What our engine is, in plain words

In May 2026 a group at Carnegie Mellon published a paper called "Base Models Look Human To AI Detectors". Its observation was simple: language models exist in two states, the base model that comes out of pretraining and the instruction-tuned version that answers questions politely. Text written by the base version reads differently, and commercial detectors treat it very differently. The authors then built a method, Humanization by Iterative Paraphrasing, that fine-tunes a base model into a paraphraser and applies it several times over. They released the code and the adapters.

Samesay runs that method as published: the Qwen3 base model with the adapter the authors released, the prompt format from their code, their sampling settings, and their number of passes at Max. We did not add prompts, rules or a second model on top. What we added is plumbing: your text is split into paragraphs so each one is rewritten on its own, headings are kept as you wrote them, and several versions run as one batch.

The server is a GPU we rent by the second. Your text is sent there, rewritten in memory, and sent back. Nothing on that server stores it. The website then runs its own checks on the result: the meaning check compares figures, names, negations and qualifiers with your draft, and the style score counts writing habits. Neither check calls any outside service.

We do not promise what any AI detector will say about the result. The paper reports its results on two detectors as of spring 2026; detectors change monthly and a later paper found the same method only partly effective against a newer one. If someone tells you a tool is detector-proof, they are selling something.

What we do promise: your text is not used to train anything, it is not shared with detector companies, and with history off it is never written to our database at all. The code of the engine is public, so anyone can check what it does.