Free Burrows' Delta authorship comparator
Implements the standard Burrows' Delta algorithm for authorship attribution entirely in your browser. Paste two text samples, get back a delta score, calibrated same-author probability, and a breakdown of the function words that drove the verdict. Pure offline JavaScript — nothing leaves your computer.
For investigators, journalists, and researchers: Burrows' Delta is the workhorse of computational stylometry, used to identify Robert Galbraith as J.K. Rowling, attribute parts of Henry VI to Christopher Marlowe, and verify newly-discovered manuscripts. The same algorithm applies to anonymous online posts, threatening letters, and disputed authorship. The 2025–2026 research wave has extended its use to detecting AI-generated text — though as a single signal among many, not as standalone proof.
For research-grade analysis with proper reference corpora and significance testing, see the stylo R package or faststylometry Python library. For forensic stylometry that incorporates additional features (sentence length, punctuation, syntactic patterns), see JStylo.
Frequently asked questions
What is Burrows' Delta?
Burrows' Delta is the gold-standard algorithm for stylometric authorship attribution, introduced by John Burrows in 2002. It compares the relative frequencies of the most frequent words (typically 100–500 function words like the, of, and, to, in) between two texts, normalises each text's frequencies into z-scores against a reference distribution, and reports the mean absolute z-difference. Lower numbers mean more similar styles.
How do I read the score?
The Δ value is on a roughly continuous scale: below ~0.20 typically indicates same author with high confidence, around 0.35 is the boundary region, and above ~0.50 typically indicates different authors. The probability shown above the score is a calibrated logistic function fitted to that range — treat it as a guide, not a definitive answer.
How long do my samples need to be?
The minimum here is 100 words per sample, but reliable results need much more — Burrows' original work used novel-length samples (50,000+ words). For forensic-style attribution, 5,000+ words per side is a reasonable floor. Below that, function-word distributions are too sparse to be stable. The tool will run on shorter input but will warn you in the explanation.
Can it tell human-written from AI-written text?
Yes, in principle. Recent research applies stylometry to detect LLM output: GPT-class models have characteristic function-word distributions that differ from human writers. The catch: light human editing or deliberate paraphrasing significantly degrades detection. For a single signal among many, this works; as standalone proof of AI authorship, it does not. Use alongside dedicated AI-text detectors and content-provenance tools.
Why are some common words missing from the analysis?
The MFW (most frequent words) are taken from the combined corpus of both samples. If a word is rare in both, it can't be in the MFW. The "Top 20 contributors" panel shows which words drove the score the most — these tend to be common function words whose frequencies diverge between the two writers, exactly what stylometry is designed to detect.
How do I defeat stylometric attribution if I'm being doxed?
Style obfuscation by hand is largely effective — research shows that purposeful style change or deliberate imitation of another known author defeats most stylometric methods. Machine-translation round-tripping, contrary to folklore, doesn't work as well — translators preserve enough authorial style to leak through. The defensive guidance from the Whonix project is the canonical reference.
Is the algorithm here the same as in academic stylometry packages?
The core delta calculation matches the standard formula. Two simplifications versus a full implementation: (1) the reference distribution is taken from the average of the two input texts rather than an external corpus, which is the standard approximation when only two texts are being compared; (2) the MFW is built from the combined corpus rather than a separate training corpus. For research-grade work, use stylo (R) or faststylometry (Python) with proper reference corpora.