AI Policy
Last updated: 18/07/2026
Scry uses AI to save you from manual note-taking, not to write your campaign for you. This page explains exactly what AI does, which models power it, what data reaches them, and what stays entirely under your control.
The short version: AI suggests, you decide. Scry organizes and surfaces what you and your table already created — it doesn't invent canon, characters, or plot on your behalf. Nothing an AI feature drafts is written to your wiki without your approval.
What We Use AI For
| What happens | Model / provider | Thinking mode |
|---|---|---|
| Transcribing session voice audio | Groq (Whisper large-v3-turbo) | — |
| Identifying who's speaking, filtering out chatter | DeepSeek V4 Flash | No |
| Summarizing sessions into wiki-ready prose | DeepSeek V4 Flash | No (Basic) / Yes (Pro+) |
| Merging per-player summaries into one session recap | DeepSeek V4 Flash | No |
| Extracting named entities (NPCs, locations, items, etc.) from sessions and uploaded PDFs | DeepSeek V4 Flash | No |
| Connecting new entities into your existing wiki | DeepSeek V4 Flash | No |
| Unseen Servant (wiki Q&A chat) | DeepSeek V4 Flash | No |
| DM Assistant (proposes wiki edits for your approval) | DeepSeek V4 Flash | Yes |
Prep Assistant (/prep) | DeepSeek V4 Flash | Yes |
| Picking a highlight line for Discord recap embeds | DeepSeek V4 Flash | No |
| Generating search embeddings for the wiki and knowledge graph | OpenAI (text-embedding-3-small) | — |
A note on DeepSeek specifically. We use DeepSeek's AI technology to power most of the pipeline above — it's fast, capable, and inexpensive, which is a meaningful part of how Scry stays affordable. That processing currently happens on infrastructure in mainland China, and unlike our other providers, DeepSeek doesn't yet offer a GDPR data processing agreement, Standard Contractual Clauses, or an EU adequacy decision covering this. If you're in the EU, EEA, or UK, that's worth knowing before you use the Unseen Servant, DM Assistant, or Prep Assistant specifically — session recording and transcription don't touch DeepSeek and aren't affected. Full detail is in our Privacy Policy.
We do not generate AI images. The Unseen Servant mascot is a fixed brand illustration, not something generated at runtime. Nothing on the Service — recap posts, entity portraits, marketing pages — uses AI-generated imagery.
What We Never Do
- We never use your session content, wiki, or uploaded PDFs to train DeepSeek's, OpenAI's, or Groq's underlying models.
- We never sell your data to anyone, for any purpose.
- We never write to your wiki without your approval. DM Assistant proposes; you approve, edit, or dismiss each suggestion individually — nothing commits automatically unless you've explicitly turned on auto-approve for high-confidence suggestions.
- We never invent canon. Every AI feature is constrained to organize and cite what your table actually said or wrote — Prep Assistant's suggested hooks, for example, require at least two references to entities that already exist in your wiki before it can propose one.
Per-Feature Controls
| Feature | Can you turn it off? | What triggers it |
|---|---|---|
| Recording, transcription, entity extraction, summarization | Core to the product — this is what builds your wiki | You start a recording with /record start |
| Unseen Servant (wiki Q&A) | Nothing runs unless invoked | You or a player asks it a question |
| DM Assistant | Pro+ only; every proposal needs your approval unless you enable auto-approve for high-confidence suggestions (Campaign Settings) | You open the DM Assistant panel |
| Prep Assistant | Pro+ only; nothing runs unless invoked | You run /prep |
| Discord recap posts | Off by default — you opt in per server | /scry recap-channel |
| AI self-improvement (learning from your corrections) | Yes — Campaign Settings → AI Learnings → master toggle | Automatic, if enabled |
| Voice recording of any individual participant | Yes, always — per-person, per-server, revocable any time | /scry consent grant / withdraw, or reacting to the in-channel notice |
Data Retention
- Raw session audio: never kept longer than 72 hours under any circumstances — this is a hard storage-level limit, not a plan-dependent setting.
- Transcripts and wiki content: kept for as long as your account and campaign stay active.
- Uploaded PDFs: deleted immediately after processing, backed by an automatic 24-hour deletion policy regardless of outcome. We never keep your uploaded sourcebooks around.
- Search embeddings: deleted the moment the entity they're linked to is deleted.
Full detail, including how consent and deletion requests work, is in our Privacy Policy.
Your Right to Leave
Your wiki isn't a hostage. You can export your entire campaign — every entity, every session, all your wiki links — as an Obsidian-compatible vault at any time, from any plan. If you cancel, your data is retained for a window after cancellation so you have time to export it, then deleted. See our Terms of Service and Privacy Policy for the specifics.
Questions about how AI is used in Scry: peter@getscrying.com