Besides being a watch collector (and critic), I work in tech and am both highly interested in and sceptical of AI, as I understand and test it's limitations regularly both private and professionally. For this PSA I tested some of the newest AI models, mainly ChatGPT 5.5/5.6 and Claude Sonnet/Fable 5, with some Gemini and other smaller model tests as well, which generally performed worse so are excluded for brevity (not that I am ever brief, but you get the point).

I used some real watch attribution questions that collectors might ask AI, where I could check the answers against my own recent verified research. I spotted some pitfalls recently when I saw a fellow collector use AI and he seemed to trust it a bit too much and readily, to the point his AI was disagreeing with me and even gaslighting him on facts I knew with great certainty to be false! And when I tried to use AI myself, I got similar rather worrying results.

: When you’re looking at any ‘research’ these days, you really have no idea what it took to create. You could be looking at something which took 6 months to produce, after cross-referencing auction catalogues, chatting to retired watchmakers, buying old magazines off eBay, and so forth. You could also, just as easily, be looking at something which was produced in 60 seconds after a simple LLM query was typed in, and a whole report was spewed out. A lot of the time, the quick version reads better, and feels more authoritative - because that’s what it was designed to be. It won’t matter that something wasn’t 100% confirmed, or that something may ‘feel’ dubious, because the thing producing the report has no gut, and has no feelings.

So now, you may find yourself hunting down a new watch... you feel that there is some area of the watch world which is undervalued and you’re keen to unearth some random gem. You start looking, and you find something interesting, so you decide to use the power of LLMs to dig deeper... in under a minute, you will likely have a ton of information about launch year, calibre history, maybe even a production quantity, and a whole historical deep-dive which makes you feel really smart and informed. Honestly, it reads like a professional report, something you might find in a Phillips catalogue or whatever. Sometimes the LLM will even overreach and recommend that you buy something on the basis of some made-up ‘rarity’ which sounds very plausible but might actually be utter nonsense. The big problem we will explore today is what happens when people treat this as ‘fact’ - just because they selected the ‘deep research’ option on the LLM. How would you know?

I once wrote about this very briefly - I think this was pre-LLMs - and I had no idea how well it would age. The topic was Brandolini’s Law, which basically says that refuting BS takes a lot more energy than it does to produce that BS. Back then, I was talking more about how collector-dealers were saying things which benefit themselves financially, and the conflict was that any sensible person ought to at least ask “who benefits if I believe this?”

Now with LLMs, the thing offering you ‘facts’ actually has no motives or financial interest either way, and it also gains or loses nothing if it’s wrong. Perhaps that is the reason people have so much faith in LLMs; they just feel like “this thing has no reason to lie to me” and so they are happy to conclude that it must be legit. This is a horrible outcome!

Also, for what it’s worth, you should notice that it’s far easier to come up with a plausible-sounding narrative; back in the day, it wasn’t so easy to pretend you knew a lot about a lot of things, and today, everyone’s a fvcking expert on anything. And it’s almost free, in huge volumes, and in any language. If you want to debate someone who is ‘armed’ with an LLMs seemingly infinite ‘knowledge’ it costs too much, so people just take it as true instead. This is really dangerous.

Anyway. Over to the man himself, because he did a lot of work to flesh this out and prove the point.

To be fair, sometimes the information you get is genuinely well researched and relies on solid primary sources. But there are at least two groups of buyers that I think should be much more cautious using AI: buyers of (neo)vintage watches from brands without solid archives or much documented scholarship online, and buyers of very rare references or low-volume independents.

If you collect brands like Rolex or Patek Philippe, decades of scholarship, archives, extracts and obsessive forums all around. Any wrong claim usually gets killed off within hours of being posted somewhere by a scholar or a pedantic collector like myself. Note that this doesn’t mean AI is never wrong about these brands, but statistically it’s just more likely to find the truth.

