PrivacyWatch
LLM data privacy, retention, and model training tracker.
Who stores your prompts? Who trains on them? Researched from primary sources — privacy policies, ToS, DPAs, and API docs.
96 provider surfaces across 85 families · all rows re-verified against live sources on 2026-09-30 (a few rows keep earlier verification dates — check each row's sourceDate).
Live: privacywatch.wyrdwerk.com
What this tracks
For each provider surface, we research and document a human-readable summary (label/detail) plus machine-readable structured fields. Every structured value is backed by an evidence[] entry: a verbatim quote from the primary source, its URL, and the retrieval date. silent means we read the provider's documents and they don't address the field — a missing field means not researched yet.
| Field | Structured as | What it means |
|---|
| Training use | training.default, training.optOut | Is your data used to train or fine-tune models? (includes default on/off and how to opt out) |
| Zero Data Retention (ZDR) | zdr.access | Does a ZDR option exist, and what tier unlocks it? |
| Retention duration | retention.kind, retention.days | How long is your data stored? |
| Data location | location.regions | Where is data processed and stored? (ISO country codes, EU, GLOBAL) |
| Compliance | compliance.dpa/soc2/hipaaBaa | DPA availability, SOC 2, HIPAA BAA |
| Incidents | incidents[] | Confirmed breaches, regulatory actions, government bans (with sources) |
| Evidence | evidence[] | Verbatim quote + source URL + retrieval date per field |
| Rating | rating | Overall assessment: Clean, Guarded, Caution, High Risk, or Unverified |
Vocabularies are defined in scripts/lib/v2-fields.mjs, mirrored in providers.schema.json, and enforced by npm run validate.
Providers covered
96 tracked surfaces across 85 provider families in 4 categories:
US Frontier: OpenAI · Anthropic · Google Gemini · xAI / Grok
Chinese: Alibaba / Qwen · Moonshot AI / Kimi · Zhipu AI / GLM · MiniMax · DeepSeek · Xiaomi MiMo
Inference: Fireworks AI · Together AI · DeepInfra · Nebius AI · SiliconFlow · Groq · Cohere · Mistral · Perplexity · Cloudflare · Amazon Bedrock · Azure · AI21 · Cerebras · SambaNova · StepFun · Baidu · ByteDance · Tencent Cloud · Upstage · Meta · Deepgram · Voyage AI · Black Forest Labs · Modal · Baseten · CoreWeave · DigitalOcean · Crusoe · io.net · GMICloud · AtlasCloud · NVIDIA · Reka AI · Thinking Machines · Inception Labs · Sakana AI · Poolside · Relace · Decart · Liquid AI · Runway · Krea · HeyGen · Fish Audio · Sourceful · Chutes · Venice · Mancer · NextBit · AkashML · Phala (Redpill) · Inference.net · Novita AI · FriendliAI · Parasail · Open Inference · Ionstream · Darkbloom · Sail Research · Perceptron · Inceptron · DekaLLM (Cloudeka) · StreamLake · Aion Labs · MARA · Nex AGI · ModelRun (Modular) · Morph
Coding Tools: Cursor · OpenCode · HyperAgent · Wafer AI · Neuralwatt · CommandCode
(Crof AI was removed from the dataset in v1.9.0 as defunct; its snapshot stays in policies/ for the record.)
Provider briefs: every row has its own page at /p/<id> (e.g. /p/openai-api) with a plain-English summary, the evidence quotes and incidents.
Public dataset: providers.json · Schema: providers.schema.json · API: /api/v1 · Licence: CC BY 4.0
Data sources: Primary research from provider privacy policies, ToS, DPAs, and API docs — every structured value carries a verbatim evidence quote with its source URL and retrieval date. Cross-referenced with OpenRouter's curated ToS index. A Cloudflare Worker watches every sourceUrl 3× a day and reports detected changes to issue #2 (see Contributing data updates). Brand marks are local PNGs in assets/logos/, refreshed from Logo.dev via npm run logos (publishable key, never shipped to the browser).
API
Read-only JSON API, no key required. Responses are CORS-enabled and cached for an hour.
