AI Detector API

The planned shape

One endpoint, JSON in and JSON out. You send text between 120 and 6,000 characters per call. You receive an integer ai_score from 0 to 100, a verdict string, a verdict_class of human, mixed or ai, and a word count. The same reading the web dial shows, machine readable, with the inconclusive band preserved: scores from 40 to 69 will arrive labeled mixed, and we encourage clients to surface that label rather than rounding it away.

POST /v1/detect
{ "text": "..." }

200 OK
{ "ai_score": 84, "verdict": "Very likely AI generated",
  "verdict_class": "ai", "words": 412 }

Principles that will survive contact with pricing

A generous free tier for spot checks and evaluation. Volume pricing for teams, with no accuracy upsells, because every tier gets the same engine. No retention of submitted text beyond processing. And documentation that states failure modes plainly: the API will ship with the same false-positive guidance the website carries, because a score consumed by software is even easier to misuse than one read by a person.

What to do today

If your volume is human-scale, the web scan already covers it. If you want a note when the API opens, say so through the contact page with a line about your use case; real use cases shape the limits we choose. For background on what the scores mean and where they fail, start with how AI detectors work.