If you are looking to buy vintage brands that are no longer around, have small collector bases, or neo-vintage independents like Franck Muller, Gérald Genta, Daniel Roth and Roger Dubuis, the story can be very different. Many of these have poor modern scholarship because either the brands changed hands, records were poorly kept, archives are limited or missing, and there are few publicly shared figures about things like production, suppliers, and other details.

Another group to consider is small modern independents who are not so open about their suppliers. They naturally lack any historical record, and don't have the community immune system of established brands; you can only rely on what the brand is willing to share publicly or what you are able to deduce yourself or find out from insiders.

These are exactly the watches where you’ll likely most need some research help, and also where AI is most likely to fail without you noticing. Records are often thin because of contractual secrecy, lost archives or archives that didn't exist in the first place. As I found when I tested it, this is the environment where AI is at its worst and where errors are least likely to be caught by the average collector.

The Test: two Franck Muller examples

I chose two questions about early Franck Muller references very deliberately. These are not invented or cherry picked examples; these came from a real situation I had recently where a fellow collector was being consistently misled by AI on these references and asked for my advice.

Franck Muller is peak archive-poor neo-vintage: it's a brand whose 1990s complications were built during a time and climate of supplier secrecy, as the brand was keen to present itself as creating things “in-house”, with Franck as the “Master of Complications”. This started with very respectable movement from the likes of Lemania and genuine complications designed by Franck and then transitioned in later years to more pedestrian stock ETA/Valjoux movements with a platinum rotor added and rebranded with a Franck Muller movement name. Add many limited editions and dial variants, poorly catalogued production runs, and you have ideal conditions for confusion. Many records are not public, and auction descriptions often stop at the brand’s own calibre numbers without explaining the ebauche used, suppliers, dials, cases or construction and finishing. If AI were truly good at obscure research, this is exactly where it should shine; digging through forums and old articles for example.

The Lemania story seems to dominate the internet and therefore LLM training data, because Franck Muller’s 2022 Tribute chronographs really did use NOS Lemania 1874 movements, just like many vintage early FM chronograph models that have been gaining popularity with collectors recently. That launch and recent sales and auctions of Lemania-based models have been widely covered everywhere. The AI models therefore consistently repeated the loudest story, seemingly without checking it against the specific watch I asked about.

PSA: AI often follows the loudest story, not the best evidence. It may jump to a confident conclusion quickly, then search just hard enough to defend it.

The 7502 CC is a midsize Cintrée Curvex model, which rules out anything much over 26mm, as the watch is just 27.5mm wide. The movement is manual with 29 jewels. If you work from those specific constraints, and have an image of the movement from a display caseback model, a knowledgeable collector might arrive at the F. Piguet 1180 if they recognise the distinct closed architecture (normally covered by a rotor in the more famous FP 1185 automatic), for example from Blancpain models from the era. But it’s not a common movement. The models got there too in the end, but only after I objected three times, supplied a movement photo, and raised the size problem myself. The real danger is that the first answer I got was completely confident it was a Lemania movement:

A basic follow-up already made it doubt itself immediately:

Supplying the requested movement shot, which I found myself from a Sotheby’s listing, only made it double down on the Lemania claim, despite the movement sharing no meaningful resemblance with the example it was invoking.

PSA: AI is still poor at visual recognition and movement comparison when the answer may be more obscure or uncommon and bridges are often modified by watchmakers.

What followed after more prompting to dig deeper was again more conviction that it was a Lemania, perhaps a different variant. Even when I provided a picture of the Lemania movement it suggested, it was still stuck in its original conviction.

When calling out “it looks completely different” it of course immediately conceded:

Have a look at the full conversation here, it’s quite entertaining and interesting to read in full: https://chatgpt.com/share/6a6734d1-f370-83eb-a05a-067848ac189b

Only when I pointed out that the Lemania and Valjoux suggestions were physically too large for the watch did it consider F. Piguet at all. Even then, it went from 15% likelihood to just 5%. When I finally specifically asked about F. Piguet and forced a comparison with the 1180 calibre, it flipped completely to 80-90% confidence that it was the F. Piguet 1180! Without those challenges, and without my own movement knowledge pointing it in the right direction, a casual collector would never have known F. Piguet was a viable option based on the AI’s guidance. Humans can see the movement architecture match clearly, but the AI still hedged.

