Why is confidence low on short texts?

There isn't enough evidence in a few sentences for a strong call either way. Rather than return a confident-looking number, the response marks confidence low, and you should treat it that way.

Will lightly edited human writing get flagged?

With some tools, often: even “extremely minor” AI polishing got 43–65% of human texts flagged by leading detectors (Saha & Feizi, APT-Eval 2025). Sentence-level scoring lets us show which sentences changed character and return mixed, instead of labelling the whole document.

Can a paraphrase fool it?

Word-level detectors are fragile: one paraphrase pass cut DetectGPT from 70.3% to 4.6% detection at 1% false positives (Krishna et al., 2023). Structural patterns hold up much better, still 96.1 macro-F1 after rewording in SlopShape v3. We'll publish our own benchmark on the research page.

How is Detect billed?

In words, from the same allowance as Rewrite: every 5 words scanned use one word, and calibrated requests count the same. Detect is included in every API plan; see pricing.

Do you keep the text I scan?

No. The text is deleted when the scan finishes, and request logs keep metadata only. We don't train models on your content. See What happens to your text.

Detect API

Detection that shows its work.

Q: Why calibrate to a writer?

Because some people naturally write in ways generic detectors mistake for AI. Perplexity-based detectors flagged 61% of essays by non-native English writers as AI (Liang et al., 2023). With calibrate_to, a text is compared with that person's own samples, so their normal style isn't the thing being flagged.

Sentence-by-sentence AI-likelihood with the pattern behind every flag, a confidence that follows the evidence, and calibration to the writer you actually care about.

See API plans API reference

Per-sentence scores
Named signals, not black boxes
1 word of allowance per 5 scanned

POST /v1/detect · granularity: sentence

In today's fast-paced business landscape, onboarding has never been more important.

In this post, we'll explore three strategies to help your team get new hires productive faster.

First, set up accounts before day one so nobody spends a morning waiting on IT.

Second, pair every new hire with a buddy for their first two weeks.

Third, write down the ten questions everyone asks, and answer them once.

By implementing these strategies, you can create an onboarding experience that drives engagement and retention.

In conclusion, great onboarding isn't just a process, it's an investment in your people.

Hover a sentence to inspect it. Example text and scores.

Document

likely_ai

verdict

medium

96 words

Sentence 1

0.96

Stock opener

A scene-setter that could open any post on any topic. Says nothing the reader can use.

Principles

Signals, not verdicts.

Detection is useful when it helps an editor look in the right place. It's harmful when it's treated as proof. The API is designed around that difference.

{

"verdict": "mixed",

"ai_likelihood": 0.58,

"confidence": "medium",

"signals": ["stock_opener",

"restated_close"]

}

Signals, not verdicts

You get a likelihood, a confidence and the evidence behind them. Nothing in the response claims to prove who wrote a text, and you shouldn't use it that way.

Every flag names the pattern that tripped it.

Calibrated to the writer

Some people simply write in ways generic detectors mistake for AI. Pass calibrate_to and we score against that person's own samples instead.

Perplexity-based detectors flagged 61% of essays by non-native English writers as AI. Liang et al., 2023.

Confidence follows length

A tweet doesn't carry enough evidence for a strong call, and the response says so. Confidence rises with word count, and short texts come back marked low.

Current thresholds: low below 150 words, medium to 600, high above.

How it works

Four passes, one readable answer.

1
Segment
The text is split into sentences and paragraphs, keeping headings and lists intact.
granularity
2
Score each sentence
Each sentence gets its own likelihood from phrase habits and rhythm, in context.
sentences[].ai_likelihood
3
Read the structure
Document-level patterns (announced thesis, restated close, summary sections) are checked.
structure[]
4
Calibrate & explain
Optionally measured against a known writer, then returned with the evidence for each flag.
calibrate_to

Interactive example

See how the score is built.

Switch samples, calibrate to a known writer, turn sentence scores off, or shorten the text and watch confidence drop.

An onboarding post drafted with an AI assistant, with three human-written steps in the middle.

In today's fast-paced business landscape, onboarding has never been more important.Stock opener

In this post, we'll explore three strategies to help your team get new hires productive faster.Announced thesis

First, set up accounts before day one so nobody spends a morning waiting on IT.

Second, pair every new hire with a buddy for their first two weeks.

Third, write down the ten questions everyone asks, and answer them once.

By implementing these strategies, you can create an onboarding experience that drives engagement and retention.Uniform rhythm

In conclusion, great onboarding isn't just a process, it's an investment in your people.Restated close

Sentences analysed7 of 7 · 96 words

Drag left to see confidence fall as the evidence gets shorter.

91%

AI-likelihood

likely_ai

verdict

medium

confidence

Signals

Stock openerAnnounced thesisUniform rhythmRestated close

Structure

Thesis announced up front
Summary section
Close restates the thesis

Example scores20 words of allowance

What we look for

The patterns are measurable.

Published research shows AI writing gives itself away in structure more than in vocabulary. These are the signals Detect names in its response, with the data behind them.