Full reference: docs/API.md · rendered page at /api-docs · OpenAPI spec at /api/v1/openapi.json
| Endpoint | Returns |
|---|
GET /api/v1/providers | Summary rows; filter with category, rating, surfaceType, training, zdr, retention, region, maxRetentionDays, incident, ids (compare), fields=full |
GET /api/v1/providers/{id} | One full row, including evidence[] (source URL + verbatim quote + retrieval date) and incidents[] |
GET /api/v1/meta | Version, licence, counts, filter vocabularies, endpoint map |
GET /api/v1/openapi.json | OpenAPI 3.1 spec |
GET /api/search?q= | Semantic search over archived policy text |
Examples:
/api/v1/providers?training=off&zdr=default,self-serve®ion=EU
/api/v1/providers?category=inference&maxRetentionDays=30
/api/v1/providers?ids=openai-api,anthropic-api,google-vertex&fields=full
Filter values: comma-separate or repeat a parameter for OR; different parameters combine with AND. silent means we read the provider's documents and they don't address the field.
Rating system
| Rating | Meaning |
|---|
| 🟢 Clean | Training off by default, ZDR documented, clear policy |
| 🟡 Guarded | Generally safe with documented caveats |
| 🟠 Caution | Training on by default, vague retention, or published pages that stay silent/opaque after a live check (not evidence they train) |
| 🔴 High Risk | Training on with weak opt-out, China storage, confirmed breach or gov bans |
| ⚫ Unverified | We have a document but it never addresses API inputs/outputs at all — not a missing review, and not the same as Caution-for-opacity |
A 🚩 incident flag is additive — it marks confirmed security breaches, regulatory actions, or government bans on top of the base rating.
Precedents applied in the 2026-09-30 audit (recorded in CHANGELOG v1.7.0–v1.12.0):
- "Improve the Services" wording that never names model training → Caution, not High Risk.
- Export-control listings and parent-company content fines go in the row's notes, not the incident flag.
- The incident flag is additive: it never changes the base rating.
- No documented ZDR caps a row at Guarded at best.
- Vague or unbounded retention → Caution.
silent (we read the document and it doesn't say) is not the same as missing (not researched); ⚫ Unverified means the document never addresses API data at all.
Contributing data updates
Provider data lives in providers.json. Each row carries human-readable label/detail text plus structured fields (vocabularies in scripts/lib/v2-fields.mjs, enforced by providers.schema.json and npm run validate). When a policy changes:
- Research from primary sources only — the provider's privacy policy, ToS, DPA, or API docs. Never infer from industry norms.
- Edit both layers: update the relevant
label/detail text and the structured fields (training, retention, zdr, location, compliance, incidents).
- Add evidence: an
evidence[] entry per field you touched — source URL, verbatim quote, and retrieval date (retrieved). A quote may be omitted only when the value is silent. silent = we read the document and it doesn't address the field; a missing field = not researched yet. Never conflate the two.
- Update
sourceDate on the row to the verification date.
- Bump the version:
meta.version in providers.json and version in package.json (they must match — npm run validate checks), update meta.lastUpdated, and log the change in CHANGELOG.md.
- Run the checks:
node --test tests/*.test.mjs, npm run validate (prints structured-field coverage), npm run build — all must pass.
- Open a PR that references the watcher report it processes: "Processes #2 up to comment
<id>".
Watcher / issue #2: the watcher Worker checks every policy URL 3× a day and posts detected changes as comments on issue #2, which it reopens. Maintainers close #2 once a research cycle has processed every comment — a closed issue means nothing is pending.
Each row links to its official sourceUrl on the live dashboard.
Deployment
PrivacyWatch is a static site hosted on Cloudflare Pages. Pushes to main trigger an automatic build (npm ci && npm run build && npm run validate) and deploy the dist/ output. No secrets are stored in this repository.
Caveats
- Not legal advice. This is a good-faith summary of public policy documents.
- Policies change. Always verify with primary sources before making compliance decisions.
- "Unknown" ≠ safe. ⚫ Unverified means the cited document never addresses API data — not that the provider is clean. 🟠 Caution on a silent/opaque host means we looked and they still don’t say; it is not a finding that they train.
- Research dates: all 96 rows were re-verified against live sources on 2026-09-30, except a few rows whose live pages blocked automated fetches or had gone unreachable, so they keep earlier
sourceDates. Always check the per-row sourceDate and the retrieved dates in evidence[] before relying on a claim.
License
- Data (
providers.json, API responses, PrivacyWatch research notes): CC BY 4.0 — credit "PrivacyWatch by WyrdWerk". Quoted policy text and logos belong to their owners. See LICENSE-DATA.md.
- Code: MIT.
Maintained by WyrdWerk
