Methodology
How WriteSignal actually works.
No black boxes. Every number on the results page comes from the formulas below.
How scoring works
Every analysis runs entirely in your browser. Nothing is uploaded. We extract six interpretable style features from your text:
- Sentence rhythm variety
Coefficient of variation of per-sentence word counts. AI drafts often keep a steady beat.
- Average sentence length
Mean words per sentence.
- Vocabulary richness
Unique words divided by the square root of total words (root-TTR).
- Stock AI phrases
Density of a curated list of LLM tell-tales (delve into, underscores, a tapestry of, and 30 more) per 100 words.
- Formal transitions
Density of scaffold openers per sentence (furthermore, moreover, in conclusion).
- Personal voice
Questions, exclamations, parentheticals, first-person, and hedges per 100 words.
Those features feed a logistic regression trained on the HC3 dataset (250 human answers, 250 ChatGPT answers). Held-out test AUC: 0.855, accuracy 76%. The heuristic score is a separate weighted anchor score you can tune yourself in Settings.
Deep analysis (optional)
Turn on "Deep analysis" and WriteSignal downloads a small language model (Xenova/distilgpt2, about 90 MB, cached after first run) and runs it in your browser. It computes two extra features:
- Perplexity. Mean next-token cross-entropy from the language model, exponentiated. Lower means the text was more predictable to the model, which is more AI-typical. Calibrated to this in-browser model, AI writing usually lands under about 55 and human writing over about 70; the midpoint sits near 60.
- Burstiness. Standard deviation divided by mean of per-sentence perplexity. It is a weak, noisy signal on short passages, so WriteSignal shows it for context but does not use it in the score.
With deep analysis on, WriteSignal blends the style-only model with the calibrated perplexity signal (a weighted average, style-weighted). Perplexity refines the verdict — it can nudge a borderline read in either direction — but it can never single-handedly flip a text that is clearly human by style and by perplexity into a strong-AI verdict. The blend is deliberately conservative because perplexity alone is a weaker separator than the style features, especially on fluent, modern-LLM prose.
Paraphrase signature
When the style-only score is above 60% but perplexity is above 55, WriteSignal shows a distinct callout. This mismatch (AI-looking style, human-looking unpredictability) is a known signature of AI text that was run through a paraphrasing tool. Treat those results with extra caution.
Model limitations
Known false positives. Formal academic prose, technical writing, and non-native English writers can look "AI-styled" to any style detector including this one. That is a well-documented failure mode of the whole category.
Training bias. HC3 is English, mid-2023 ChatGPT, mostly question-answering. Prose in other genres or later model generations may score differently.
Length matters. Below ~40 words, none of these features are stable. WriteSignal blocks analysis of shorter text on purpose.
Adversarial rewrites. Paraphrasers reduce perplexity signal; human editing of AI drafts flattens most style tells.
Responsible use
No AI detector, including this one, is proof. WriteSignal is a conversation starter and a coaching tool.
- For writers: use it to see where your draft sounds generic, then rewrite in your own voice.
- For educators: treat scores as evidence to prompt a supportive conversation, never as sole grounds for an accusation. Weigh rubric evidence, drafts, and process artifacts above any style score.
- For everyone: results describe the text, not the person.