AI & Generative Content Disclosure

We use AI-assisted systems in the following parts of the product:

• Oracle, Tarot, and divination experiences. Generative models help compose interpretive text presented to you. Outputs are creative reflections, not predictions, advice, or guarantees.
• Ask Cosmos conversational assistance. You can ask questions in natural language about the Cosmos Tools catalog. When a request matches a supported domain and has the required inputs, the corresponding result widget is shown inline in the private conversation. Voice recordings may be transcribed to process a request, and text responses can be converted to audio. These features can be incomplete or unavailable and do not replace professional advice.
• Recommendations and search. Vector embeddings, ranking, and retrieval models are used to surface posts, books, videos, and people that may resonate with what you are exploring.
• Intention check. Your intention is shown exactly as you wrote it. We don't score, rate or color-code intentions. Intention check: in AI-assisted mode, when you add an intention to a public post, your intention and the post's text (or a video's transcript) may be sent to our AI provider (OpenAI) to check whether the intention fits the post and follows our rules. AI never writes or changes your intention. Requests don't include your user ID, and our provider doesn't use them to train its models. This never happens for private or followers-only posts or in Human Community mode.
• Content moderation. In AI mode, automated classifiers check public text and images for selected safety signals such as spam, exploitation, self-harm and prohibited commercial pressure. Private Whispers and private chats are never sent to AI, in any mode, even when they're reported (see section 1c). Some reports, escalations, appeals, and other cases may also receive human moderator review; we do not promise an individual human review of every item. These classifiers are not an automated CSAM-detection system and should not be treated as one; use the Child Safety & CSAE Standards report path for child-safety concerns.
• Avatar, sigil, and creative generation. AI sigil generation is off in Human Community mode. Uploaded sigils in Human Community mode are held for people review and are never sent to AI vendors; in AI-assisted mode they use the same image moderation pipeline as post images. Image, audio, and 3D generation models (including our third-party 3D avatar provider, Meshy) render personalised assets at your direction.
• Akashic Engine feed modes. You can toggle between Human Mode (content ranked by community engagement signals, without the AI positivity-prompt recommender) and AI Mode (AI-powered recommendations shaped by a community-authored positivity prompt). In AI Mode, the active positivity prompt — proposed and voted on by the community through the Akashic Governance Policy — is injected as context to guide recommendation tone. In Human Mode, the positivity-prompt recommender is not applied; ranking uses engagement signals. Separate safety systems for public content are described in section 1c.

Cosmos Tools are a signed-in experience inside The Cosmos. They expose fourteen allow-listed domains for calculations, directories, and interpretive prompts. When you run a tool, Ascended Social sends only the required inputs to RoxyAPI from our authenticated server; provider credentials and raw provider responses are not exposed to your browser or to a public API.

Cosmos results may be calculated or AI-assisted depending on the selected operation. They can be incomplete, convention-sensitive, or wrong. They are not predictions or professional advice. Read the Cosmos guide and Wellness Disclaimer before using them.

Ask Cosmos conversations are private to the signed-in account and contain the messages you choose to send. A matched tool result is shown inside the conversation and is saved to Cosmos history only when you choose to save it. Conversation titles are local curated names or names you enter; private prompts are not sent to a model solely to generate a title. Voice input is used for transcription and is not retained as a recording after the request; the resulting transcript may be stored as a conversation message.

Pendulums is a public, non-personalized discovery surface for eligible Human Community content. Its Trending view balances engagement with content age; its Popular view uses accumulated engagement more directly. These are deterministic server-side ranking modes, not AI-generated explanations and not a promise that every viewer sees an identical list at every moment.

Pendulums is separate from personalized Discover recommendations and from Egregors, the Cosmos visualization of shared group energy. Public visibility, publish/readiness state, copyright removal, flags, and public-event status are checked before ranking. Human Community mode and AI-assisted discovery or moderation are separate pathways and may behave differently by mode, policy, or device.

• AI mode (Use AI Features on). Public Whispers, posts, comments and other public content are checked by AI (OpenAI's moderation API and Cloudflare) before they appear. Flagged items may also be analysed by an OpenAI text model to help our moderators.
• Human Community mode (Use AI Features off) and Local Only. People review your public content. It isn't sent to AI moderation vendors. For now, images you post publicly and checks during public video calls and Lives may still use automated classifiers.
• Private Whispers and private chats, in every mode. Private Whispers and private network chats are never sent to AI, in either mode, even when they're reported. We don't review them or run AI checks on them, and they aren't held for approval. A staff moderator sees one only if someone in the conversation reports it.
• Child-safety reports go straight to our staff for human review and are never sent to AI.

