How accurate is AI antique identification? Questions about confidence and limits
AI antique identification is a reading of a photograph, not an examination of an object, so its reliability changes with the item, the light and the mark. Patina is built to say how sure it is on every result, and to point you toward a vetted appraiser when it is not sure enough.
Patina is not open yet, so anything on this page about what it does is what it is being built to do.
Why accuracy is the first question collectors ask
You are standing in someone's dining room with a lidded tureen in your hands, the mark underneath is half worn away, and you have a few minutes to decide. An answer that arrives in seconds is only useful if you know how much weight it can carry. That is the real question behind "how accurate is AI antique identification", and it does not have a single number for an answer.
An AI vision model reads pixels. It can recognise a form, a glaze, a joinery style or the shape of a hallmark, and it can associate those with makers, eras and places of origin it has seen described many times. It cannot feel the weight of the piece, turn it to the light, tap the porcelain, or check whether the screws are hand cut. Anything that depends on those things sits outside what a photograph can settle.
So Patina is built around a different promise than a verdict. Every result comes back in plain English with maker, era and origin, alongside a confidence level and a persistent note that this is an AI estimate rather than an expert appraisal. When confidence falls below our threshold, the result carries an appraisal triage flag and a referral link to a vetted appraiser, because the honest next step for a doubtful object is a person holding it.
The questions below are the ones we would want answered before trusting anything like this with a find.
Questions
How accurate is AI antique identification, really?
There is no honest single percentage, and we will not invent one. Accuracy depends almost entirely on how much evidence the photograph contains: a crisp, legible factory mark on a well known form gives a model a great deal to work with, while an unsigned country chair in poor light gives it very little. That is why Patina returns a confidence level with every identification rather than a flat answer, so you can tell the difference between a reading that is well supported and one that is a best guess. Treat it as a knowledgeable second opinion that narrows your search, not as a conclusion.
What does the confidence level actually mean?
It is the model's own account of how well the image supports the maker, era and origin it has proposed. High confidence usually means the photo contains something distinctive, a clear mark, a recognisable pattern, a form with few near neighbours. Lower confidence means the visible evidence is consistent with several possible answers, and the identification is one of them rather than the only one. We show it on the result screen in plain language because a confident collector and a hesitant model should not look the same on your phone.
What happens when Patina is not confident enough?
Below our threshold, the result carries an appraisal triage flag along with a referral link to a vetted appraiser. That is the whole point of the design: the useful thing to know about an ambiguous object is that it is ambiguous, and that a human should look at it. We would rather hand you a clear "this one needs an expert eye" than a smooth paragraph that sounds settled and is not. Referral links may be affiliate links, and we log the clicks so we know which referrals are worth keeping.
Which objects is AI identification weakest on?
Anything where the answer lives in the hand rather than the image. Reproductions that copy a genuine form and mark, unmarked or country made furniture, heavily worn or partially struck hallmarks, later marriages of old parts, and restored surfaces that hide construction evidence are all cases where a photograph is a thin basis for a conclusion. Condition and repair judgements are similar, since they usually need raking light and close inspection. Expect lower confidence on those, and expect the triage flag to fire more often.
Will it tell me what my item is worth?
No, and not as a limitation we plan to remove. Patina never returns a dollar value, because value depends on condition, provenance, completeness and a live market, none of which a single photo settles. Identification is our job and valuation is an appraiser's, and mixing the two is how people talk themselves into bad decisions. When you need a figure, the referral link is there for exactly that.
Can I rely on a result to buy or sell something?
Please do not. Every result screen carries a persistent disclaimer that the identification is an AI generated estimate and not an expert appraisal, and it is not a basis for a purchase or sale decision. What it is good for is orientation: a name to search, an era to sanity check, a reason to look harder at the base or the back before you commit. If money of any consequence is involved, the confidence level and the triage flag are telling you whether to get a human opinion first.
Has the identification been tested on real finds yet?
The identification path is built and the model call is wired up, but we have not yet published a verified end to end run of a real photograph resolving to maker, era and origin on the deployed app. We are saying that out loud rather than implying a track record we have not earned, which is also why the site is a waitlist and not a purchase page. When we have that evidence, it will appear here in place of this paragraph. Photos are processed by Anthropic's model on our behalf, and stored with your scan results while your account exists.
How do I give the model the best chance of a good reading?
Photograph the most informative surface rather than the prettiest one: the mark, the hallmark, the signature, the underside, the joint. Fill the frame with it, get indirect daylight on it if you can, and wipe off dust that reads as a scratch. A legible mark tends to move a result from a plausible guess toward a supported one, and it is the single biggest thing you control. If the mark is genuinely unreadable, expect lower confidence, and take that as information rather than a failure.