Reading the result

The API rates every sentence on its own. This is how to read the verdict, the shares and the positions.

Verdict

Value	Meaning
`none`	No signs of AI found.
`some`	Some passages stand out.
`mostly`	At least half of the text is in the two upper levels.

There is deliberately no “human” class: a text without flags is not proven to be written by a person. The API returns no raw probability either.

Levels

`band`	Label	Meaning
`very_likely_ai`	Very likely AI	The sentence carries clear patterns of a language model.
`likely_ai`	Likely AI	Several sentences in a row read as machine-written.
`possibly_ai`	Possibly AI	A weaker hint. It does not count towards `ai_share`.
`not_flagged`	Not flagged	Only with `all_sentences`: the sentence carries no flag. It does not say who wrote the sentence.

A flag is a hint, not proof. Show it to your users as a reason to look at the passage again.

All sentences

With "all_sentences": true, sentences holds every rated sentence in text order. Whitespace between sentences and fragments shorter than three characters belong to no entry. Sentences without a flag carry the level not_flagged. That does not mean a person wrote the sentence: the API flags strictly, a single sentence only on a very clear signal and otherwise several conspicuous sentences in a row. Verdict, shares and billing stay the same.

Excerpt

"sentences": [
  { "start": 0, "end": 118, "band": "not_flagged" },
  { "start": 119, "end": 247, "band": "likely_ai" },
  { "start": 248, "end": 391, "band": "likely_ai" },
  { "start": 392, "end": 470, "band": "not_flagged" }
]

Use it when you show the whole text sentence by sentence or want to work with the same sentence split.

Positions

start and end count from 0; end is behind the last character. You choose the unit with offset_unit:

Value	Counts	Fits
`codepoint`	Unicode code points	Python, Go (runes), Rust (chars)
`utf16`	UTF-16 code units	JavaScript, Java, C#

The two differ only for characters outside the basic plane, such as emojis.

Highlighting passages

parts, last = [], 0
for s in result["sentences"]:
    parts.append(html.escape(text[last:s["start"]]))
    parts.append(f'<mark class="{s["band"]}">{html.escape(text[s["start"]:s["end"]])}</mark>')
    last = s["end"]
parts.append(html.escape(text[last:]))
marked_html = "".join(parts)

let out = "", last = 0;
for (const s of result.sentences) {            // request with offset_unit: "utf16"
  out += escapeHtml(text.slice(last, s.start));
  out += `<mark class="${s.band}">${escapeHtml(text.slice(s.start, s.end))}</mark>`;
  last = s.end;
}
out += escapeHtml(text.slice(last));