Detection Threshold

A detection threshold is the numerical boundary that separates outputs labeled 'likely AI' from outputs labeled 'likely human' in a detector's results. If a detector produces a continuous probability score between 0 and 100, setting the threshold at 50 means any input scoring above 50 is classified as AI. Setting it higher at 80 makes the detector stricter — fewer human texts get wrongly flagged (lower false positive rate) but more AI texts slip through (higher false negative rate). Lowering the threshold has the opposite effect.

Most end users never see or adjust the threshold directly; it is baked into the detector's default output. But the threshold is where detection policy decisions get made. A university using Turnitin may implicitly accept its default threshold; a content platform using Originality.ai at scale may calibrate a custom threshold against their own acceptable error rate. No threshold setting eliminates errors entirely — the choice is about which type of error you prefer to accept.

For users producing AI-assisted content, understanding detection thresholds informs the revision target. We are not going to give you a target to write under. A threshold belongs to whoever set it, and coaching people to slip beneath one is a different product from the one we are trying to build — it is also unbuildable honestly, since the platforms that set those thresholds forbid the testing that would tell anyone where the line actually is. Coda One's AI Detector reports a 0-100 score rather than a verdict, and deliberately not a percentage: it is an ensemble reading of the whole passage, not a probability and not a share of the text. The threshold its published error rates are measured at is 50, and both rates at that threshold are on /ai-detector/accuracy so the number can be argued with.