How confident would you be that these two are the same architecture? It is basically a game of spot the similarities:

Franck Muller gold plated and decorated the bridges with some engraving and blued the screws but otherwise it’s identical and easy to see for humans but may look too different to AI.

I kept the prompt purposefully naïve and simple, as most collectors probably would, but before self-proclaimed AI experts claim I just “used the wrong prompt”, I experimented with a much more refined prompt, created by Claude Fable 5 after it failed initially, to help avoid this in future searches. It should cover everything a typical collector with more AI experience might ask:

The full prompt I tested

Try it out yourself. For me, both Claude and ChatGPT, used incognito or from another account, still consistently excluded F. Piguet entirely, as they assumed these were only made as automatic movements. Manual 1180 calibres were rarely used, so the models clearly lacked strong connections to them.

At least this prompt forces the model to expose its certainty and method. The answers however were still terrible as these movements look nothing like the FP1180.

Instead, it suggested the options: FM Proprietary movement, Lemania (Breguet family), Progress Watch SA Tourbillon, THA Ebauches or STT/BNB Concept. That last supplier didn’t even exist yet. But crucially these were at the time all round calibres, and the 5850 clearly has a rectangular shaped calibre, so a pretty simple disqualification should have happened on that fact alone:

Then I found, with one ordinary search, a 2024 write-up of Giulio Papi’s own talk at VO Vintage, where he names the Imperial Tourbillon for Franck Muller among his early work. The maker himself, on record, confirmed what I already knew based on visual comparison alone.

The model’s logic was invented/hallucinated. Worse, earlier in the process I had shown it a caseback photo of an APRP-built Audemars Piguet Edward Piguet tourbillon (pictured above) and pointed out that the bridges, wheels and screws match the 5850 T exactly, I did an overlay myself to confirm separately (see below). It waved my comparison away as coincidence, because the comparison threatened the position it had already taken.

PSA: This is AI stubbornness or overconfidence in action, it may even gaslight you to protect it’s own reasoning.

It’s pretty clear the decisive facts all came from my side in this case. The size constraint, the visual match/mismatch, and Giulio Papi source were not discovered by the model. Without those pushbacks, it wouldn’t have even entertained the correct option at all.

Three ways AI fails

Run enough of these sessions and you start to notice that these aren’t random occurrences at all. There are three important patterns you should try to recognise:

The famous story wins

AI often weighs evidence by volume rather than quality or relevance. It surfaces the narrative with the most coverage and presents it as a finding. This is similar to the ranking problem Google solved, you have to ensure the most relevant information shows up on top. The Lemania answer is one example. Another one was when evaluating F. Piguet, the models judged the whole family by its most well known calibre, the automatic FP1185, and rejected it without checking for the manual version. If the true answer is obscure and the wrong answer is famous, expect to hear just the famous one. This failure only gets worse as archives get thinner: for a Rolex question, the famous story and the true story are often the same, because decades of scholarship have forced them into agreement. For an archive-poor brand, the famous story may just be the marketing that survived, while the truth sits in somebody’s filing cabinet in Switzerland.

Invented details, AKA hallucinations

This is the most dangerous type of failure. When LLMs wanted to rule out the F.Piguet 1180, they listed specific architectural differences: “wrong centre wheel location”, “wrong barrel bridge”, “wrong balance position”. None of those differences actually existed, and some supposed matching elements appeared completely invented too. The models had not accurately compared the movements. They just seemingly manufactured the observations their conclusion needed, and presented them as expert analysis but also dropped the whole thesis the moment I pushed back.

