How AI Identifier Apps Actually Work (And Where They Stop)

You hold up your phone, tap once, and two seconds later something tells you the painting on your wall looks like a late-19th-century landscape, or that the ring in your hand is probably a round brilliant in white gold. It feels like magic, which is exactly the problem — magic doesn’t come with an instruction manual, so nobody tells you which answers to trust. Here’s what is really happening between the shutter and the result, and where the whole approach hits a wall it cannot get past.

What happens between the shutter and the answer

Three things, in order, and it helps to keep them separate because they fail in different ways.

• Find the object. The app first works out which pixels are the thing and which are your kitchen table. Clutter, shadow, and a busy background all make this step guess — and everything downstream inherits that guess.
• Turn the picture into numbers. This is the actual AI part. The model converts your image into a long list of numbers describing edges, proportions, textures, colour relationships, and repeating patterns. Similar objects end up with similar lists.
• Find the nearest known things. Your list gets compared against a vast set of reference objects. The closest matches become the suggestions, and how close they are becomes the confidence score.

Value estimates are a fourth, separate step layered on top: once the app has settled on a likely category and likely attributes, it looks up what things in that bracket have been selling for and returns a range. That number is modelled from the guess, not read off the object — which is why it comes back as a range and should stay one.

Great at “what kind,” shaky at “exactly which”

This is the single most useful thing to understand, and it explains almost every result that feels off.

Recognising a category is what this technology is genuinely excellent at. Dive watch with a rotating bezel and a date window. Quilted flap bag with chain hardware. Emerald-cut stone in a bezel setting. Impressionist landscape in oil. These have strong, repeated visual signatures, and a model that has seen a million examples reads them faster and more consistently than most people can.

Pinning down the exact instance is a different sport. Which of forty near-identical reference numbers. Which year of a production run. Which of two artists who trained in the same studio and painted the same harbour. The visual difference between those often comes down to a few millimetres of printing, a font weight, or a mark hidden under the strap — details that either weren’t in your photo or were three pixels wide when they were.

A recognition model answers “what does this look most like?” — never “what is this?” Most of the time those are the same answer. When money is involved, the gap between them is the whole story.

The wall: what a photo physically cannot contain

No amount of model improvement fixes this next part, because the information was never in the file. A photograph records light bouncing off a surface. It does not record mass, density, chemistry, conductivity, or smell. So every check that depends on one of those is off the table:

• Diamonds — telling a diamond from moissanite or cubic zirconia relies on thermal and electrical conductivity, refractive index, and density. A photo carries none of them. Separating a lab-grown diamond from a mined one needs lab equipment reading growth structure and trace elements, and the two look identical to any camera ever made.
• Paintings — pigment chemistry sets a hard “not before” date, and craquelure has to be read for depth through the paint layers, not just pattern. A good giclée print on textured canvas can beat a photo and lose instantly to a fingertip.
• Watches — the movement is the most conclusive evidence there is, and it lives behind a case back. Weight is the second, and grams don’t photograph.
• Handbags — hardware heft, the smell of real leather, the stitch density on an interior seam. Three of the strongest tells, all unavailable through glass.

Notice the pattern: every one of those is a measurement, not a look. That’s the boundary line. Anything you could establish by looking, a good model can help with. Anything requiring a scale, a loupe, a probe, or a lab is a different job with a different tool.

Your lighting is doing more work than the model

Genuinely: the gap between a poor result and a good one is usually the photograph, not the AI. Five minutes of care moves the answer more than any other thing you control.

• Clean it first. Skin oil on a stone flattens its brilliance and reads as a poor cut. Dust on a canvas reads as surface damage.
• Use indirect daylight. A window with the sun off it beats every lamp in your house. Direct sun and your phone’s torch both blow out highlights and destroy the detail that identifies things.
• Fill the frame, then stop. Get close, tap to focus on the object itself, and back off if focus won’t lock. A sharp small subject beats a blurry large one every time.
• Shoot it more than once. Straight on, at an angle, and in profile. Different angles surface different features, and depth almost never shows head-on.
• Photograph the small print separately. Serial numbers, hallmarks inside a band, a signature in a corner, a heat stamp inside a bag. A dedicated close-up of one of these is often the most informative frame in the whole set.
• Give it a scale. A coin or a ruler in frame turns “roughly this big” into an actual measurement, and size drives value harder than almost anything else.

Why we put the limits in the copy

An identifier that overstates itself is worse than useless, because you’d act on it. So our identifier apps return estimates labelled as estimates, and we keep a much deeper reference for that whole product line at The Lovenhall Registry — the house we publish the identifier line under. It holds the free calculators, the field guides, and a maker-by-maker authentication reference covering what to check on a specific maison, manufacture, grading lab, or artist, weighted by how conclusive each check actually is.

If you want the same argument made from the other side of the lens, the Registry’s own piece on how accurate diamond identifier apps really are goes through the four Cs one at a time and explains why each comes back as a band rather than a grade.

Reliable at

Category, style, shape, era, materials you can see

Indicative at

Exact model, exact attribution, value range

Can’t do

Anything needing weight, chemistry, or conductivity

Biggest lever

Your lighting and focus, not the model

Next step

A lab, appraiser, or specialist before money moves

The takeaway

An AI identifier is a fast, cheap, surprisingly good first read. It turns “I have no idea what this is” into “this is probably a mid-century oil landscape worth a few hundred” in the time it takes to hold your phone steady — and that single jump is what decides whether you frame it, insure it, or take it to a specialist. What it can’t do is finish the job. Treat the answer as your starting point, shoot a better photo than you think you need, and when the number gets big enough to matter, hand the object to someone who can put it on a scale.