Detecto
Trust

Calibration is a posture, not a footer note.

A detection score is a probability, not a verdict. Detecto reports the score, the confidence band, and the signals that drove it. We refuse to claim 100% accuracy because no detector is 100% accurate.

Calibration

What the number means.

Read this once and the score stops feeling like a black box.

Score
AI likelihood

A 0-100 probability that the input was AI-generated. Reported alongside the confidence band.

Confidence
Variance band

The variance across resampled detection passes, read as ± on the score. Wider band, less commit.

Verdict
Plain-language label

Likely human, likely AI, mixed signals, or inconclusive. Not the headline. The headline is the score and the band.

Signals
Vendor-agnostic features

Perplexity, burstiness, lexical diversity, and modality-specific features. The labels never depend on the vendor.

Calibration
Release cadence

Calibration plate refreshed against new generators on a regular cadence. Last refresh is in the dashboard footer.

Disclosure
Per-input limitations

When input falls outside the calibrated range (very short, low quality, atypical register), the report says so.

What we measure

Confidence variance, not just a point estimate.

The variance band is shipped alongside the score because the variance is part of the answer.

For text, Detecto runs the input through a calibrated language-pattern model. The model produces a probability, and the same input is resampled multiple times across slightly perturbed prompts. The variance across those passes is the confidence variance reported in every report.

For image, audio, and video, Detecto runs the file through a multi-signal media analyzer. The aggregate generation likelihood is reported alongside media-specific evidence: image signals, aggregate audio signals with waveform playback, and per-scene video likelihoods. Each exposed segment value carries its own variance band.

The point estimate is a starting point. The confidence variance is the room you have to act on the score.

Where Detecto explicitly does not commit

If you find these limitations buried in our marketing copy, we have failed.

  • Detection is probabilistic. A 92% AI score is not proof. A 6% AI score is not exoneration.
  • Very short input (under ~120 words for text; under ~5 seconds for audio) widens the band.
  • Translated text, heavily edited human writing, and heavily compressed media all degrade signal.
  • Newer generators reach our calibration plate before we update; expect lag during release cycles.
  • Detection signals can interact with non-native register and creative writing.
  • Detecto does not commit to any binary verdict. It reports a likelihood and the evidence behind it.
FAQ

Common questions on accuracy.

Read the related posture pages

False positives. Privacy. Security. All written down.