Illustrative example

Our false-alarm target for Forensic mode, the strictest setting: out of 100,000 texts written by people, no more than one wrongly flagged.

Catch AI text
without accusing
human writers.

Borzoi marks the words that look AI‑written and names the model that most likely wrote them.

Contact us See it work

See exactly which words look AI‑written.

And what most likely wrote them.

Mixed

AI evidence in 2 of 4 paragraphs · most likely GPT family

Strictness: Forensic

When we read Frankenstein in class I assumed it would be a horror story, mostly because of the films. It isn’t, really. Most of the book is people writing letters and feeling guilty about things they can’t undo.

More than a verdict.

Every report shows the evidence behind it, and what most likely wrote the text.

You choose how sure we must be.

Three strictness levels: Standard, Strict and Forensic. The level used is shown next to every verdict. Forensic targets one false alarm in 100,000 human‑written texts.

Names the most likely source.

A model family, a specific model, and up to three alternatives, each with its confidence.

Humanizers change the words. Not where they came from.

Borzoi recognizes AI text run through humanizer and paraphrasing tools, and names the tool where it can.

Finds the AI inside human writing.

It marks where each AI‑written section starts and ends, and flags AI text that a person lightly edited.

Works on a single paragraph.

Short answers, comments and reviews get the same word‑level evidence as long documents.

Explains a human verdict, too.

When a text reads as human, the report says what kind of writing it most resembles.

1 in 100,000

False‑alarm target in Forensic mode, the strictest setting

Up to 20×

Cheaper per word than leading detectors1

1,900+

AI models and tools it can name as the most likely source

Price per million words

Published prices in US dollars1

Borzoi Detect$50
Originality.ai 5.0$119
Pangram 4.1$500
Originality.ai 5.0 Max$1,193

One request.
Every word scored.

The API returns the verdict, the strictness used, the most likely source, where each section starts and ends, and a score for every word.

REST APIJSON results, ready to store or display
Batch uploadFiles and CSV
Per‑word scoresThe same evidence you see in the report

Example responseStudent essay

{
  "verdict": "mixed",
  "strictness": "forensic",
  "likely_source": [
    { "family": "GPT family", "confidence": 0.72 },
    { "family": "Claude family", "confidence": 0.20 },
    { "family": "Gemini family", "confidence": 0.08 }
  ],
  "humanizer": { "detected": false },
  "sections": [
    { "start": 0, "end": 210, "label": "human" },
    { "start": 213, "end": 495, "label": "ai" },
    { "start": 498, "end": 672, "label": "human" },
    { "start": 675, "end": 849, "label": "mixed" }
  ],
  "tokens": [
    { "text": "profound", "evidence": 0.84, "source": "GPT family" },
    { "text": "consequences", "evidence": 0.78, "source": "GPT family" },
    …
  ]
}

{
  "verdict": "mixed",
  "strictness": "forensic",
  "likely_source": [
    { "family": "GPT family",
      "confidence": 0.72 },
    { "family": "Claude family",
      "confidence": 0.20 },
    …
  ],
  "humanizer": { "detected": false },
  "sections": [
    { "start": 0, "end": 210,
      "label": "human" },
    { "start": 213, "end": 495,
      "label": "ai" },
    …
  ],
  "tokens": [
    { "text": "profound",
      "evidence": 0.84,
      "source": "GPT family" },
    …
  ]
}

Contact us.

Borzoi Detect is in research. If you check writing for AI, tell us what you check and what you need.

A person reads every message.

One in 100,000

Catch AI textAI text without accusing human writers.