Check how
a text reads
A check reads a text with our AI detector and reports how likely it is to be AI-written, which paragraphs carry that, and the writing signals behind it. Checks run in the background, like runs. Each one is charged a small fixed price, held when it starts.
On this page
Starts a check and returns 202 Accepted. The API key needs the detect scope (see Authentication); without it the response is 403 insufficient_scope. Limited to 10 requests per minute per key.
curl https://humanizeword.com/api/v1/detections \
-H "Authorization: Bearer $HUMANIZEWORD_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"text": "Your text, in plain text or Markdown."
}'| Field | Description |
|---|---|
text | Required. Plain text or Markdown. Longer than restrictions.max_chars_per_run from GET /v1/me returns 400 text_too_long; text with no prose returns 400 no_prose. |
{
"id": "detection_01J9ABCDEF",
"status": "queued",
"words": 68,
"created_at": "2026-10-13T09:30:00+00:00",
"price_usd": "0.010000",
"charged_usd": null,
"result": null,
"error": null,
"progress": null,
"poll_url": "/api/v1/detections/detection_01J9ABCDEF"
}price_usd is held from your balance until the check finishes. If the available balance cannot cover it, the response is 402 insufficient_balance and nothing is queued. words counts prose only: Markdown structure and code are not charged.
Check statuses
| Status | Meaning |
|---|---|
queued | Waiting for the detector. |
running | Being read. progress shows the current stage. |
succeeded | result holds the report. charged_usd is the price, or 0.000000 when the text was too short to read. |
failed | error says why. Nothing is charged; the hold is released. |
Returns the check. Poll poll_url about once a second until status is succeeded or failed; most checks take a few seconds, a long text up to a minute. A check is kept for one hour after it was created and then returns 404, as does a check made by another account. The text you sent is never stored.
{
"id": "detection_01J9ABCDEF",
"status": "succeeded",
"words": 68,
"price_usd": "0.010000",
"charged_usd": "0.010000",
"result": {
"version": 2,
"score": 0.98,
"verdict": "ai",
"words": 68,
"sentences": 4,
"mean_sentence_words": 17.0,
"paragraphs": [
{"start": 0, "end": 286, "words": 35, "score": 0.98, "verdict": "ai"},
{"start": 288, "end": 556, "words": 33, "score": 0.98, "verdict": "ai"}
],
"signals": [
{
"key": "function_words",
"label": "Function-word density",
"value": 0.25,
"unit": "share",
"human": 0.5369,
"ai": 0.4625,
"lean": "ai",
"human_percentile": 1.0,
"reliability": 0.692,
"evidence": [{"start": 0, "end": 162}, {"start": 407, "end": 556}]
}
],
"models": 3,
"elapsed_ms": 4220,
"basis": "whole_text"
},
"error": null,
"progress": null,
"poll_url": "/api/v1/detections/detection_01J9ABCDEF"
}The report
Every start and end is a character offset into the exact text you sent, so you can highlight paragraphs and evidence in your own copy. Keep that copy: the report holds no text.
| Field | Description |
|---|---|
score | How likely the text is to be AI-written, from 0 to 1. Normally the reading of the whole text. When every scored paragraph reads as human, it is the paragraphs' word-weighted score instead (see basis). |
verdict | human below 0.35, ai at 0.65 or above, mixed between. insufficient when the text has fewer than 40 words, which is not charged. When every scored paragraph reads as human, the verdict is human, unless the whole text reads 0.9 or above. |
paragraphs | Each paragraph's span, word count, score and verdict. A paragraph under 20 words has score and verdict null. |
signals | Writing signals that separate human from AI text on our measured corpus: this text's value, the typical human and AI values, which way it leans, how reliable the signal is (0–1), and evidence spans. |
models | How many classifiers read the text. |
basis | whole_text, or paragraphs when the paragraphs decided the verdict and score. Added in report version 2. |
whole_text_score | Only when basis is paragraphs: the whole text's own reading, 0 to 1. |
Every detector, ours included, can be wrong about a given text, and other detectors will score the same text differently. Use the score as one signal, not as a verdict about who wrote something.
Returns what a check of this text would hold and charge now, without starting one. No API key is needed.
{
"text": "Remote work has fundamentally transformed the modern workplace."
}{
"prose_word_count": 8,
"price_usd": "0.010000"
}Pricing
A check is paid from the same balance as runs. Its price depends only on length, so it is known before you start: currently $0.01 per check up to about 5,000 words, then $0.002 per 1,000 words. POST /v1/detections/estimate always returns the current price. You are charged only when a reading lands: a check that fails, or a text too short to read, is not charged.
Limits
| Limit | Details |
|---|---|
| Rate | 10 checks per minute per key. A 429 includes a Retry-After header. |
| Length | Up to restrictions.max_chars_per_run characters (35,000 by default). At least 40 words for a verdict. |
| Retention | A check and its report are kept for one hour. The text is never stored. |