Structure

AI postsHuman posts

Thesis announced up front
93%
51%
Summary section
88%
27%
Close restates the thesis
77%
12%
Old-vs-new framing
76%
26%
No way for the reader to act
97%
38%

Share of B2B posts showing each pattern. Source: Sitefire / SlopShape v3, arXiv 2609.15369 (Sept 2026). Structure alone reaches 97.0 macro-F1, and 96.1 after the AI rewords its own posts.

Trailing “-ing” clauses

5.3×

GPT-4o's rate of sentence-final analysis clauses (“…, highlighting the need for…”) vs human writers.

Reinhart et al., PNAS 2025

Tell-tale vocabulary

28×

How often “delves” appeared in 2024 PubMed abstracts vs its expected rate.

Kobak et al., Science Advances 2025

Metronome rhythm

Sentence lengths that barely vary. Measured as spread, then compared with the writer's own baseline when you calibrate.

Measured per request

Tech specs

Endpoint: POST /v1/detect · model detect-2026-10
Base URL https://api.intactvoice.com/v1
Input: 40 to 25,000 words of plain text per request
Granularity: document one score · sentence adds per-sentence scores and signals
Response: verdict likely_human | mixed | likely_ai
ai_likelihood 0–1 · confidence low | medium | high
sentences[] and structure[] with named signals
Calibration: calibrate_to any voice_id; scores are relative to that writer's baseline
Counts the same as an uncalibrated request
Available on Pro API and Ultra API
Confidence: Low below 150 words, medium to 600, high above
Sync or async: Synchronous by default: the call waits up to 120 s
Long texts: pass async: true for a job and a webhook
Availability: Async jobs wait for worker capacity
Current service status
Rate limits: Basic API 60 req/min, Pro API 180 req/min and Ultra API 600 req/min
Enterprise: by agreement
Billing: 1 word of allowance per 5 words scanned
Calibrated requests count the same
Your text: Deleted when the scan finishes
Request logs keep metadata only, never the text

API

The contract.

One POST, a readable JSON answer. Here's the request, its parameters and the endpoints you'll use with it.

Request

curl -X POST https://api.intactvoice.com/v1/detect \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "text": "In today'\''s fast-paced world, side hustles are more than just extra cash. In this post, we'\''ll explore…",
  "granularity": "sentence",
  "explain": true
}'

Response

{
  "id": "dt_01JD4QB2M9",
  "object": "detection",
  "model": "detect-2026-10",
  "verdict": "likely_ai",
  "ai_likelihood": 0.94,
  "confidence": "medium",
  "sentences": [
    {
      "text": "In today's fast-paced world, side hustles are more than just extra cash.",
      "ai_likelihood": 0.97,
      "signals": [
        "stock_opener",
        "inflation_frame"
      ]
    }
  ],
  "structure": [
    {
      "feature": "announced_thesis",
      "present": true
    },
    {
      "feature": "restated_close",
      "present": false
    }
  ],
  "usage": {
    "words_billed": 19
  }
}

POST/detectapi.intactvoice.com/v1

Scores how likely a text is to be AI-generated, sentence by sentence, and returns the reasons: structural patterns, phrase habits and rhythm. Optionally calibrated against a known writer's own samples.

Parameter	Type	Description
textrequired	string	Text to analyse, 40–25,000 words. Short texts return confidence: low.
granularity	document \| sentence	sentence adds per-sentence scores.Default `sentence`.
explain	boolean	Return the signals behind the score.Default `true`.
calibrate_to	string	A voice_id. Scores are reported relative to that writer's own baseline.

Billed per input word at the Detect rate. Calibrated requests cost the same.

POST/voices

Create a voice

Builds a reusable voice profile from 2–20 writing samples by the same person. Samples are used for style only; facts in them are never copied into rewrites.

GET/jobs/{job_id}

Retrieve a job

Returns an asynchronous rewrite (or long detection) and, once it has succeeded, its result. Every rewrite is a job by default. Prefer webhooks in production; poll every few seconds otherwise. We don't keep results: the first response that includes a result (or a 2xx from your webhook) deletes it, and results not collected within 60 minutes are deleted. After that this endpoint answers 410 result_expired.

GET/usage

Get usage

Words and requests by day, endpoint and API key, plus what is left of this month's allowance and any pay-as-you-go use.

Full reference for POST /detect

FAQ

Questions, answered.

Something missing? Email support@intactvoice.com.

No, and please don't. A score is a signal about how text reads, not evidence of who wrote it. Use it to decide what to look at more closely: the response tells you which sentences and which patterns, so a person can make the call.

Know what to look at.

Score your own content sentence by sentence. Detect uses 1 word of allowance per 5 scanned, so 200,000 words on Basic API cover 1,000,000 words of detection.

See API plans Read the quickstart

Annual: 2 months free
Failed fidelity checks use no words
Cancel anytime

Detection that shows its work.

Signals, not verdicts.

Signals, not verdicts

Calibrated to the writer

Confidence follows length

Four passes, one readable answer.

Segment

Score each sentence

Read the structure

Calibrate & explain

See how the score is built.

The patterns are measurable.

Structure

Tech specs

The contract.

Questions, answered.

Know what to look at.