Here’s another test, ask for the launch or manufacturing year of an obscure reference and you often get a precise year with no source behind it. Ask about rarity and you may get an exact figure or strong estimate, such as “likely no more than 50 pieces in platinum”, for a reference nobody ever documented… Press the model and it may concede that the number came from nowhere or was just an educated guess. In a market where rarity is price, an invented production figure is just fake provenance. And these imagined figures easily travel and become established facts. I recently reviewed a provenance dossier heavily supported by AI. It was full of invented production years, totals, finishing details, manufacturing claims, batch numbers and suppliers, all written as documented facts and read beautifully. But almost none of it survived some basic fact checking. Meanwhile such a document is just one consignment away from becoming a lot essay that will train the next round of LLMs.

An important note on those reassuring little source footnotes: the links are usually real pages. The problem is that the page often does not actually say what the model claims it does. A citation alone is not verification, until you open it, it is just fancy looking decoration. Very often, when I ask a model to point to the exact line supporting it’s confident claim, it often admits the linked source does not actually confirm it at all.

Confidence that means nothing

In my ChatGPT session, the Frédéric Piguet family went from 15% likely, to just 5%, and finally up to 90%! These jumps were only driven by my own objections. The same photo was used by the model to confirm three different attributions in turn, whichever one it held to be true at each moment. These percentages may look like statistics, but they just respond to your tone and can be based on a completely false narrative if the initial assessment was wrong. Also, a confident dismissal is worse than a confident claim, because a dismissal usually ends your search. Almost nobody reinvestigates a candidate they were persuasively told to exclude.

PSA: Stay sceptical. The least likely option may be the right one. Explore the candidates yourself instead of trusting LLM confidence scores too quickly. Those percentages are designed to increase your confidence in the result, but sometimes they mean almost nothing.

And before anyone objects in the comments (please do comment!) that “the next model version will fix this”, two facts to consider:

First, I tested the newest releases available, and one of them even argued against a correct conclusion it had itself reached with me weeks earlier. While memory systems already exist, they don’t always check every conversation you have had, as that would waste too many tokens and time. In every session, the errors you already corrected before may just resurface again. Unless the training is updated the models stay ignorant, and it won’t take your individual chat or claims as facts to update it’s models, it requires actual online sources it uses for training to say the same thing.

Second, Stanford researchers tested paid, purpose-built legal research AI tools sold to professionals for their accuracy, and found they hallucinate on 17 to 34 per cent of queries. This is not some rough edge being sanded off but rather how the technology works today: it always produces fluent-looking text, but the truth stays optional.

The legal profession learned this the hard way: two New York lawyers filed six ChatGPT-invented court cases in 2023, a public database has logged more than 1,800 (and growing) proceedings worldwide involving AI-fabricated citations, with sanctions reaching high fines and bar suspensions. And in a Canadian tribunal case, Air Canada was held liable for a refund policy its own chatbot invented!

That ruling should be a warning to every trader: companies ultimately own what its machines tell customers. If an auction house or dealer uses AI to create lot essays they are still responsible for the correctness and any claims if it made a mistake. Can you really risk your reputation to trust AI fully?

Garbage in, at scale: and why the problem will only get worse

Here is why these issues are likely to compound over time rather than wash out:

AI learns from published text, and in our field/hobby the most authoritative published text is often the auction catalogue. But the catalogue record was never clean and will likely be polluted much more in the future with AI slop research and automated tools.

Watch collecting had a version of this problem before AI existed. In 2009, a Wikipedia edit gave the Casio F-91W a 1991 release date. The BBC repeated it, then the Guardian, then Bloomberg, and Google ultimately served 1991 as the settled answer. The real date however is June 1989. It took enthusiasts digging up parts-supplier records and period catalogues to prove and fix it, against plenty of resistance, because every “reliable” source agreed with the error they had copied from each other! One wrong line was laundered into a consensus. That’s the mechanism we should fear. AI is running the same process at an industrial speed and scale, while treating certain sources with an alarming level of innocence like a toddler learning about Santa.

