How AI detection works in Syft
What Syft Detect measures when it reads writing, what a result contains, and how to read one well.
Published Syft
What the result is
An estimate of resemblance, shown with its evidence.
Syft reads a text and reports how strongly it resembles machine-generated writing on a set of measurable features. The reading is presented as an estimate with the passages behind it, never as a fact about who wrote the text.
Every analysis produces a signal band, the passages that carry the signal, and any notice about the input, so you can see what drove the result, not just a number.
The bands are: Low signal (the writing sits mostly in the range of human drafting), Moderate / elevated signal (some AI-writing signals are present and worth reviewing), High signal (strong AI-writing signals across the text), and Insufficient signal (too little or unsupported text for a reading). High is reported only for 160 words or more. A marked passage carries stronger AI-writing signals; a detector result is not proof of authorship.
AI-writing signal band · High signal
3 of 9 sentences in a sample document carry the signal. Read live by Syft’s engine on this page.
The reading
One reading, and every result carries its evidence.
Syft Detect reads whatever you give it the same way, on this site and in the workspace. There is nothing to choose and nothing hidden: the result names the detector version that produced it.
What it reads
Syft Detect reads the text passage by passage, identifies how strongly each passage carries patterns associated with machine-generated writing, and reports a signal band rather than a percentage.
It needs about 80 words before it will estimate anything. Below that there is not enough to read, and Syft says so rather than inventing a reading.
What it reports
A signal band for the text as a whole — Low signal, Moderate / elevated signal, High signal, or Insufficient signal — and the passages that carry the signal, marked on the text itself with the features behind each one. Never a percentage.
Notices about the input travel with the result: text that is too short, not running prose, or longer than the reading is calibrated for is described as such rather than scored anyway.
What it measures
Features of the writing, not fingerprints of a model.
Syft Detect does not look for a particular AI model. It identifies how strongly the writing carries patterns that machine-generated text tends to share, which careful human writing can sometimes share too.
Sentence-length variation
Human writing tends to vary sentence length; generated text often settles into a narrower rhythm. Measured across the whole text and passage by passage.
Model-favoured phrasing
Connective and framing phrases that appear far more often in generated prose than in edited human prose. Their presence raises a passage's reading; it is never decisive on its own.
Contraction rate
How often the text uses contractions. Formal registers use few, which is one reason formal human writing can read as AI-like on this feature.
Punctuation variety
The spread of punctuation the text uses. Uniform punctuation is one signal among several.
Lexical diversity
How varied the vocabulary is relative to the text's length.
Passage evidence
Each sentence is scored, and the ones at or above the AI-like threshold are shown with the features that drove them, so the reading can be checked against the text.
What is deliberately not described here: the exact weights and phrase lists. They are in the product, and publishing them as a recipe would help someone tune text to a detector rather than help anyone understand a result.
What a result is evidence of
Evidence about the writing, shown with the passages behind it.
A result describes patterns in the text in front of it, and the marked passages show where they concentrate. A detector result is not proof of authorship, in either direction; it belongs beside the record of how a text came to be.
- Built for essay-like prose
- The reading is designed for essay-like and academic prose in English of about 160 words or more. Between 80 and 159 words it is limited and is never reported as high.
- No reading for some inputs
- Short notes, lists, code, bibliographies, quotation blocks and repeated text get no reading. Syft says so instead of scoring them.
- A signal, not proof
- A detector result is not proof of authorship. It is a statistical estimate of how strongly the writing carries patterns associated with machine-generated text.
- It can be wrong in both directions
- Human writing is sometimes reported as elevated or high, more often for reviews, forum posts and summaries than for essays. Edited or paraphrased AI text is often missed. Text from some models is missed more often at the high band, including Llama-family models, older GPT-3-era models and, at the margin, GPT-4.
- Populations we watch
- In Syft's evaluation, ESL essays (English as a second language) were reported as high slightly more often than native-speaker student essays, though still rarely; short scientific abstracts were not tested in enough numbers to state a rate. Both are monitored during the beta, and a high result on such writing deserves extra care.
- Marked passages show where the signal concentrates
- A marked passage carries stronger AI-writing signals. It is not a claim that the passage, or the person, used AI, and a human paragraph inside a mixed text can be marked.
- Results change with the model version
- Every result records the detector version that produced it. A later version may read the same text differently.
- Never the sole basis for a decision
- Do not use this result as the sole basis for academic, disciplinary, employment, or legal decisions.
- If a result is contested
- A reading is not proof, in either direction. What can be reviewed is the record of how a text came to be: drafts, revision and version history, notes and source material, and the writer’s own account. Wrote this yourself? How to show your process.
Measured performance
What Syft has and has not published about its own accuracy.
UNKNOWNSyft has not yet published a measured accuracy or false-positive rate for the detector on this site. No figure is stated here because no approved measurement of the production engine exists. When one does, it will appear in this place with its engine version, method, sample, date and limitations.
OBSERVEDIndependent research on AI-text detectors as a class reports meaningful false-positive rates, lower accuracy on paraphrased text, and bias against non-native English writers. The references below are the studies those statements come from.
INFERREDThose findings are about detectors generally, including ones built very differently from Syft’s. They are cited as context for how to read any result, not as a measurement of Syft.
References
- 01Testing of detection tools for AI-generated text Weber-Wulff et al., 2023 · International Journal for Educational Integrity · 2023
- 02GPT detectors are biased against non-native English writers Liang et al., 2023 · Patterns · 2023
- 03RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors Dugan et al., 2024 · ACL · 2024
Reading a result well
Four habits that make a result useful.
- 01
Read the passages, not just the band
Open the marked sentences. If they are formulaic transitions and stiff summaries, that is what Syft reacted to. If they are the writer's own specific claims, weigh the reading down.
- 02
Compare drafts, not verdicts
A reading changes as text changes. Checking an earlier draft and a later one tells you more than either alone.
- 03
Treat 'uncertain' as the answer it is
A moderate / elevated signal means the evidence is mixed and wants review. Forcing it to a yes or no adds certainty the reading does not carry.
- 04
Ask for process evidence before conclusions
Drafts, version history and notes show how a text came to be. A detector reading cannot.