AI Data & Training Policy

Effective date: June 17, 2026

This policy explains how AI features across the APERTURESyndicate platform process your data, which models and providers receive it, and — most importantly — our firm commitment that we never use your personal content to train or fine-tune AI models unless you explicitly opt in.

It applies to ASAI (AI chat, roleplay, the ASAI API and AI image generation, including persistent AI "memory" features), to ETHERANE, and to AI-assisted functionality elsewhere on the platform — the AI tools in ASVentor and machine tagging of Resonance tracks.

AI features are available only to registered users aged 13 or older.

This policy is part of, and should be read together with, our Privacy Policy, our Data Retention Policy, and our AI Output & Limitations notice. Company and registry details are published at Company Registration.

This policy is a working text pending review by qualified EU/Estonian legal counsel; it may be refined for legal precision without weakening the user commitments stated here.

1. Controller

The data controller for AI features is APERTURESyndicate OÜ (registry code 17384111, VAT EE102972654), Priisle tee 8, Lasnamäe linnaosa, Tallinn, Harju maakond, 13914, Estonia. Our single point of contact for all privacy, AI, and data matters is [email protected].

2. What Our AI Features Process

When you use an AI feature, we process the data you provide and the data needed to generate a response:

• Prompts — the text, instructions, and questions you send to an AI feature.
• Conversation history — prior turns in a chat, roleplay, or coding session, used so the AI can maintain context within that session and across sessions you choose to keep.
• Roleplay content — characters, scenarios, and dialogue you create or interact with.
• Uploaded files — documents, images, code, and other files you attach to an AI feature for it to act on.
• AI "memory" — facts, preferences, and context that AI features derive from your use over time so they can respond more helpfully. Memory is off by default; when you start a paid subscription it switches on for a short trial and then switches itself off again unless you keep it on. You can view, edit, and delete this memory in your account settings, and disabling memory stops new memory from being created.
• Your platform account, when you ask about it — your profile, badges, what your subscription unlocks, who you follow and what you listen to on Resonance. ASAI reads only your own account, only when a question needs it. The switch is in Settings → Privacy and is on by default.
• Approximate location — city, region and country, so answers can be local. It is worked out in your browser: from the device's location if you allow it, otherwise from your IP address through an IP-lookup service. The switch is in Settings → Privacy and is on by default.

We process this data to deliver the feature you requested, to maintain your account, to keep the service secure and reliable, and to comply with our legal obligations. The legal bases for this processing are set out in our Privacy Policy.

3. Which Models, and Who Receives Your Data

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).

To deliver AI features, the content you submit (and the context required to answer it) is sent to the following recipients:

Recipient	What is sent, and when
Anthropic (Claude Sonnet, Claude Opus)	The scene — characters, world, recent turns and your message — for every roleplay turn on the Prime or Max tier, including the scene's hidden calls (deciding the turn, dealing a scenario's hidden roles).
OpenAI (GPT Image, embeddings)	Your picture request and, for an edit, the picture you supplied; for character portraits and sprites, the character's description and reference art; memory summaries and the message used to look them up, to compute their numerical representations — only while memory is on.
Model hosting providers	Prompts, conversation context and attached images — for every request to ASAI Mini and ASAI Arc, and for the automatic safety checks described in Section 5.
Speech providers	Audio you dictate, attach or link (recognition), and the text of a reply you ask to have read aloud (synthesis) — only when you use those features.
Web search and page-reading providers	Your search query, or the URL you supplied — only where a feature looks something up live to ground its answer.
Document parsing providers	The uploaded file — only where a PDF or document has to be converted into text before an AI feature can read it.
Code execution providers	The code and the files involved — only when ASAI runs code for you in an isolated sandbox.
Vector memory storage	Excerpts of your conversations and their numerical representations — only while memory is on.

Every one of these recipients is contractually bound as our GDPR sub-processor: they act only on our instructions, process the data only to perform the operation requested, and may not use your content for their own purposes or to train their own models. Transfers outside the EEA are covered by Standard Contractual Clauses or an adequacy decision, as set out in the sub-processor register.

What we publish, and what we do not. Which models run under the ASAI Mini and ASAI Arc names, and which companies host them and provide the infrastructure categories above, we do not publish — we describe those recipients by category, which is what Article 13(1)(e) of the GDPR permits. We improve and replace what runs under our own names, and change infrastructure providers, as the technology moves, without that being a change to this policy; what does not change is the categories of processing described above and the protections attached to them. A third-party model is always shown under its maker's name.

If you are a business customer and your own compliance obligations require the named list, ask us at [email protected] and we will provide it under confidentiality.

Your own key. In Settings → Roleplay you can add your own key from a provider you chose yourself. Scenes then go to that provider under your own account with it — and, if you switch that on, so does the scenes' background work. That provider is a recipient you chose, not our sub-processor: its own terms and privacy policy apply, including whether it may train on what you send. Your key is stored encrypted on our server until you remove it. See Recipients you choose yourself.

