Under the hood
How SentenceBase reads your writing
No black box and no verdicts. SentenceBase shows you which phrases make text sound automated, so you can decide what to change.
From robotic to real in three steps
- 1
Paste your text
An essay, blog post, cover letter or email. As many documents as you like.
- 2
See the tells
Every AI-sounding phrase is highlighted, counted and ranked, with an overall score.
- 3
Rewrite in your voice
Swap clichés for plain, natural wording and re-check until it sounds like you.
Built to fix, not to accuse
Phrase-level highlighting
See exactly where the robotic phrasing is, not just a vague percentage. Every flagged phrase shows how often it appears.
Word credits
Unlimited documents. You only spend credits on the words you check.
Private by design
Your text is analyzed for your result and never used to train models.
Rewrite suggestions
Coming soonOne-click, plain-English alternatives for every flagged phrase, in the tone you pick.
The method
SentenceBase uses two models. A phrase model finds the stock phrases AI tools lean on, and those are the highlights you see. The score itself comes from a second model that reads the writing style: it looks at short letter patterns across the whole text and combines them into a 0 to 100 chance that the text was written by AI.
The style model was trained on AI-written text from ChatGPT, Gemini and Claude (including essays and academic writing) and on human text written before about 2015, such as old blogs, forum posts, news, Wikipedia and research papers. Older text is used for the human side so it cannot be AI-written.
How accurate is it?
On held-out test texts the model caught about 87% of AI-written samples and wrongly flagged about 2% of human ones. Those figures come from our own test sets, not from independent testing, and they are lower on some kinds of writing, so treat them as a direction, not a guarantee. It does worse on very short text (under about 60 words), AI text that has been rewritten to hide it, very casual writing and non-native English, and it will need retraining as new AI models appear.
No AI detector is fully reliable. A score should never be the only evidence that someone used AI, and it must not be used to accuse anyone.
Privacy
Text you paste is analysed to produce your result. It is not stored and is not used to train models. Files you upload for review are different: we keep the file, its text, the result and your IP address for 30 days, then delete them. The phrase-swap suggestions run in your browser. If you use the AI rewrite, your text is sent to an outside AI provider to be rewritten.