Uncensored AI and privacy go together. If you’re asking a model something a mainstream assistant would refuse, you probably don’t want the conversation stored, reviewed, or used for training. Almost every provider in our index says something like “no logs.”
So we read every privacy policy and scored it against a fixed rubric. The rubric is on the methodology page, and the core rule is simple: only documented claims earn points. If a policy doesn’t say a provider won’t train on your prompts, it gets no credit for not training.
The rubric (100 points)
| Item | Points |
|---|---|
| Prompt retention: none / short-lived / retained | 20 / 10 / 0 |
| Prompts not used for training | 15 |
| Technical guarantee (attested TEE or end-to-end encryption) | 20 |
| Signup: no personal info / email / account / ID | 8 / 5 / 3 / 0 |
| Accepts crypto | 7 |
| Legal jurisdiction disclosed | 10 |
| Upstream hosts disclosed | 10 |
| Independent audit or open source | 10 |
The results
| Provider | Score | Why |
|---|---|---|
| OpenRouter | 65 | No logging by default, no training, named jurisdiction, each request routed to a named upstream |
| DeepInfra | 58 | Memory-only processing, no training, California company, upstream disclosed |
| Hugging Face | 58 | Doesn’t store bodies when routing, no training, named upstreams |
| NextBit | 58 | Zero-retention terms, own EU hardware in Spain, no training |
| Venice.ai | 55 | No logging, no training, Wyoming jurisdiction, crypto accepted |
| Unfil.ai | 50 | Strong claims and no-email signup, but anonymous operator |
| Abliteration.ai | 48 | Transient processing, no training, Delaware company; card-only, account required |
| AionLabs | 48 | No prompt storage, no training, Polish company; metadata kept 12 months |
| Novita AI | 48 | Zero-retention terms, no training by default, Delaware |
| Parasail | 48 | No logging beyond generation, no training, US company |
| DemonRoute | 45 | Strong claims and crypto-only, but unnamed upstreams and no formal policy page |
| HackAIGC | 45 | Strong claims, crypto accepted, but no named company or jurisdiction |
| PayPerQ | 45 | Chats decrypted only in an open-source, attestable enclave; no account needed, Lightning and Monero; training unstated |
| Electron Hub | 40 | No content logging, upstream disclosed; moderation-scans every request; no named entity |
| Infermatic | 40 | No retention, Delaware company, crypto accepted |
| Unjailed AI | 40 | API not stored; web chat expires after 24 hours; Monero accepted |
| Nebius | 38 | Retains prompts by default (opt-out available), no training, EU company |
| Requesty | 38 | Logging on by default for 30 days (can be disabled), no training, UK company |
| Chub AI | 35 | API not logged, no training, crypto only; no named entity |
| NanoGPT | 35 | No retention, no account needed, but training and jurisdiction unstated |
| Uncensored.chat | 35 | 24-hour auto-delete, crypto only, but no privacy policy document |
| Arli AI | 33 | No retention, all inference in-house |
| NovelAI | 33 | Doesn’t store prompts, client-side encrypted stories, Massachusetts company; training unstated |
| NoFilterGPT | 32 | “Purged after receipt” claims, but its policy logs IPs and usage |
| NoTrack | 30 | Policy says chats are kept 30 days; API page says nothing is stored |
| AI Horde | 28 | Open source, but volunteer workers can see prompts |
| DreamGen | 28 | Stores stories, training opt-in, Swiss jurisdiction |
| Perchance | 28 | Not stored server-side, no account, but anonymous operator |
| NoLimitGPT | 25 | Conversations stored to your account; no training |
| Featherless.ai | 23 | Clear no-logging claim; nothing else documented |
| Agnai | 18 | Generic policy: collects content, may use it for research |
| RedModel | 18 | Keeps content as long as useful and trains on it; no account needed for chat, Texas governing law |
| MegaNova | 13 | Collects prompts and outputs, kept “as long as necessary” |
| FreedomGPT | 3 | Collects prompts; “no expectation of privacy” |
| AI/ML API | — | Policy says nothing about prompts |
| Dolphin | — | No privacy policy found |
| Mancer | — | No logging policy published |
Why nobody breaks 70
1. Nobody offers a verifiable guarantee across the board. Every provider’s “no logs” is a promise in a policy document. That’s worth something, but you can’t check it. The 20 rubric points for technical guarantees go to attested confidential computing, where hardware proves what code is running, or end-to-end encryption. Venice offers end-to-end encryption on some models. PayPerQ now decrypts every chat request only inside an enclave running open-source code you can verify, which stops PayPerQ reading your prompts, but the model host behind it still sees them. Nobody protects the prompt all the way to the model on every request, so nobody gets those points yet.
2. “We don’t log” says nothing about upstream. Many uncensored providers don’t own GPUs. They rent from cloud hosts or resell other APIs. A provider can truthfully keep no logs while the GPU host or upstream API does. OpenRouter is the clearest about this: every request names its upstream. Others mention upstream policies only in passing, if at all.
3. Anonymous operators can’t be held to anything. A large share of the index doesn’t name a legal entity or jurisdiction. Two anonymous services, DemonRoute and HackAIGC, score mid-table on the strength of their claims alone. A privacy policy from an anonymous operator is a promise with no one to enforce it. That’s why disclosing a jurisdiction is worth 10 points.
4. Silence earns zero. Featherless has one of the clearest no-logging statements in the market but says nothing about training, jurisdiction, or payment methods. Under a rubric that only rewards what’s written, it scores 23. Providers can raise their scores the cheap way: write it down.
What to actually do
- For sensitive work, prefer providers with named jurisdictions and clear policies. A promise someone can be held to beats an anonymous one.
- Check which model you’re using. At Venice, “Anonymized” frontier models still pass your prompts to OpenAI, Anthropic, or xAI.
- Assume upstream can see it unless the provider says otherwise.
- Never send anything to AI Horde you wouldn’t post publicly. Its own FAQ says the same.
- Watch for end-to-end encrypted options as they spread. That’s the first real step from “trust us” to “verify us.”
We re-check every policy monthly against archived copies. When a provider changes its terms, its score changes, and we’ll say so here.