About the site
Editorial policy
This page explains how information gets onto Uncensored AI Atlas and how we keep it accurate. It applies to every page, including the ones about our commercial partner.
Sources we use
| Information | Primary source | Not used |
|---|---|---|
| Model method, capabilities, claims | The model’s own Hugging Face card | Third-party listicles, social media posts |
| Parameters, context, architecture | config.json and the Hugging Face model API | Estimates or rounded figures from elsewhere |
| License | The license tag, plus the license text | Our own reading of an unlinked license |
| Prices and limits | The provider’s pricing page or public model list | Prices quoted by other sites |
| Concepts and methods | Papers, official documentation, original repositories | Unsourced blog explanations |
Every external source is linked and was checked to load on the date shown. Model, price and provider records carry the date they were checked against their sources.
Four kinds of statement
We label what kind of statement you are reading:
- Measured: we ran it under a published protocol. As of today there are none.
- Sourced: read from a primary source we link.
- Claimed: asserted by a creator or provider and not checked by us. We write “the creator reports…” and link the source.
- Inference: our reasoning from sourced facts, written as such (“this suggests…”).
Missing information stays missing. We show “Not stated” or “Not applicable” instead of a zero, a blank or a guess, and calculators refuse to produce a total when an input they need is missing.
How models and providers are chosen
A model enters the catalogue when its card and metadata are publicly readable and its card names the base model it was derived from. We prefer the original release to re-uploads, and link quantized copies under the original. A provider is listed when it publishes the models it serves and a per-token price. Neither is an endorsement, and neither can be bought.
Rankings and recommendations
We do not rank models without measurements. Editorial picks on Best uncensored AI state their criteria at the top, are labelled as qualitative, and include the strongest argument against each pick. When we publish measured results, the page will say which statements are measured.
Use of AI tools
Drafts of pages on this site are written with the help of AI tools, working from the sources listed on each page. Facts, numbers and links are checked against those sources before publishing, and code in the guides is run before it is published; each guide states what it was run against. The site owner approves every release before it goes live. Bylines read “the Atlas editors” because we do not attribute pages to people who did not write or review them.
Commercial relationship
The site has a commercial relationship with MolochAI and carries a paid advertisement for TartarusAI. Our rules for both are in the partner and advertising disclosure. In short: partner placements and the advertisement are labelled, partner rows follow the same data standards as others, links to both carry rel="sponsored", and neither reviews or approves content. Paying for an advertisement does not get anything into our tables, picks or articles.
Content boundaries
We explain how refusal-reduction techniques work at a conceptual level. We do not publish step-by-step procedures, code or parameters for removing safety behaviour from a model, prompts designed to get around a service’s safety systems, or examples of harmful use. Model cards we link may contain more; we link them because they are the primary source for each model.
Dates
“Published” is the day a page first went live. “Updated” changes only for substantive edits, not for typo fixes. “Facts checked” is the day the page’s data was last compared with its sources.
Corrections
Report errors through the corrections page. Material corrections are logged there with the date and what changed.