AI card grading explained: how accurate is it really?

Every card-scanning app now claims "99% AI grading accuracy." That number sounds great in marketing, but it doesn't survive contact with a real PSA submission. Here's what AI grading actually does, where it's reliable, and where it gets things wrong.

What AI grading can actually see

A good AI grader for Pokémon and Yu-Gi-Oh cards analyzes four signals from a clean photo or scan: centering offsets (front and back), edge whitening, corner wear, and surface defects like scratches, dimples, holo scuffs, and print lines.

Centering

Pixel-level measurement of the inner border on all four sides. The most reliable subscore an AI grader can give — geometry is geometry.

Surface

Detects scratches, holo scuffs, print dots, and dimples that catch light. Reliable on raw photos, much less reliable through a sleeve or top-loader.

Edges & corners

Whitening and soft corners show clearly. Tiny dings sometimes need a magnified shot to register.

Back of card

Often the deciding factor for a PSA 10. Many apps skip the back entirely — if yours does, treat its grade as an upper bound.

Where AI grading disagrees with PSA

Print defects vs. damage. A vertical print line baked in at the factory is not damage, but most AI graders penalize it like a scratch. PSA reads context.
Light reflections. A holo reflecting your room light reads as "scratch" to an AI that wasn't trained on a controlled lighting rig.
Off-condition edge cases. Slight warping, miscut borders, and factory-rough cuts confuse most AI models — humans handle them in seconds.

How to use an AI grade well

As a pre-submission filter. If the AI estimate is a confident 9 or 10 with clean subgrades, it's worth submitting. If it's a 7 with a centering flag, a PSA 10 is extremely unlikely — don't waste the fee.
As a triage tool for big collections. Scan 200 cards in an evening and let the AI flag the 10 worth a second look.
Not as proof of value. No grading company recognizes an AI grade. When you sell or trade, the buyer wants a real slab.

PocketVault's approach

PocketVault gives every scan a 1–10 estimate, 9 subscores, a defect heatmap, and a plain-English explanation of what the model saw. We log every grade and compare it against the PSA outcome when you report back — so the model improves on real data, not on marketing claims. We won't promise 99%, because nobody can verify that number honestly. What we will promise is that you'll see exactly why the model scored the way it did.

Try it: scan a card and see the full subgrades + defect heatmap. Open the scanner →