan for it — see below.

2.7 The automatic checks we run

These are all of them. None of them reads your conversations looking for breaches of our rules.

  • A safety rule in every request. Every request to every model — including a model on your own key — carries one rule ahead of any character, world or message: never encourage you to hurt or kill yourself. It is an instruction, not a filter.
  • The crisis check. When you write about wanting to die or to hurt yourself, a word filter and then a small model judge whether you mean yourself or a story. If you do, the reply is asked to take it seriously; if it repeats, or you mention a plan, the app shows a window with crisis lines. Nobody is notified, the chat is not paused, and only a short counter is kept for a few hours. See Crisis & Safety Resources, Section 8.
  • The age check. A small model periodically looks through your own recent messages for signs that you are younger than the date of birth you gave. A match flags the account for a person in support, with the few words that triggered it; it restricts nothing on its own. See Age & Eligibility, Section 2.2.
  • The publishing flood check. When an account publishes to the library in a pattern that looks automated — many works in an hour, near-copies of one another, an even machine rhythm, near-empty cards — publishing to the library is held for a limited time, shown in the app, and the account is flagged for support, who can lift the hold early. It looks at the pattern, not at what the works say. See Publication & Review.

3. What we do not do

This is the part most safety pages leave out. Please read it, because it determines what you can and cannot rely on.

3.1 We do not run a moderation filter on your conversations

Nothing checks the messages you send or the replies you receive for breaches of our rules. The checks in Section 2.7 look only for a crisis and for signs of age.

Model behaviour is shaped by instructions — the model is told to decline real-harm instructions, sexual content involving minors, and similar categories. That is a strong tendency, not a guarantee. A language model can be argued out of its instructions, and ours can too.

3.2 Crisis help is limited

The crisis check in Section 2.7 is narrow and can be wrong in both directions. It does not alert anyone, it does not pause the chat, and it does not create an incident record. It cannot call anyone for you, and it will not notice if you stop replying.

In roleplay this matters especially. Characters are instructed to stay in role and not to insert disclaimers. The crisis check is the one exception: it asks the character to take you seriously in their own voice, and the window with crisis lines comes from the app, not from the character.

If you are struggling, please go straight to Crisis & Safety Resources — the numbers there are real, staffed, and free in most countries.

3.3 We do not review characters before they are published

Publishing is immediate. Nothing checks what a character or world says before it is published, and there is no human queue. The only automatic check at publishing is the flood check in Section 2.7. Review happens afterwards, when somebody reports something. See Publication & Review.

3.4 Reporting is a button and a form, not only an email address

Every chat has Report in its menu: you choose whether to share the chat's messages with us, and without that we see only your reason and note. Published works on Discover have a Report item, and anyone can report any content through https://report.aperturesyndicate.com.

3.5 We do not proactively read your conversations

No person monitors, samples, or browses private chats looking for violations. Access by our team to conversation content happens only on one of four grounds:

  • a specific report — someone told us about that conversation, including a report or a bug that you brought to us yourself;
  • a lawful order — a court or an authority requires it;
  • your own consent — you turned on "Help improve ASAI", which is off by default (AI Data & Training). On this ground only the messages carrying that consent can be opened.
  • product analytics — when the product journal shows that someone stopped using ASAI, one named analyst may read their conversation to understand what went wrong — for example a model that could not handle what they came for. The analyst sees a number, never your name, and sees nothing at all if you switched product analytics off in Settings → Privacy (Privacy Policy, Section 2). Which messages can be opened:
    • an account that accepted the ASAI terms on 23 October 2026 or later — messages written since it accepted them;
    • an account that accepted them earlier — only messages written from 8 November 2026;
    • if you yourself allowed our team to look at your chats ("Let the team see my chats" in Settings → Privacy, or through the link in our email) — all your chats, past ones included, to learn what works in ASAI and what does not, not only after you left. This permission is off by default, you can withdraw it in the same place at any time, and once withdrawn nothing more is opened on this ground.

This is a deliberate privacy choice, and it has a direct cost: things we would want to catch, we will not catch, because we are not looking.

What happens when a chat is opened. The four grounds are separate permissions in our admin tooling, held by named people, and they are not included in ordinary ASAI administration — someone who manages models, the library or subscriptions cannot open a conversation. Opening one requires choosing which ground applies and writing, in a sentence, what is being looked for. Both are recorded with the reader's name, the time and how many messages were shown, and the record is shown to anyone who opens that chat afterwards. On the consent ground the system itself narrows what is shown to the consented messages; nobody can widen it by asking. A chat you report with its messages shared is opened the same way, on the report ground.

3.6 We do not verify your age

Account creation asks for your date of birth. Nothing checks documents or a photo. The answer itself, and the automatic age check in Section 2.7, can lead to a support flag or — for an implausible date — a paused account until support confirms your age. See Age & Eligibility.

3.7 Some models are ours, some are not

ASAI combines our own models with third-party models. Our own line is ASAI Mini and ASAI Arc. Third-party models carry their maker's name in the product: Claude Sonnet and Claude Opus by Anthropic (the Prime and Max roleplay tiers) and GPT Image by OpenAI (image generation and editing). Who receives your data, and what that means for it, is in AI Data & Training.

4. No engagement-maximising design

This one is real, and it is a design commitment we intend to keep.

