We split text into small blocks and score each with a supervised classifier trained on real vs AI writing.
Paragraph colors reflect the weighted mix of block scores inside each paragraph.
If ML isn’t available, we fall back to style metrics like burstiness and repetition.
The short version
AuthentiScan doesn’t “prove” authorship. Instead, it looks for statistical patterns that are more common in AI-written text than in human writing. Our ML block classifier scores short spans of text, and the app summarizes those scores into an overall estimate and a paragraph heatmap. Use the results together with context and source credibility.
Step 1 — Split into blocks
We tokenize the document and slice it into small, fixed-size blocks. Short blocks help the model focus on local patterns (tone, cadence, connective language) without being thrown off by very long paragraphs or formatting quirks.
Step 2 — Score with ML
Each block goes through a supervised classifier trained on a mix of real and AI-generated samples. The model learns the kinds of features that often distinguish AI writing—things like unusually even sentence rhythm, over-use of connective phrases, hedging, and template-like phrasing.
A block score is a probability-like value: higher means “more AI-like.” It’s not a certainty, just the model’s estimate based on patterns it has seen.
We calibrate scores on a validation set so that, in aggregate, similar values mean similar risk across different kinds of text.
Step 3 — Paragraph heatmap
To make results readable, we map block scores back to paragraphs. If a paragraph contains several high-scoring blocks, its tint trends warmer. The slider in the app lets you set a minimum score and optionally show only “suspicious” paragraphs (≥ 60%).
- Green: low AI-likelihood
- Amber/Orange: medium risk, worth a closer look
- Red: strong AI-like signals
Step 4 — Document summary
Finally, we calculate an estimated AI-written fraction (the share of blocks at or above the threshold) and an overall score that blends block scores and confidence. These appear alongside the heatmap and per-signal breakdown.
Why not only heuristics?
Older detectors relied on simple style rules (“burstiness,” type–token ratio, etc.). Those are helpful, but easy for modern models to dodge. Our ML approach still uses stylistic cues—but it learns how they interact, which improves accuracy on subtle edits or mixed-authorship documents. When the ML model isn’t available, we clearly label the heuristic fallback.
Interpretation tips
Scores guide your review. Combine them with context, drafting history, and any available provenance.
It’s normal to see some paragraphs score higher than others if edits came from multiple sources.
Very short inputs don’t give the model much to work with—treat results as low-confidence.
Editing for clarity can raise some signals. Look for clusters, not single outliers.
Limitations
- Paraphrasers and heavy manual edits can mask AI patterns.
- Very formulaic human writing (templates, corporate style guides) can look AI-like.
- Language, genre, and topic can affect scores; we continually expand training data.
FAQ
We process inputs to compute scores. See our Privacy page for current retention settings.
We display paragraphs ≥ 60% as “suspicious” by default. You can adjust this in the UI.
We train on a mixture of popular systems and human corpora, and refresh the data over time.
No—use it as a decision aid, not a verdict. Always combine with additional evidence.