AI Detector

How this AI detector's scoring actually works

This tool combines four separate signals into a single 0-100 score, all calculated from the text itself with no external service involved. First, it measures "burstiness" the natural variation in sentence length. Human writing tends to mix short, punchy sentences with longer, more complex ones, while AI-generated text often produces sentences of more uniform length; the tool calculates the statistical variation across your sentence lengths and scores unusually low variation higher. Second, it checks for a list of roughly three dozen phrases that show up disproportionately often in AI-generated writing things like "delve into," "tapestry," "in today's world," "leverage," and "it's important to note." Third, it looks at how often multiple sentences start with the same word, since repetitive sentence openers are a common AI writing tic. Fourth, for passages of 150 words or more, it measures vocabulary diversity (the ratio of unique words to total words) a narrower vocabulary contributes to a higher score.

What the three result categories mean

A score below 35 is labeled "Likely Human-Written" few or none of the four signals triggered. A score between 35 and 65 shows as "Mixed Signals / Uncertain," meaning some patterns are present but not conclusively. A score of 65 or above shows as "Patterns Consistent with AI," meaning multiple signals lined up at once. None of these labels are a verdict they describe how closely your text's statistical fingerprint resembles typical AI output, nothing more.

Why isn't any AI detector free or paid 100% accurate?

Paid, professional-grade detectors use machine learning classifiers trained on massive datasets of known human and AI text, and often analyze "perplexity" (how predictable each word choice is to a language model) in addition to burstiness. Even those tools have documented false-positive and false-negative rates. This tool uses transparent, rule-based heuristics rather than a trained model, which makes it fast, free, and fully private, but also means it's working with less nuance a human writer who favors formal, structured prose can score as "AI-like," while AI text that's been edited or run through a paraphrasing tool can score as "human." Treat the result as a data point, never as proof.

When a tool like this is actually useful

It's most useful as a quick sanity check flagging text that might be worth a closer look, or reassuring you that a draft doesn't read as obviously AI-generated before you publish it. It's not appropriate as the sole basis for an academic integrity case, a hiring decision, or any other high-stakes call, since both false positives (flagging genuine human writing) and false negatives (missing lightly-edited AI text) are common with any detector, paid or free.

FAQ

Can this be used to catch students cheating?
We don't recommend using any free AI content detector as the sole basis for an academic integrity decision. False positives happen, especially with formal or heavily edited writing, and treating a score as proof rather than a signal can lead to unfair outcomes.

How much text do I need?
At least 40 words, though longer passages (150+ words) give a more reliable read since there's more sentence and vocabulary data to work with the vocabulary-diversity signal specifically only activates above 150 words.

Is my text stored or shared?
No. Everything runs locally in your browser using JavaScript your text is never uploaded or saved anywhere.

What do the three result categories mean?
"Likely Human-Written" (below 35) means few AI-associated patterns showed up, "Mixed Signals / Uncertain" (35-65) means some did but not decisively, and "Patterns Consistent with AI" (65+) means several signals lined up together. All three are probability-based descriptions, not certainties.

Can editing or paraphrasing AI text fool the detector?
Often, yes. Rewriting AI output in your own words, varying sentence length manually, or running it through a paraphrasing tool can lower the score significantly, since the tool is only measuring surface-level patterns rather than the actual origin of the text.

Does it work on languages other than English?
Partially. The sentence-length and vocabulary-diversity signals work in any language, but the AI-phrase list is English-specific, so text in other languages will only be scored on two of the four signals rather than all four making the result less reliable than for English text.