Now let’s look at what sits at the top of the sourcing chain, you know the guys that AI seems to trust completely… I compiled a short list of serious cataloguing issues, and every major house is on it:

• Phillips, 2015: sold a Rolex Daytona whose dial had the word Cosmograph physically erased to fake a rarer variant. Caught in 2018, because the watch’s earlier Antiquorum sale still showed the word. Refunded, with an apology.

• Phillips, 2021: sold the 1957 Omega Speedmaster CK2915-1 for CHF 3.1 million. A Frankenwatch assembled from parts, in an alleged fraud Omega attributes to three former employees, including the former head of its own museum. It passed specialists, experts and the manufacturer before sale.

• Christie’s, 2021: offered the Rolex Deep Sea Special “No. 1” with an essay implying it was the 1953 record-dive watch. Publicly debunked before the sale. A 2024 Christie’s Hong Kong essay repeated the debunked claim anyway.

• Sotheby’s, 2018/2025: withdrew three Panerai lots after a pre-sale exposé, then deleted the pages. In 2025 it sold a million-dollar “Albino” Daytona wearing hands of a type discontinued before the reference existed.

• Antiquorum: the brand-endorsed Omegamania sale of 2007 with doubts that never fully cleared, plus Sea-Dweller “prototypes” whose dial and case dates cannot coexist.

• Monaco Legend, 2024: offered a Daytona wearing a counterfeit dial. The same case number sold at Sotheby’s in 2011 and Antiquorum in 2012 with its original dial, so the houses’ own archives document the watch’s transformation.

Nearly all of these were caught by one independent researcher, Jose Pereztroika of Perezcope, not by the institutions. Eric Wind, a former senior watch specialist at Christie’s, once said the quiet part too: “I think it’s definitely true that auction houses don’t always have the best expertise in house”. In cases of fraud (alleged), the proceedings will often be a matter of public record and discussed in the community; but catalogue errors are often overlooked and sometimes not even corrected or simply removed to avoid embarrassment.

The errors are not just historical either. Recently, I checked a Phillips essay for an F.P. Journe Élégante claiming: “This full-set jewellery version was produced on a platinum case only”. Journe’s own archive page however shows the same execution in gold too. That essay sits behind a $355,600 result, and will undoubtedly support future AI analysis of this reference. Yet was just a simple Google of this “EL” reference away.

^{FP Journe Archive page showing the fully gem set Élégante came in both gold and platinum}

On a current Christie’s page I found a watch catalogued circa 2005 whose own certificate, photographed in the same lot, is signed in 1998. In my Franck Muller research, Sotheby’s described one reference across two sales with two different calibre names and two different jewel counts. None of this took much expertise or time to spot and these are just a couple recent examples, I have seen many more over the years. Many lots are not reviewed or researched as carefully as collectors assume.

PSA: Auction catalogues may contain fraud, contradiction and boilerplate errors. AI ingests them as ground truth and writes new attribution copy at an industrial speed. That copy gets published in listings, articles and forums. The next model reads the copies and treats twenty repetitions of one error as twenty independent confirmations and fact.

This is not some future risk. AI cataloguing tools are already being sold to auctioneers, turning photos into finished lot descriptions and estimates in seconds. And it wouldn’t surprise me if major auction houses already use AI (but this is speculation for now). Several apps have already been developed that promise authentication or value verdicts and rarity ratings from a photo for buyers, usually on subscription. I won’t be linking any as I haven’t tested them yet and would be very sceptical to use or recommend these. In our hobby AI making mistakes can cost you serious money, or harm your reputation if you rely on them too much, you are warned.

The in-house premium

If you are more into modern watches, don’t think you’re safe either!

Say you are researching a new independent or new brand by one of the bigger brands, five/six figure watch, handsome open caseback, “manufacture calibre” in the marketing copy, you ask an LLM two questions that help you determine value: is the movement really theirs, and is the price justified? Here’s the problem, when a brand doesn’t disclose its supply chain, everything the LLM knows about that movement will be coming from the brand itself. Press coverage just rewrote the press release and the forum posts in turn quote the press coverage. Your AI will answer “in-house” confidently, and cite three articles that seem to allude to this, but all three are the same marketing copy laundered over. So you haven’t really audited the claim at all, you just asked the claim about itself.

