A.I. as a way to authenticate watches

Posts
35
Likes
9
For those of you who are tempted to ask AI about the authenticity of your watch, don’t bother.


Here are my experiences.

1. I took three different pictures of my Seamaster and Chat GPT identified it as authentic, fake and inconclusive in three attempts.

2.Here is what it says about my brand new Speedmaster purchased directly from Omega 3 weeks ago

I asked:
Is this authentic or fake.



ChatGPT said:
This appears to be an Omega Speedmaster Professional "Moonwatch", but with some unusual features that raise questions about authenticity:
🔍 Observations:
1 Dial Color & Text
◦ The dial is white with red “Speedmaster” text.
◦ Omega has made limited white-dial Speedmasters (e.g., the Alaska Project, Mitsukoshi, or Panda dials), but most have very specific subdial layouts and text fonts.
◦ The font and spacing of “PROFESSIONAL” and “SPEEDMASTER” here look slightly off compared to authentic models - the “Speedmaster” script seems too thick and a bit misaligned.
2 Subdial Spacing
◦ Authentic Speedmaster Professional models using the Caliber 1861/3861 have distinct subdial spacing. In your photo, the subdials look too close to the center, which is common on replicas using non-Omega movements (like the Seagull ST19 or Valjoux 7750).
3 Bezel and Case
◦ The tachymeter bezel font looks somewhat shallow and uneven in engraving depth. Genuine Speedmasters have crisp, precisely printed bezels.
◦ Case shape appears close, but the lug bevels and brushing seem less refined than on an authentic Omega.
4 Hands
◦ The hands look slightly thicker than genuine Omega ones. The central chronograph seconds hand tip appears shorter than usual.


⚠️ Verdict:
Based on the dial printing, subdial spacing, and case finishing, this watch is very likely a replica or a modified/custom build (e.g., a “Frankenwatch” combining genuine and aftermarket parts).


If you want to be certain:


• Remove the caseback and check for an Omega Caliber 1861 or 3861 movement with proper engravings.


• Verify the case reference number between the lugs or inside the caseback.


• Compare it against Omega’s official parts database or bring it to an authorized Omega service center.


Would you like me to point out the exact reference it seems to be imitating (e.g., Mitsukoshi or Alaska
 
Posts
726
Likes
918
I wonder if AI can access the content from OF easily. One would think that ingesting pictures and commentaries, as well as these great threads deep diving into references would be a pretty strong starting point!
 
Posts
23,024
Likes
51,474
Also, discussed here. I agree that LLMs are not well-suited to this task, even using in-focus photos.
 
Posts
35
Likes
9
Also, discussed here. I agree that LLMs are not well-suited to this task, even using in-focus photos.
Thanks for the link. I hadn’t found it on a search before posting mine.
 
Posts
35
Likes
9
I can understand that uploaded images may not be high res or great quality and result in poor analysis, HOWEVER:
Here was my biggest gripe about the AI answer.

“The dial is white with red “Speedmaster” text.
◦ Omega has made limited white-dial Speedmasters (e.g., the Alaska Project, Mitsukoshi, or Panda dials), but most have very specific subdial layouts and text fonts.”

Clearly the AI was not up to date on the White dial Speedy released in 2024. Not good enough searching by the AI. Not even close to good enough.
 
Posts
1,050
Likes
1,125
I can understand that uploaded images may not be high res or great quality and result in poor analysis, HOWEVER:
Here was my biggest gripe about the AI answer.

“The dial is white with red “Speedmaster” text.
◦ Omega has made limited white-dial Speedmasters (e.g., the Alaska Project, Mitsukoshi, or Panda dials), but most have very specific subdial layouts and text fonts.”

Clearly the AI was not up to date on the White dial Speedy released in 2024. Not good enough searching by the AI. Not even close to good enough.

One of the shortcomings with LLMs, they are only as good as their training data. LLMs break down REALLY quickly if they are trained on AI generated data. As a result, many of the AI companies are VERY resistant to training on more recent data. Its also why web forums like this one (and Reddit/etc) are so important to them, human generated content is VERY VERY valuable in the AI era.

Second, and most importantly: They don't actually KNOW anything, they apply no logic/etc. They are a really really large database of words/connectors and a random number generator. They aren't actually analyzing the photo, they are using it to set an internal state for their word/connector database random number generator.

When analyzing the photos, they do a 'nearness' search (or feature search, like google photo search), and use that as a 'begin state' for their word generation.

Anyway, the side effect of this is they tend to generate more or less the 'average' of their training data. So an LLM trained on Omega Forums would, obviously, note that just about every single watch when asked about authenticity, is a repainted dial (because thats what happens here!).

As far as the White Speedmaster: It doesn't really matter IF it is trained on newer materials; the signal-ratio on "white speedy==fake : white speedy==new thing" is going to be decades to the former, and only a couple of years to the latter, so the random number generator is going to lean very heavily to the former.
 
Posts
23,024
Likes
51,474
I can understand that uploaded images may not be high res or great quality and result in poor analysis, HOWEVER:
Here was my biggest gripe about the AI answer.