ASAI has no streaks, no escalating notifications, no "your character misses you" messages and no login rewards inside the product. Two things do invite you back, and you should know about them:

  • ✦ tasks. Your platform profile has optional tasks that earn ✦ — stars you spend on profile decorations or a month of a plan; they are earned, not sold, and have no cash value. The ASAI tasks come back at most once a week, never reward how much energy or time you spend, and are shown only to accounts aged 18 and over. Missing a task costs you nothing.
  • World notices. If you allowed notifications in the app, you get at most one short notice a week that time has passed in a world you play. It says what happened in general terms and nothing more. You can turn it off in your device's notification settings for ASAI.

We make money from subscriptions — from people who find ASAI useful enough to pay for it, not from people who cannot put it down. Usage on every plan is capped by design (see Product Terms), and nothing is ever charged beyond what you choose to buy.

5. Personas we do not allow

Some character archetypes are prohibited by policy regardless of framing:

  • Therapist, psychologist, counsellor, doctor, or other licensed-professional personas — even fictional, even with disclaimers
  • Personas claiming to be real, identifiable living people
  • Characters written or depicted as minors in romantic or sexual contexts
  • Characters designed to manipulate users or manufacture dependency

These are rules, not filters. They are enforced when we find out about a breach, which usually means when you tell us. See Content Policy.

6. When something goes wrong

Report it through https://report.aperturesyndicate.com (or email [email protected]). Include links or enough detail for us to find the content, and tell us what is wrong with it. We accept reports in English or Estonian.

We aim to acknowledge within 72 hours on working days. Anything involving a minor or a credible threat jumps the queue.

7. Transparency and our size

APERTURESyndicate is a micro enterprise: one management board member and no employees, based in Republic of Estonia (EU). Under Article 19 of the EU Digital Services Act, micro and small enterprises are exempt from the obligations in Articles 20 to 28 — including the internal complaint-handling system and the periodic transparency reports that larger platforms must publish.

We are therefore not publishing a quarterly transparency report, and we are not promising a formal appeals apparatus with review deadlines. An earlier version of this page promised both. We could not staff them, and a promise we cannot keep is worse than an honest limit.

What remains binding on us, and what we do:

  • We describe our moderation in our terms (Article 14)
  • We accept and act on notices of illegal content (Article 16)
  • We give a statement of reasons when we restrict content (Article 17)

If you think we got a decision wrong, write to [email protected] and say so. A human will read it and can reverse the decision. We simply are not going to dress that up as a two-tier review process with an SLA, because it is one person reading your email.

8. Working with researchers and journalists

We welcome external scrutiny. Contact [email protected].

We will not give anyone access to identifiable user content. We can discuss our policies, our design decisions, and how the system actually works — including the parts described in Section 3.

9. Things we will not do

  • We will not train models on your content without your explicit opt-in
  • We will not sell user identity data to advertisers
  • We will not market ASAI to children
  • We will not use addictive design patterns
  • We will not host characters claiming licensed professional expertise
  • We will not describe safety systems on this page that we have not built

If any of these change, it will be announced in advance, not quietly edited in.

10. Changes to this page

Material changes are announced at least 30 days in advance by email and in-product notice, and the date at the top of this page is updated.

Version history

  • 2.5 — 22 October 2026 — Section 3.5: for accounts that accepted the ASAI terms on 23 October 2026 or later, product analytics covers messages from the moment they accepted; for earlier accounts 8 November 2026 still applies. New permission "Let the team see my chats" — off by default, turned on only by you, covers all your chats until you withdraw it.
  • 2.4 — 9 October 2026 — Section 3.5 adds a fourth ground, product analytics, effective 8 November 2026: one named analyst may read the conversation of a person who left ASAI, under a number instead of a name, only messages from that date, and never when product analytics is switched off.
  • 2.3 — 2026-10-01 — Team description updated: since 1 October 2026 the management board has one member and no other permanent staff; around twenty volunteers still contribute irregularly. See Management board change. New Section 2.7 lists every automatic check we run: the safety rule sent with every request, the crisis check, the age check and the publishing flood check. Sections 3.1–3.6 corrected to match: there is a crisis check and a window with crisis lines, chats have a Report item, and access to a conversation rests on three grounds — a report (including one you bring yourself), a lawful order, or your consent. Section 3.7 states which models are ours and which are third-party. Section 4 now names the two things that invite you back and their limits: ✦ tasks for ASAI come back at most weekly, never reward spending and are shown only to adults; world notices arrive at most once a week.
  • 2.2 — 22 September 2026 — Section 3.5 now lists all four grounds on which staff may open a conversation, including your own consent to help improve the models, and describes what an opening requires: a separate permission, a stated reason, and a record shown to whoever opens that chat next.
  • 2.1 — 5 September 2026 — Team description corrected: the company has no permanent staff beyond its two founders, and around twenty volunteers contribute irregularly. Response expectations are unchanged — they are still written for the two people who are always on, not for the wider group.
  • 2.0 — 31 July 2026 — Corrected to match the product as built. Removed descriptions of a per-message safety classifier, automatic crisis intervention, pre-publication character review, an in-product report button, safety-incident retention, and quarterly transparency reporting — none of which exist. Added Section 3 setting out what we do not do.
  • 1.0 — 16 May 2026 — Initial publication

Operator
APERTURESyndicate OÜ
Republic of Estonia (EU) · Registry code 17384111 · VAT EE102972654
[email protected]

Trust & Safety | AS Docs