AI Watch Authentication - Where is your data from, and how clean is it?

If there is one element that AI needs more than anything else, it is data. And, according to AI itself, that data needs to be ‘clean, consistent, and well-organised input to reduce errors and bias.’

Most AI tools we come across are Large Language Models (LLMs) - not best suited for image comparison, but quite good at aggregating what the web says about watches, if that’s all you need.

Using an LLM, mid-authentication response, it may switch to ‘Rolex Marketing Language’, drop in common fallacies about the brand, hallucinate a Datejust's smooth bezel as the more common fluted variety, and even identify a collector’s Omega as a 100% genuine Rolex! If they are right as often as they are wrong, you might as well save time and flip a coin.

Convolutional Neural Networks (CNNs) are better at dealing with images (because, let’s face it, that’s all you have), producing detailed feature maps to compare, but they still need to be trained as to what is right and what is wrong to allow them to build the category sets, either by an expert directly, or by showing them huge data sets that have already been sorted into real and fake…by an expert.

The profusion of authentication services now springing up are unsurprisingly coy about where their data comes from. If they are scraping Chrono24 or eBay, both of which are littered with fakes and Frankenwatches, who is deciding whether the data is right or wrong?

Datasets, of course, need balance, so where are the thousands of images of fake, Franken, and customised watches, from every angle, being obtained from?

As part of a consultancy project, I have been examining and documenting the taxonomy of watches across all major brands produced over the last 70 years. The fact is, the data doesn’t exist in any unified form. It is scattered across myriad websites, forums, and blog posts, often disputed, and certainly not presented alongside the kind of comparison images a CNN would require.

And that is just for real watches – a taxonomy of fakes is next!

I am not a Luddite. Do I use Google image search to help identify watches with no model number more quickly (Thanks… Audemars Piguet!)? Yes, I do, but it is a human (me) double-checking that the served-up suggestions are correct and ignoring the Google search results, which I know from experience are wildly out.

At The Watch Scholar, we teach a process, both online and offline; we don’t offer a tool. We and our fully trained experts gather and record evidence from all parts of the watch and, on a case-by-case basis, compare it to determine whether the story of the whole watch hangs together.

This may not be easily scalable, but the process can be applied to any watch, from any brand, from any era. Not an LLM or a CNN, but good, honest research.