Interpretable AI-writing analysis
Evidence,
not verdicts.
WriteSignal shows the exact patterns that make writing sound AI-generated, then helps you fix them. Proofread, strengthen, humanize, and check for AI, all in one editor.
No AI detector is proof. We show evidence you can read, argue with, and teach from, never black-box verdicts.
The report recieved strong feedback. It is important to note that the team delved into the data and utilized every insight.
Try it, right here
A working analyzer, on the landing page.
Edit the paragraph. The scores update as you type. Nothing is uploaded, nothing is saved. This is the same analyzer you'll use inside the app, in a stripped-down form.
For the full experience, per-sentence flags, coaching, humanizer, PDF export, comparison, and history, create a free account.
Style-only model (AUC 0.855). Enable Deep analysis in-app to blend in a calibrated in-browser perplexity signal.
How the model works
Six interpretable features. One honest model.
No opaque neural verdict. Every feature is a number you can inspect, dispute, and reproduce. The trained model is a plain logistic regression fit on the HC3 dataset (250 human answers, 250 ChatGPT answers). Held-out AUC 0.855.
Turn on Deep analysis and a small language model runs in your browser to compute perplexity. WriteSignal blends that calibrated signal with the style model to refine the read — perplexity nudges the verdict, it never overrides it.
Full methodology- 01Sentence rhythm variety
Coefficient of variation of per-sentence length. AI keeps a steady beat.
- 02Average sentence length
Mean words per sentence. AI drafts skew long.
- 03Vocabulary richness
Unique words over sqrt of total (root-TTR).
- 04Stock AI phrases
35+ curated tells: delve into, underscores, a tapestry of.
- 05Formal transitions
Scaffold openers per sentence: furthermore, moreover.
- 06Personal voice
First-person, hedges, questions, parentheticals per 100 words.
- 07Perplexity (deep only)
Mean cross-entropy from an in-browser language model.
- 08Burstiness (deep only)
Sentence-to-sentence variation in perplexity.
For writers and students
Sound like yourself, not like a model.
Writer mode surfaces every sentence that reads AI-typical and offers a concrete rewrite. Vary your rhythm, swap the stock verb, put your stance back in.
- › Per-sentence coaching, not just a score.
- › Rewrite, humanize, and clean text in one click.
- › Compare drafts and export a clean PDF.
For educators
A fair conversation, not an accusation.
Educator mode reframes results as evidence and ships with an integrity banner, talking points for a supportive student conversation, and a batch queue with review status.
- › CSV or file upload for a whole class.
- › Sortable review list with queued / analyzed / reviewed.
- › Every score records the exact settings used.
Integrity commitment
"Style signals are not proof. Formal academic prose and non-native English writers trigger false positives. Use this to start a conversation, never as sole grounds for an accusation."