Definitions, not answers to questions — if you are looking for how ScribeLens works or what it costs, the FAQ covers that. AI-writing detection has its own vocabulary, and the terms get thrown around loosely. This glossary defines the core concepts in plain language — what a score actually measures, where detectors go wrong, and what a fair result looks like — so you can read any detection report, including ours, with a clearer eye.
An AI detection score is a number or grade a detector assigns to a piece of text, describing how closely its measurable patterns — word choice, sentence rhythm, and structural predictability — resemble text generated by a language model. A score describes statistical similarity to AI-typical writing, not authorship: it does not identify who wrote the text, only how machine-typical its patterns look. Reading the evidence behind a score matters more than the score itself.
A false positive is a detector flagging human-written text as AI-generated. It happens most often with formal, heavily-edited, or highly structured writing, since academic prose, technical documentation, and non-native English usage all share surface features — evenness, predictability, formulaic phrasing — with AI-typical text. A false positive is a known limitation of pattern-based detection generally, not evidence that the writer did anything wrong.
A false negative is AI-generated or AI-assisted text that a detector does not flag. This happens with lightly-edited AI drafts, heavily paraphrased AI output, or hybrid writing where a person substantially reworked machine-generated content. Because both false positives and false negatives occur, a single score should never be treated as a final verdict on its own.
Sentence-level evidence is a breakdown that classifies each sentence in a document individually — for example as human-like, mixed, or AI-like — instead of collapsing an entire document into one score. It lets a reader see exactly where a document's overall signal comes from and check it against the actual text. ScribeLens reports every scan this way, highlighting flagged sentences directly in the document rather than reporting a single number.
Reason keywords are short labels attached to a flagged sentence that name the specific pattern behind the flag, such as uniform sentence length, predictable phrasing, or low lexical variation, instead of leaving the reader to infer why a sentence was marked. They turn a verdict into something a person can independently check against the sentence itself. ScribeLens attaches reason keywords to every flagged sentence in a report.
Probability vs proof is the distinction between what a detection result can and cannot establish. A high score indicates that a passage shares statistical patterns with AI-generated writing — a probability — not that a specific person did or did not write it, which would be proof of authorship. Treating a detector's output as proof, in either direction, is a misuse of what the underlying measurement supports; it is evidence for a human to weigh, not a standalone verdict.
Perplexity measures how predictable a piece of text is to a language model — essentially, how "surprised" the model is by each word choice given the words before it. Text made of common, expected words in expected sequences has low perplexity, a pattern strongly associated with AI-generated writing, while text with more unusual or varied word choices scores higher. Perplexity is one of several statistical signals detectors draw on, not a measurement used in isolation.
Burstiness describes how much sentence length and structure vary across a piece of writing. Human writing tends to mix long and short sentences unevenly, described as "bursty," while AI-generated text often maintains more uniform rhythm and structure from paragraph to paragraph. Like perplexity, burstiness is one signal among several, since very uniform or very uneven writing can also occur naturally in human prose.
An AI humanizer is software that rewrites AI-generated text to alter its statistical fingerprint, typically by adding sentence-length variation, swapping in less common words, or introducing small irregularities, with the goal of making the output read as less AI-typical. The underlying content, argument structure, and reasoning of a passage usually remain unchanged by this kind of rewriting, which is why detectable patterns often persist even after a passage has been run through a humanizer.
Academic back matter is the non-argumentative part of a scholarly document: reference lists and bibliographies, in-text citation apparatus, DOI/ISSN lines, copyright and licence text, journal running headers, and declaration sections such as funding, conflicts of interest, and ethics statements. This text is the most formulaic in any paper by construction, so scoring it inflates AI-like signals on genuine scholarship. ScribeLens detects and excludes academic back matter automatically before scoring, while still displaying it in the report for transparency.
A graded verdict presents a detection result as one of several descriptive categories — for example likely human-written, mixed signals, a blend of human and AI writing, substantial AI involvement, or strong AI-like signal — rather than collapsing it into a single pass/fail label or bare percentage. Grading better reflects that AI involvement in a document exists on a spectrum, from none, to light editing assistance, to fully generated text.
The base rate is how common a condition already is in a population before any test is applied — for instance, what share of submitted essays actually contain undisclosed AI writing. Base rate matters because even a well-calibrated test produces more false positives in absolute terms when the condition it looks for is rare, which is why a single flagged result should prompt review rather than an automatic conclusion.
Non-native and ESL writers face a disproportionate risk of false positives because instructional English writing conventions — simpler sentence structures, more common vocabulary, and formulaic transitions — overlap with patterns associated with AI-generated text. This is a documented limitation of pattern-based detection generally, not evidence that non-native writers use AI more often, and it is a strong argument for reviewing sentence-level evidence rather than trusting a score alone.
AI disclosure is a statement, usually attached to a piece of academic or professional writing, that explains whether and how generative AI tools were used — for drafting, editing, brainstorming, or research — and to what extent. Clear disclosure norms are becoming standard at many institutions and publications as a way to make AI assistance transparent up front, rather than something to conceal or later have to defend.
Yes. Every definition on this page is free to read, and so is a basic check of your own writing: ScribeLens offers a free guest scan with no signup required, up to 2,000 words per check.
No. A score reflects how closely a text's statistical patterns resemble AI-generated writing — that's probability, not proof. Treat it as evidence for review alongside context like drafts and revision history, never as a standalone verdict.
Reference lists and other academic back matter are the most formulaic text in any paper by construction, so scoring them inflates AI-like signals in genuine scholarship. ScribeLens detects and excludes reference lists, citation apparatus, and declaration sections automatically before scoring.
A false positive flags human-written text as AI-generated; a false negative misses AI-generated or AI-assisted text and lets it read as human. Both happen, which is why sentence-level evidence matters more than trusting a single number.
Run any text through ScribeLens and see sentence-level evidence, reason keywords, and a graded verdict for yourself. Free, no signup required.