What Is a Culling Confidence Score?

Why a yes/no answer is not enough

An AI cull that only outputs keep or reject hides the most important information it has: which decisions were easy and which were coin flips. A black frame is a certain reject; a formal where one aunt’s eyes are half-closed is genuinely arguable. The confidence score surfaces that difference, turning one undifferentiated pile of decisions into a triage: trust the certain ones, look at the uncertain ones.

How it changes the review pass

Without confidence scores, reviewing an AI cull means re-checking everything, which defeats the point. With them, review becomes targeted. Say a full wedding cull produces a few hundred low-confidence calls out of several thousand frames: your review pass is those few hundred, plus whatever spot-checking makes you comfortable. Your attention lands exactly where the software is weakest, which is the honest answer to being afraid the AI will cull the wrong photos.

The uncertainty bar

1. The software scores each frame and produces a decision plus a confidence value.
2. Decisions above the uncertainty bar go to KEEP or REJECT.
3. Decisions below the bar are routed to a REVIEW queue for a human call.
4. Your review decisions are recorded, and in learning systems they become training data.

Limitations

Confidence is the model’s self-assessment, not a guarantee. A model can be confidently wrong, especially on styles or situations it has not learned yet, and no threshold catches every mistake. Confidence also says nothing about story value: the software can be entirely sure a frame is technically weak while you keep it for reasons no algorithm can see. Treat the score as a routing signal for your attention, not as proof.

How Cull AI Studio handles this

Cull AI Studio attaches a confidence score to every single decision. Below the uncertainty bar, 0.65 by default, a photo is routed to REVIEW instead of silently trusted, so you judge only the close calls. Reviewing is not wasted effort either: corrections are logged with reasons and carry double weight when your personal model trains, so today’s uncertain calls make tomorrow’s cull more certain. Combined with preference-based selection, the review queue shrinks as the model learns which calls you actually consider close.