AI Rewrite & Paraphrase Detector — free AI rewrite, rephrase & text-similarity checker

Lexical similarity measures how many of the same words two texts share. Semantic similarity measures how close they are in meaning, even when the wording is completely different. The two often disagree, and that disagreement is the whole point of comparing two texts properly.

Take "I have no coffee left" and "My cup is already empty". They share zero content words, so lexical overlap is near 0 — yet they mean almost the same thing, so semantic similarity is high. The reverse happens too: "a light lunch" and "a light bulb" share the word light but mean nothing alike.

To measure meaning, the tool converts each sentence into a 512-dimensional vector — an embedding — using the Universal Sentence Encoder. Sentences that mean similar things land close together in that vector space regardless of vocabulary, so a synonym swap barely moves the score. Sentence similarity is the same idea applied one sentence at a time, which is how the tool pinpoints exactly which lines were rephrased. A pure word-matching checker sees paraphrasing as original writing; a semantic checker sees straight through it.

• Embedding cosine (the semantic layer). The angle between the two 512-dimensional meaning vectors. It runs from -1 to 1; for normal prose it sits between 0 and 1, where 1 means the same meaning. This is the metric that catches paraphrasing.
• Lexical cosine. The same angle idea applied to word-frequency counts rather than meaning — high when both texts use the same words in similar proportions.
• Jaccard. Shared unique words divided by total unique words, from 0 to 1. Two sentences with 6 words in common out of 20 distinct words score 0.3.
• Edit distance (Levenshtein). The number of single edits — insert, delete, replace — needed to turn one text into the other, normalised to a 0–1 similarity. It is sensitive to small character-level changes.

Why use all four? Any single metric is easy to game: a synonym swap defeats Jaccard but not embedding cosine, while two unrelated texts on one topic can fool lexical overlap alone. The tool weights meaning at 70% and the three surface metrics at 30%, so the verdict reflects both what the texts say and how they say it.

Swapping a handful of synonyms is the rewrite this tool flags fastest: it keeps semantic similarity high while lexical similarity barely moves, which is the exact fingerprint of a paraphrase rather than original thought. To write something genuinely new:

• Rebuild the sentence, not the vocabulary. Reorder clauses, merge or split sentences, and switch between active and passive voice so the structure no longer mirrors the original.
• Work from understanding, not the text. Read the passage, look away, then write the idea in your own framing. Copying the source sentence and editing it in place almost always leaves a match.
• Quote directly when the wording matters. If the phrasing is the point, use quotation marks and a citation instead of a near-copy.
• Always cite. Paraphrasing removes the words, not the obligation to attribute the idea.

Then re-run the comparison. A strong rewrite should fall into the SEMANTIC SIMILAR or ORIGINAL CONTENT band, and the sentence-level view should stop flagging pairs as rephrased. If lines still light up, they are the ones to rework.

What it does well: it quantifies how closely two texts you already have match in meaning and wording, highlights the specific sentence pairs that were rephrased, and places the result on a six-tier scale from IDENTICAL to ORIGINAL CONTENT.

Where it stops:

• It does not search the web or any database. You provide both texts and it compares only those two — it is not a source-finding plagiarism scanner.
• A high score is evidence, not proof of intent. Two writers can independently land on similar phrasing, especially for definitions, instructions or common facts.
• Very short inputs are noisy. A single shared sentence can swing the percentage, so aim for a paragraph or more on each side for a stable reading.
• It measures similarity, not authorship. It can flag a rewrite's fingerprint but cannot name which model produced it.

Read the verdict as a strong prompt to investigate, then confirm with the sentence-by-sentence breakdown and your own judgement before drawing a conclusion.

Astra's AI Text Similarity Detector is a free, ad-free, no-signup tool that compares two texts to detect AI paraphrasing, rewriting and semantic similarity directly in your browser. It runs on top of TensorFlow.js and the Universal Sentence Encoder — nothing is uploaded.

The 5-layer pipeline

• Layer 1 — Semantic similarity. Sentence embeddings on cleaned text capture meaning regardless of wording.
• Layer 2 — Concept preservation. Jaccard overlap on extracted key terms tells us whether the same concepts survived the rewrite.
• Layer 3 — AI fingerprints. Counts of formal/transitional phrasing, sentence-length uniformity, structural markers — patterns characteristic of LLM output.
• Layer 4 — Sentence mapping. Each suspect sentence is matched to its best original counterpart (length-filtered, 25% tolerance), then flagged as AI rephrased when semantic ≥0.5 and lexical ≤0.4.
• Layer 5 — Topic vs. derivation. If both texts discuss the same topic but show no real derivation signal, the score is reduced — coincidence isn't plagiarism.

Score formula

The final score is semantic-first: 0.7 · semantic + 0.3 · (cosine + jaccard + edit)/3, with a +0.1 bonus when the semantic-lexical gap exceeds 0.6 (and +0.05 when it exceeds 0.5). Identical inputs (similarity ≥0.999) shortcut to 100%.

Verdict bands

• ≥ 96% — IDENTICAL. Essentially the same text.
• ≥ 85% (gap > 0.6) — AI CONFIRMED. Strong rephrasing signature.
• ≥ 75% (gap > 0.55 or pattern signals) — AI LIKELY.
• ≥ 65% (gap > 0.51) — AI POSSIBLE.
• ≥ 39% — SEMANTIC SIMILAR. Topic overlap without strong derivation evidence.
• < 39% — ORIGINAL CONTENT.