How AI Detectors Actually Work (and Their Limits)

The core idea

AI detectors estimate how "predictable" a passage is to a language model. Two signals dominate: perplexity (how surprised a model is by each next word) and burstiness (how much sentence length and complexity vary).

Why humans look different

Human writing tends to be bursty — a short punchy line, then a long winding one. Machine text is often smooth and uniform. Detectors lean on this.

The honest limitation

These are probabilities, not proof. Edited AI text, or naturally uniform human writing, can flip the result. That's why Sahihly frames scores as guidance for improving style — never as evidence for accusations.

Using scores well

• Treat a high score as "worth a second look," not a verdict.
• Read the highlighted sentences — they show why.
• Combine detection with your own judgment.