Outside AI generation, the platform is supported by payment, hosting, network security and email providers — those are named in the sub-processor register, which also records any change to these lists.

AI output is generated, not retrieved from other people. Responses are produced by a model computing over your input — they are not pulled from, and do not expose, another user's chats, files, or memory.

4. Training: Opt-In Only

We do not use your personal content to train or fine-tune AI models unless you have explicitly opted in.

This is an unconditional commitment. By default — and unless you actively choose otherwise — none of the following is ever used to train, fine-tune, or otherwise develop AI models, by us or by our sub-processors:

• your prompts;
• your conversation, chat, or coding history;
• your roleplay content;
• your uploaded files;
• your AI memory.

One thing does not depend on your switch alone: cards you publish. A character or a world you publish is played by other people. When a player who has allowed training plays a scene with it, that scene — including your card as ASAI received it — may be used in the anonymised dataset, whoever wrote the card. Publishing grants this use, as set out in Ownership & Licence, Section 2; for works published before 1 October 2026 it applies from 31 October 2026. Your private cards and your own scenes stay under your own switch.

4.1 How you are asked, and how to change your mind

• When you first open ASAI, a short setup window asks — on a step of its own — whether you want to help. The two answers, Allow and Not now, are the same size and weight; nothing is pre-selected, and saying no changes nothing about what you can use. Accepting the Terms of Service is a separate checkbox and does not include this consent.
• If you said "Not now", we may ask once more, and only once: the moment you reroll a roleplay reply and carry on with the one you liked. Closing that card counts as an answer, and we will not ask again.
• In Settings → Privacy, the "Help improve ASAI" switch shows your current choice and changes it in one tap. It is off by default.
• Scene by scene: after allowing, you can open a scene's menu and choose "Keep this scene out of training". That scene — its past turns as well as future ones — is then excluded.

Every decision is recorded with its date, the version of the text you saw and where you made it (setup, settings, or the one-time reminder), so that we can show what you agreed to (GDPR Article 7(1)).

You may change your choice at any time. Withdrawing stops any further use of your content: turns already marked for training lose that mark at once and are left out of every future dataset. Content already incorporated into a trained model cannot be extracted from that model, and your withdrawal does not affect the lawfulness of processing carried out while your consent was active.

4.2 What is used when you allow it — and what never is

Used: your roleplay scenes — the scene's context (characters, world and scene state as ASAI received them) and the turns of the dialogue — and which reply variant you kept: when you reroll a character's reply and carry on with one of the variants, the kept variant and the ones you passed over form a comparison that teaches the model what a good reply looks like.

Never used, even with consent:

• regular (non-roleplay) chats, coding sessions and research;
• your email address, username or handle, IP address, payment data;
• scenes you marked "Keep this scene out of training";
• uploaded files and your AI memory;
• anything from an account under an open review for suspected under-age use.

4.3 How the data is prepared

Once a month we assemble a dataset from the previous month's eligible scenes. Before anything leaves the database, email addresses, phone numbers, @handles and IP addresses in the text are replaced with placeholders; each scene is keyed by a one-way hash instead of its chat or account identifier. Only people whose consent is active on the day of the export are included. The dataset is stored encrypted in our private storage in the EU and is never published.

4.4 What we train, and who sees it

We use these datasets to train and evaluate our own roleplay model for ASAI — so that characters keep their voice and follow the scene. The datasets are never sold and never given to other companies for their own use. Where the training computation itself runs on rented infrastructure, the provider acts as our GDPR sub-processor under the same contractual terms as in Section 3: it may process the data only to run our training job and may not keep it or use it for anything else.

If you allow training, we use the content only to improve model quality and only under the safeguards above; it does not change any of the other protections in this policy, including Section 5.

Allowing training also means a named person on our team may open those specific messages in our admin tooling — to judge the quality of an answer, or to act on something dangerous found in them. Only the messages your consent covers can be opened on this ground; each opening requires a written reason and is recorded with the reader's name and time. This is described in Trust & Safety, Section 3.5. Turning the setting off ends that access from the moment you turn it off.

5. Inferences We Will Not Make About You

We design our AI features to assist you, not to profile you. We do not use AI features to:

• provide medical, psychiatric, or psychological diagnoses about you; or
• infer your sexual orientation, gender identity, political opinions, religious or philosophical beliefs, financial status, health conditions, immigration status, or criminal history.

If you choose to state any of this information yourself, the AI may use it within your session to respond to your request, subject to the rest of this policy and your settings. We do not attempt to derive these special categories of data from your behaviour, and we do not build advertising profiles from your AI activity.