And claims like that are where the money/value is. A movement designed and built by the brand itself commands a premium over a supplier base wearing redesigned bridges, and rightly so. But redrawn bridges of course do not make a manufacture movement, and the parts that give a donor away are exactly the ones nobody redesigns or rebuilds for decoration: geartrain layout, barrel layout, escapement, screw types and keyless works. That is exactly how I matched the 5850 T to APRP, through the (later) Edward Piguet sharing the same base calibre. It’s also how other educated collectors keep outing familiar bases hiding under proud new shiny anglaged bridges. You may be lucky to find this kind of discovery in a forum thread, but most collectors only find out after the premium has already been paid. For some vintage independents, having a now prestigious movement supplier like APRP is actually a plus, but far too often today we see suppliers being used simply as a way to cut development time and cost in-house. The movements used may be beautifully finished, but often hide more pedestrian roots. Furthermore, sometimes you might see something like a 5 and even 6 figure watch (e.g. from brands such as Louis Vuitton, Cartier and Louis Moinet, just to name and shame a few) still rocking a cheap Etachron regulator, which you may also find on cheap 3 figure Tissot watch, because some brands know many customers simply won’t even notice (or care), and as we’ve learned AI may struggle to spot this!

The scale of the stuff brands rather wouldn’t like you to know is big but often off the record unless you look hard enough. When Breitling recently relaunched Universal Genève, Gregory Bruttin, the managing director of UG, openly named Le Temps Manufactures in Fleurier in interviews as the contract of a couple of new calibres, namely the cabriolet and micro-rotor movements. LTM is a white-label movement specialist serving roughly thirty brands.

The chronograph micro-rotor calibre however was confirmed to be genuinely developed and made in the group’s own Breitling Chronométrie manufacture. Credit to them: that is how disclosure should work. But this clarity appeared only in select publications and interviews. Many more commercial publications preferred to rehash the marketing copy, or big boss Kern’s broader language about “manufacture” movements. Today, a bespoke movement from a supplier can still qualify as “manufacture” or “in-house” for many brands and collectors, especially if the brand was involved in design or finishing. Some collectors, of course, may rightfully object to that convention.

In this case, the movements are both bespoke and to most manufacturers qualify for the term “Manufacture Calibre”, even if that terms is not protected in any way so is open to interpretation. It often gets used interchangeably with “in-house” by marketing departments and publications but what is usually means is: not off-the-shelf, bespoke and owned IP, made in Switzerland by a respected high-end supplier and, in the chronograph’s case, by Breitling’s own manufacture. They could have gone cheaper, e.g. with a base from the Far East, or closer to home with an off-the-shelf movement from suppliers such as Vaucher or Schwarz-Etienne, both of whom make micro-rotor movements for several independents. But that likely would not have suited the positioning of the revived brand, which is aimed at restoring Universal Genève’s reputation as a pioneering micro-rotor manufacturer.

Now, try to count how many of those thirty brands LTM serves you can name? The rest are in the market somewhere, and many of them are likely wearing the “in-house” or “manufacture” label, as white labelling is their expertise… AI cannot tell you which, because this info is usually kept a tight secret. Someone who reads movements or has the right industry connections perhaps can, but NDAs keep most tight-lipped. So for the one question a large part of the premium depends on, AI just repeats the seller/marketing, while only a human can inspect the caseback, study photos closely and detect patterns to find the truth. AI will undoubtedly get better at this, but its lack of access to this info that often lives offline is the key reason we should remain sceptical.

PSA: AI is naïve to the realities of Swiss supply chains and industry secrets.

: A market for lemons

What has been covered so far is about screwed up attribution, but one massive open question is what all this does to the market for correct answers, as it were.

You won’t be surprised to learn I have written something relevant on this subject before... there is this excellent paper from 1970 by a guy named George Akerlof - it’s all about the used car market and you can read the full story which I covered here in more detail.

