How AI detectors actually work
Most score two signals: perplexity (how surprising the next word is) and burstiness (how much sentence length varies). Human text is uneven; AI text trends toward the middle.
Detector comparison
| Detector | Method | Strength | Weakness |
|---|---|---|---|
| GPTZero | Perplexity + burstiness | Popular, education focus | False positives on careful human text |
| Writer | Perplexity model | Fast, API | Tunes to its own model |
| Copyleaks | ML ensemble | Multilingual | Opaque scoring |
| Glint AI (free) | Local heuristic | Private, no upload | Lightweight check |
Why scores disagree
Different training data, different thresholds, and different definitions of “AI.” A human-written, polished essay can score “AI”; a short AI draft can score “human.”
How to use a detector responsibly
Use it as a signal, not a verdict. Pair it with the Glint AI Content Detector for a private first pass, then edit for clarity.