“The dial is white with red “Speedmaster” text.
◦ Omega has made limited white-dial Speedmasters (e.g., the Alaska Project, Mitsukoshi, or Panda dials), but most have very specific subdial layouts and text fonts.”

Clearly the AI was not up to date on the White dial Speedy released in 2024. Not good enough searching by the AI. Not even close to good enough.
There may well be pre-2024 reddit threads in the training data. The model doesn't know what is correct, at best it's just spewing back the training data, and even then there can be unexplainable hallucinations.

My strong conviction is that LLMs are the wrong type of ML/AI for authentication. Convolutional neural nets, specifically trained for classification, are extremely well-established and the obvious approach for this application. I am peripherally involved in a research project that uses a CNN for classification of undesirable by-products created during pharmaceutical processing, and it is very effective. I understand that it's tempting to use LLMs because they are readily available and "free." But really they are not expected to be good at this.

Edit: As I mentioned in the other thread, Entrupy has a product that uses CNNs to authenticate handbags and sneakers, and claims to reject over 99% of fakes. I suspect that they do have a higher (i.e. greater than 1%) rate of false fake identifications. And other companies, e.g. IBM, are developing similar solutions.
Edited:
 
Posts
2,643
Likes
4,218
I posted my thoughts in another thread.

Not sure what CNN's are. but I have a 30 year old book on Neural Nets in the complex domain. Googled the author a few years back when LLMs became popular. Still seems to be peddling the same get rich in the stock market with NN predictions.

I still want to write a system using the silhouettes in the best fit catalog. Mostly perimeter tracing. Which is what NNs are good at finding and learning anomalies in the frequency domain.

When my interest in watches re-awakened it was after giving a lecture on player piano roll scanning in Bern. I was starting to play around with recovering audio from the grooves in LP disks. Have a paper where someone scanned such with a flat bed scanner. Really surprised no one made an app that could use a phone camera to take a picture and play the LP.

Someone in Japan did train a data set to convert LP recordings to MIDI events. Requires a lot of post processing to actually make piano rolls from the records. One can run the program, which goes out and downloads the training data. But that will put one on the naughty list for exceeding data and time limits.

I wrote the bulk of the color image test suite for the apple color laser printer. A lot of it was perception testing. Color image processing is a three dimensional non linear solid. Most people (and a lot of color scientists) think tri stimulus values are fixed. Like audio decibels, they are relative to each other. So cameras have white point settings for the user to calibrate to. (there is also a corresponding black point.) I find it really annoying the latest apple stuff is attempting to auto enhance my photos, background removal. I do not want anything between me and my pixels. There was also a desire to monetize the color calibration profiles (same as fonts.) Kodak was part of this cash grab too.

Note how many ask to see photographs of watches taken in full sunlight.

Much of what I call 'expert systems' tend to simply use brute force. The goal being the maximum reward for the least amount of effort.
 
Posts
2,334
Likes
3,024
...
I wrote the bulk of the color image test suite for the apple color laser printer.
I remember these! I took a graphics design course where the lab had several. They were really good at the time.

Re: authentication via current "AI" tools. I don't believe it's possible at all, except for the most obvious of examples. There's plenty of fakes that experts struggle to pick out from arm's length without super close scrutiny. You'd need to crack them open and see what horrors are inside.

As I've mentioned previously, I haven't found a model that can reliably figure out the differences between Moonswatch/Speedmaster if you simply remove identifying Swatch text on the dials. It will still provide an extremely convincing summary of why your Speedmaster is authentic. Pretty significant shortcomings.
 
Posts
23,024
Likes
51,474
Re: authentication via current "AI" tools. I don't believe it's possible at all, except for the most obvious of examples. There's plenty of fakes that experts struggle to pick out from arm's length without super close scrutiny. You'd need to crack them open and see what horrors are inside.
I assume you're referring to publicly available LLMs, in which case I totally agree. However, if by "current AI tools", you include all standard ML algorithms, I think the situation is less clear. I see no reason why CNNs, or something similar, won't be very good at authentication with appropriately high-quality/uniform image inputs and good training sets. In principle, it's a pretty basic classification problem, and ML is very good at that.

Where ML will be less good is the details that matter to vintage collectors, where experts know exactly what dial variation was used with which hand style in a given serial number range. But this might not be necessary if the focus is on new/modern models.

Still, it's notable to me that Entrupy indicated a couple of years ago that they were going to add luxury watch authentication to their product line, and that hasn't happened yet. Instead, they came out with a sneaker product. This suggests to me that there may be something unexpectedly challenging about watches, either in the imaging or computational piece of the process.
 
Posts
2,334
Likes
3,024
I assume you're referring to publicly available LLMs, in which case I totally agree
Yeah, I should've been more clear that I meant OpenAI/Anthropic/etc, which people are asking about watches.
I see no reason why CNNs, or something similar, won't be very good at authentication with appropriately high-quality/uniform image inputs and good training sets. In principle, it's a pretty basic classification problem, and ML is very good at that.
I wonder if training on movements and inner case dimensions could prove more successful here. Much more difficult to fake, due to technical complexity and proportions. Counterfeits tend to focus on exterior.