AI Text Similarity

What it measures

Paste two passages, one on each side, and the page scores how close their meaning is. The score is a percentage, with a label underneath: very similar meaning, somewhat similar, or different meaning. It is a semantic comparison, not a word-for-word diff.

That distinction is the whole point. Two sentences can share almost no vocabulary and still mean the same thing — "the film was far too long" and "the movie dragged on forever" — and the tool should score them highly. Conversely, two sentences can share most of their words and mean opposite things.

The model

The Universal Sentence Encoder turns a whole sentence into a fixed-length vector. Texts with similar meaning end up with vectors pointing in similar directions, and cosine similarity measures that angle. Because the model was trained on paraphrase pairs and natural language inference data, it learns relationships rather than word overlap.

The quantised variant is smaller and faster than the full model, with a small loss of precision that is rarely visible in a two-text comparison. The bundled model is the English version, so text in other languages is embedded poorly even when the characters are accepted.

The scale, and why 50 percent is not half

Cosine similarity runs from −1 to +1. The page maps it to 0–100 with (similarity + 1) / 2 × 100, which means a raw similarity of zero is displayed as 50 percent. The midpoint is neutral, not half-similar.

In practice, unrelated sentences tend to sit in the middle of the scale rather than near zero, because ordinary English sentences share a lot of structure. A pair that has nothing to do with each other may still show 55 or 60 percent. The labels are generous: "Very similar meaning" starts just above 70, which corresponds to a cosine of about 0.4.

Use the number to compare pairs, not to judge a single pair in isolation. If you are ranking candidate paraphrases against one reference, the order is meaningful even when every score sits in a narrow band. If you need a cut-off, collect a handful of known pairs and choose a threshold from those results rather than from the label text.

Step by step

1. Paste the first passage into Text A and the second into Text B.
2. The Compare button enables only when both sides contain at least three characters.
3. On the first comparison, the runtime and embedding model download; the panel reports each stage.
4. Press Compare. The result appears as a large percentage badge with a label.
5. Edit either side and press Compare again; the model stays loaded.

The Load AI Model button preloads the embedder ahead of time. When it succeeds it shows "Model Ready" and disables itself.

Where it is useful

• Paraphrase checks. Confirm that a rewrite preserves the original meaning.
• Duplicate detection. Spot two FAQ answers or two product descriptions that say the same thing.
• Translation review. Compare a translation against the source meaning, with the caveat that the model is English-only.
• Content audits. Find pages in a set that overlap heavily in meaning.
• Prompt testing. See whether two instructions really ask for the same behaviour.

It is not a plagiarism checker. A close paraphrase scores high precisely because the meaning matches, even when no words were copied. For copied text, use a string-matching or n-gram tool.

Where it fails

Negation is the biggest weakness. "I like this product" and "I do not like this product" share almost every word and sit in the same semantic neighbourhood, so they often score very high despite opposite meanings. The same applies to "recommended" versus "not recommended" and to sentences with a single crucial "no".

Numbers, names and dates are easily blurred. "The deadline is Monday" and "The deadline is Friday" differ in one token but in the fact that matters, and an embedding model tends to treat them as close. Product SKUs, dosages and legal references have the same problem.

Antonyms in identical frames — "the room was hot" versus "the room was cold" — can also score higher than a human would expect, because the surrounding structure dominates.

Very short inputs are unstable. A single word or a two-word phrase gives the model little context, and the score can swing wildly. Long inputs are truncated at the model's input window, so a comparison of two long articles may effectively compare only their openings.

Semantic similarity is not equivalence

A high score means the texts are used in similar ways, not that they assert the same facts. Two sentences about the same topic can score well while disagreeing. Two sentences with different topics can score lower even when they express the same logical claim with different entities.

That is why the tool is best used as a ranking signal. It narrows a set of candidates so a human can check the ones that matter. It does not decide whether two statements are logically equivalent, whether one entails the other, or whether a claim is true.

What it does not do

No word-level highlighting, no diff view, no entailment or contradiction labels, no fact checking, no cross-language comparison, no batch matrix and no export of vectors or scores. It does not explain which words drove the similarity, and it does not store a history of comparisons.

Privacy

Both texts are embedded locally in WebAssembly. Nothing is uploaded, and no comparison is retained after the page is closed. The runtime and model come from public CDNs on first use, which reveals that the tool was loaded but not what you pasted.

Limits

Two texts at a time, English only, one comparison per click, no threshold control, no download and no numeric export. Long passages are truncated by the model window. The page needs a modern browser with WebAssembly and network access on first load.

Common mistakes

• Reading 50 percent as half the meaning. It represents a neutral cosine, not a halfway match.
• Assuming a high score means the facts agree. Similar wording and similar topics can hide disagreement.
• Missing negation. A single "not" can reverse the meaning while leaving the score high.
• Comparing single words. Short inputs are unstable and context-free.
• Using it as a plagiarism tool. Paraphrases score high by design.
• Comparing long articles. Only the model's input window is considered.

Frequently Asked Questions