Guides

Calibrating detection to a writer

Score text relative to a known writer's own baseline instead of a population average.

Generic detectors compare text against an average writer. Some real people sit far from that average: very formal, very consistent, writing in a second language. Calibration compares text against that person's own writing instead.

When to calibrate

You check work from a known, recurring set of writers (staff, contributors, agency clients).
A writer's genuine work scores mixed or likely_ai uncalibrated.
You want to spot a change in how someone writes, not a population-level guess.

1. Build a baseline

Calibration uses a voice profile. Create one from writing you're confident is the person's own, ideally from before they had AI tools in their workflow. The response includes their baseline:

{
  "id": "voice_3m9qa2",
  "object": "voice",
  "name": "Priya — research notes",
  "samples": 8,
  "words": 9840,
  "baseline": {
    "ai_likelihood_mean": 0.41
  }
}

A mean of 0.41 tells you a generic detector already finds this writer somewhat “AI-like”. That's the false positive calibration removes.

2. Detect with calibrate_to

curl -X POST https://api.intactvoice.com/v1/detect \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "text": "…", "calibrate_to": "voice_3m9qa2", "explain": true }'

3. Interpret relative scores

Calibrated scores answer “how unlike this writer's own work is this?”. The same text, scored both ways:

	ai_likelihood	verdict	Reading
Uncalibrated	0.62	`mixed`	Formal register looks AI-like to a population model.
Calibrated to voice_3m9qa2	0.18	`likely_human`	Typical for this writer.

Look at the reasons too. If a calibrated score is high, the sentences and structure arrays show what changed, for example an announced_thesis and restated_close the writer never uses.

Keeping baselines fresh

Rebuild the voice when a writer's role or format changes (from long reports to short posts, say).
Never add text you're trying to evaluate to the baseline.
Calibration costs the same as a normal detection.

Calibration reduces false positives; it doesn't turn a score into proof. The rules in Signals, not verdicts still apply.