Two narrow automatic safety checks do read what you write, and you should know about them:

• The crisis check. When a message contains words about dying or self-harm, a small model judges whether you are talking about yourself or about a story, so that the reply can take it seriously and, if it repeats, a window with crisis lines can appear. It keeps only a short counter for a few hours and stores no record about you — 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 only flags the account for a person in support, together with the few words that triggered it; it does not restrict anything — see Age & Eligibility, Section 2.2.

6. Operational Signals Used to Improve the Service

Separately from your content, we may generate and analyse anonymised, aggregated operational signals about how the system itself is performing — for example, overall request volumes, error and latency patterns, feature usage trends, and aggregate quality measures.

These signals describe the behaviour of the service, not your identity or the substance of your conversations. They are aggregated across many users and stripped of information that could identify you or reveal what you wrote, and we use them only to keep the platform fast, stable, and reliable and to improve features. This is distinct from the training described in Section 4 and does not rely on, or require, your opt-in.

7. AI Transparency (EU AI Act)

When you use an AI feature, you are interacting with an artificial-intelligence system, not a human. We label AI features clearly so this is never ambiguous.

AI systems can produce output that is inaccurate, incomplete, outdated, or misleading, and AI-generated images or text may not reflect reality. Do not rely on AI output for medical, legal, financial, or other professional decisions without independent verification by a qualified professional. For the full explanation of what our AI can and cannot do, see AI Output & Limitations.

8. Your GDPR Rights

As a data subject in the European Union, you have the right to access, rectify, erase, restrict, and port your personal data, to object to certain processing, and to withdraw consent (including your training opt-in) at any time without affecting prior lawful processing.

You can exercise most controls directly in your account settings (including viewing and deleting AI memory and changing your training choice), or by contacting [email protected]. Full details on how to exercise each right, and your right to lodge a complaint with the Estonian Data Protection Inspectorate (Andmekaitse Inspektsioon), are set out in our Privacy Policy.

9. Retention

AI-related data — including prompts, conversation history, roleplay content, uploaded files, and AI memory — is retained only as long as needed to provide the feature and meet our legal obligations, and is deleted or anonymised after that. Deleting a conversation, clearing your AI memory, or deleting your account removes the associated content according to the schedule in our Data Retention Policy.

10. Changes to This Policy

We may update this policy as our AI features, sub-processors, or legal obligations change. Material changes will be communicated through the platform, and the "Last updated" date above will always reflect the current version. We will never retroactively weaken the opt-in-only training commitment in Section 4 for content already processed under this policy.

Contact

For any question about this policy or your data, contact [email protected].

Version history

• v1.6 — 2026-10-12 — Section 3: numerical representations (embeddings) of memory are now computed by OpenAI instead of model hosting providers.
• v1.5 — 2026-10-01 — Section 4: a card you publish can be part of the scenes of players who have allowed training (works published earlier: from 31 October 2026). Section 3 now states which models are ours and which are not: our own line is ASAI Mini and ASAI Arc; Claude Sonnet and Claude Opus by Anthropic (the Prime and Max roleplay tiers) and GPT Image by OpenAI (images) are third-party models and are named as recipients. The recipient table gained speech synthesis, code execution and vector memory storage, and a paragraph on your own key. Section 2 now covers account data ASAI reads when you ask, approximate location, and the memory trial that starts with a paid subscription. Section 5 describes the two automatic safety checks. Product scope corrected.
• v1.4 — 2026-09-27 — Section 4 rewritten to match how consent works now: asked on its own step in the ASAI setup window with two equal answers, one reminder at most, a per-scene exclusion, and a record of every decision. Training now covers roleplay scenes only (including which reply variant you kept); regular chats are never used. Added how datasets are prepared (monthly, anonymised, hashed identifiers, consent checked on the day of export, suspected under-age accounts excluded) and what they are used for (our own roleplay model; never sold or shared).
• v1.3 — 2026-08-04 — Removed two dead product names: ASChat (the user-to-user messenger was withdrawn — see Terms of Service v2.2) and ASAIStudio (renamed ETHERANE). Company details, the minimum age and the contact address are now placeholders resolved from the single fact registry instead of being written out by hand, which is what let them drift out of step with the rest of the documents in the first place.
• v1.2 — 2026-08-01 — Section 3 now describes recipients by category — AI inference, speech-to-text, web retrieval, document parsing — instead of naming individual vendors. Our products are ASAI Mini and ASAI Arc; the models and infrastructure behind them are confidential to our business. Category-level disclosure is what Article 13(1)(e) of the GDPR requires, and business customers who need the named list for their own compliance can request it under confidentiality.
• v1.1 — 2026-08-01 — Section 3 corrected to match what the service actually does: which stage of processing goes to which category of provider, and that speech-to-text is transcription only. Added an explicit statement that we do not publish which models sit behind the ASAI Mini and ASAI Arc product names.
• v1.0 — 2026-06-22 — Initial publication.

---

This document is provided in English, which is the controlling language; any translation is for convenience only.