How Accurate Is AI Nail Biting Detection?

What is actually being detected

It is worth being precise, because the honest answer to "how accurate is it" depends entirely on what the question means.

Stop Biting does not detect biting. It detects a fingertip being close to your mouth. Two models run on each sampled frame — a hand landmark model that locates twenty-one points on the hand, and a face landmark model that locates the mouth. The system measures the distance between each of the five fingertip landmarks and the midpoint of the inner lips, and treats a fingertip inside a threshold distance as a positive.

So the accurate description is proximity detection used as a proxy for biting. That proxy is good, because the hand-to-mouth position is a necessary precondition for biting and is not a posture most people hold accidentally for long. It is not the same thing as observing teeth on a nail, and any product claiming otherwise from a standard webcam is overstating things.

Why a dwell requirement exists

A single frame showing a fingertip near the mouth means very little. People touch their faces constantly — scratching a nose, resting a chin, adjusting glasses, covering a yawn.

So a positive frame alone does not trigger anything. Detection runs at five frames per second, and three consecutive positive frames are required before an alert fires. That is roughly 600 milliseconds of sustained proximity.

That single design choice removes the large majority of incidental face touching, because brief contact does not survive the dwell requirement while actual biting comfortably does. It is also why the alarm arrives a fraction of a second after your hand arrives rather than instantly — that delay is doing useful work.

The sensitivity trade-off, concretely

Detection thresholds are expressed as a fraction of the frame, and the three sensitivity settings correspond to different distances.

High sensitivity uses the tightest threshold: the fingertip has to be very close to the lips before it counts. This produces the fewest false alarms and will miss episodes where you bite with your hand held slightly away from your mouth. Low sensitivity uses the widest threshold, catching more real episodes at the cost of flagging more chin-resting and face-touching. Medium sits between them and is the sensible default.

There is no setting that catches everything and flags nothing, and any product that claims otherwise is describing marketing rather than engineering. The useful framing is which error you would rather have. In the first week, most people are better served by tolerating some false positives, because the goal at that stage is building awareness — and being prompted to notice where your hand is turns out to be useful even when you were not about to bite.

What reliably degrades accuracy

Failures are mostly not model failures. They are input failures, and they are largely fixable.

• Poor lighting. Both models need to resolve landmarks. A backlit face against a bright window is the single most common cause of missed detections.
• The hand leaving frame. A standard laptop webcam has a limited field of view, and a hand approaching from below the frame edge may not be fully visible.
• Extreme head angles. Looking sharply down at a phone or desk can take the mouth landmarks out of reliable range.
• Occlusion. A hand covering most of the mouth, resting a face fully in a palm, or holding an object can interfere with landmark placement.
• One hand at a time. The configuration tracks a single hand, so simultaneous two-handed activity is not fully covered.

Why imperfect detection still works

This is the part that matters, and it is easy to lose sight of while comparing accuracy numbers.

The therapeutic mechanism is not comprehensive surveillance. It is interrupting enough episodes, often enough, that the cue stops reliably predicting the reward. Habit reversal training works through repeated interruption over weeks, and a system catching a substantial majority of episodes supplies that.

Compare it against the actual alternative rather than against perfection. Unassisted, most nail biters notice well under half their own episodes, and the ones they notice are systematically the memorable ones rather than the typical ones. A detector that catches most episodes, including the ones during deep focus that you would never have registered, is a large improvement over that baseline even with its failure modes fully acknowledged.