Reading a confidence score without over-reading it

Two numbers, two questions

The AI likelihood score answers: how strongly does this text match patterns associated with AI generation? The confidence level answers a different question entirely: how much should you trust that first answer, for this particular piece of text? They are independent. A long, well-structured essay scoring 88% carries high confidence. A 320-word fragment scoring 88% might carry low confidence — same score, radically different weight. Reporting the first without the second is the single most common way detection results get misused, because a bare number reads as settled whether or not it is.

What raises confidence

Confidence rises when the measurement conditions are good. Four things matter most. Length. Detection is statistical, and statistics need samples. Above roughly 600 words the estimate stabilises; between 300 and 600 it is usable but noisier; below 300 Bypass AI does not scan at all. Distance from the middle. A score of 92% or 8% sits well clear of the decision boundary. A score of 52% sits on it, where small changes in input flip the result. Signal agreement. If predictable phrasing, sentence uniformity and low specificity all point the same way, that is three independent measurements converging. If two say AI-like and two say human-like, the total is an average of a disagreement. Structure. Continuous prose in real paragraphs gives the model context to work with. Lists, stacked headings, keyword strings and OCR noise do not.

What lowers it

The mirror image, and each one is worth recognising in the wild. A score between 41% and 59% is the classic low-confidence case — the model is genuinely undecided, and presenting that as a result is misleading regardless of how it is worded. Contradictory signals mean the text has properties of both classes, which frequently happens with human writing that has been edited by a tool, or AI drafts that have been substantially revised. Broken formatting — OCR errors, PDF extraction artefacts, mixed content pasted from several sources — corrupts the features the model reads. Template-shaped content like email, marketing copy and abstracts is inherently regular, so it looks AI-like for reasons that have nothing to do with authorship. And very few sentences, or mostly sentence fragments, leaves nothing to measure variation across.

How Bypass AI words each level

The app shows fixed wording per level rather than generating a new explanation each time. The phrasing is deliberately flat — no escalation, no alarm.

1. 1High — "High confidence · Clear writing patterns found."
2. 2Medium — "Medium confidence · Some patterns are clear, but results may vary."
3. 3Low — "Low confidence · The writing patterns are not clear enough for a strong result."

What to do at each level

High confidence with a high score: the text does strongly resemble generated writing. That is still not proof of authorship, but it is a reasonable basis for asking questions — and for a writer, a clear signal that an editor's detector will likely read it the same way. High confidence with a low score: the most genuinely reassuring combination, and the one worth keeping a record of if you expect to be questioned. Medium confidence: treat the score as directional. Useful for deciding whether to revise your own draft; not sufficient for a claim about someone else. Low confidence: treat as no result. Not a weak yes — no result. If you need an answer, supply better input: more text, cleaner formatting, continuous prose rather than a list. If you cannot, the honest conclusion is that this text cannot be assessed, and acting as though a 55% meant something is how people get wronged.

The uncomfortable implication

Taking confidence seriously means accepting that a meaningful share of real-world submissions cannot be assessed at all. Short answers, heavily-formatted documents, translated text, anything under a few hundred words — the honest output is 'unknown'. Most detectors do not surface this, because 'we cannot tell' is a poor product demo. But a tool that returns a confident number on input it cannot actually evaluate is not more capable than one that declines. It is just less willing to say so, and the cost of that unwillingness lands on whoever is on the wrong end of the number.