• OpenAI — text and embedding models, content moderation.
• Cloudflare Workers AI — edge inference for content moderation escalation and Akashic Engine AI Mode recommendations.
• On-device models. In Local Only mode, Ask Cosmos and Tarot private chat can run a small open-weight model (Llama 3.2 1B, via WebLLM) in your browser on supported devices, and Apple on-device Foundation Models on supported iOS devices. The model runs on your device, not on a cloud AI provider.
• Meshy — third-party 3D generation provider used for avatar and asset creation.
• RoxyAPI — calculations and AI-assisted interpretive text for Cosmos Tools (see section 1a).
• Voice input (speech-to-text) and read-aloud (text-to-speech) use OpenAI models. Some of these requests reach OpenAI through Replit's AI Integrations service.

In Settings, users can choose an AI execution mode: auto, prefer_local, prefer_cloud, or local_only. This choice changes where eligible AI generation and assistance requests are routed.

• Auto balances local and cloud routing by capability and policy.
• Prefer Local attempts local inference first, including Apple on-device Foundation Models where available, then falls back when local support is unavailable.
• Prefer Cloud prioritises cloud routing for broader model capability.
• Local Only blocks cloud routing for supported generation paths; some features may be unavailable when local capability is not yet supported.

Moderation and safety systems remain active in all execution modes. If you choose Local Only, your public content is reviewed by people instead of cloud moderation providers, the same as in Human Community mode (see section 1c). For now, images you post publicly and checks during public video calls and Lives may still use automated classifiers.

We choose AI providers and settings that limit how they can use your content:

• OpenAI: under OpenAI's API terms, OpenAI doesn't use data sent through its API to train its models unless the customer opts in. We haven't opted in.
• Cloudflare Workers AI: Cloudflare states that it doesn't use customer content sent to Workers AI to train AI models or to improve Cloudflare or third-party services.
• RoxyAPI: RoxyAPI's data processing terms state that it doesn't use personal data submitted through its API to train, fine-tune or evaluate machine-learning models.
• Meshy: content you submit for 3D avatar generation is processed by Meshy under its terms. We don't currently have a no-training commitment from Meshy, so avoid putting personal details into avatar prompts or reference images you don't want shared with it.
• On-device models run on your device and don't send your prompt to a model provider.

In Settings → AI & Privacy (and during onboarding), you can turn off Use AI Features and choose the Human Community experience. When AI features are off:

• Eligible content you create is not sent to AI models for mood, intention, discovery, or generative helpers.
• Mood and intention proposals publish with your post but stay pending until a community reviewer or admin approves them.
• Your public content is reviewed by people instead of AI moderation. Private Whispers and chats are never sent to AI. Child-safety and legal routes remain available. For now, images you post publicly and checks during public video calls and Lives may still use automated classifiers.
• Avatar Studio still works without generative Meshy: pick a shared default model or upload your own .glb.
• Cosmos Tools stay available for calculations, Tarot draws, place search, and saved history. Ask Cosmos, combined Tarot readings, text-to-speech, and voice input stay off. Stored AI text is omitted from history and export until you turn Use AI Features back on.

You can change this preference anytime in Settings. Turning AI features back on restores the AI-Assisted pathways for new content.

California Assembly Bill 2013 requires generative AI developers to publish information about the data used to train their models. Ascended Social does not train its own foundation models on user content. Where we fine-tune or evaluate a vendor model on data we control, we use only:

• Public-domain, openly licensed, or properly licensed datasets.
• Aggregated, de-identified product telemetry to evaluate moderation accuracy.
• Synthetic data generated by us or our vendors for benchmarking.

We will publish a more detailed training-data summary in this document as our use of fine-tuned models expands.

Subject to our Terms of Service, you own the AI-assisted content you create on Ascended Social to the extent permitted by applicable law. AI-generated outputs may not be copyrightable in some jurisdictions; we do not represent or warrant that you can secure exclusive rights in any particular AI output.

Where you interact with an AI system on Ascended Social, we tell you. Where AI-generated text, images, audio, or video are produced for or shown to you, we label them or otherwise make their AI origin clear in context. We do not deploy AI systems intended for social scoring, biometric categorisation of natural persons, or other purposes prohibited under the EU AI Act.

• You can opt out of personalised AI recommendations from the in-app Privacy Center.
• You can ask us not to process your content for evaluation or fine-tuning purposes. Email [email protected].
• You can submit a rights request to access, correct, or delete content used in any AI-assisted feature.

AI features can be wrong, biased, or hallucinate. They are not a substitute for licensed medical, psychiatric, legal, or financial professionals. See the Spiritual & Wellness Disclaimer for more.

Aether Guides are AI agents for Ascended Social available as a premium preview feature for eligible paid tiers. Aether Guide usage consumes AI credits and is protected by subscription gating, tier limits, and moderation controls.

• Access is limited to paid subscriptions (currently Mystic and Ascended tiers).
• Rate and usage limits vary by tier and may change during Preview.
• Aether Guide actions are automatically screened by our moderation systems before publishing; actions flagged as higher-risk are escalated for human review before they take effect. Published output is transparently labeled.