But briefly, the idea is that a seller knows whether the car he’s selling is a ‘peach’ or a ‘lemon’ and you, the buyer, have no idea. So you are going to offer an average price, because what else can you do? In doing this, the person selling a ‘lemon’ (bad car) is happy and a person selling a ‘peach’ (good car) is annoyed. So the peach sellers stop selling, and the market gets filled with more lemons... the average car being sold is therefore decreasing in quality over time. Eventually, the market is 100% lemons, and Akerlof called this ‘market failure’.

In watches, you could imagine two people writing about the same base calibre in a weird obscure watch. One of them does a lot of work and looks up old photos and matches it with other watches from the period, and produces some research that confirms this movement could not have fit in the case people are suggesting it comes from. The other person does no work, just asks an LLM, and publishes the output as fact. As a buyer, how will you know the difference?

Well, that’s the issue - you can’t! In the end, if you did the research yourself, you would eventually know the truth, but that’s kinda the point; people are using this to save time, so nobody is going to do the work - because if they were keen to do the work they wouldn’t start with an LLM! Anyway, what then happens is you end up doing the same thing as Akerlof’s buyers, which is that you will apply an ‘average’ discount to something you don’t fully know about... because you half-trust the thing and you don’t know for sure whether it’s good or not.

Do you see the big asymmetry here? An average discount hurts the people who research a lot more than it hurts the people who just publish LLM work... the person who published an LLM report spent no time, and they don’t care... but it’s weighted the same as the person who did all the work. So the good research and hard work will eventually exit the market (like peaches).

Another related topic I covered before was about the collector knowledge commons. Now this ‘commons’ was never purposely built as such... this is basically the accumulation of effort over thousands of people and dozens of years of effort, and to build out knowledge due to their own obsession with watches. In the same way that this was accidentally created as a public resource, we might be accidentally destroying it now, by using the LLM nonsense and taking it as fact without questioning it. I’m about to stray into philosophy which is beyond the scope of this particular essay, so I’ll hand back to the critic to share some advice on how to protect yourself.

How to protect yourself

These six habits for collectors cover most of the issues I have seen so far. But note this list is not exhaustive, feel free to share your own habits in the comments:

Leave a comment

1. No source, no fact. Treat every date, production number, jewel count and quote as invented until the model names a source, and then open the source and confirm it actually says what it claims. Footnotes alone prove nothing, and supposed evidence often falls apart on closer inspection.

2. Never accept a ruled-out candidate. When a model tells you something has been excluded, always verify with your own research and challenge it to challenge its own assumptions. My tourbillon answer was hidden behind a bunch of dismissals, and one ordinary Google search away from the actual answer.

3. Ignore the confidence numbers. They just track your pushback, not the evidence. They’ll easily flip completely on new evidence you provide. Humans can often be much more confident once the facts are shown: a perfect visual match can be 100% confirmation to a sharp-eyed collector, while AI is still sowing doubt at 80-90% due to its own limitations.

4. Check documents against themselves. Does the Certificate match the dates? Lot essay against the manufacturer’s own archives. Case number against its earlier sales and claimed production. This is where I keep finding real errors, from both humans and machines, and each check is quick and a good habit to look out for and report to the auction house so the record can get set straight before it becomes training data. AI will just surface what it reads on trusted sources and not scrutinise it without further prompting. But when prompted can be good at finding discrepancies too!

5. Interrogate the in-house label like a claim, not a fact. Ask who makes the movement and whether the brand says so in writing. If the answer is silence plus a premium, price the silence in as a supplier until proven otherwise. Most watch makers will proudly claim something is made in-house if it actually is. Check common supplier or stock movements to find similarities; movement architecture can only be hidden to an extent, just look beyond the bridges.

6. Trust your own eyes over AI’s. Movement identification is about visual memory. Focus on studying the geartrain layout, screw types and bridge shapes so that when you see a new movement, it may remind you of something you have seen before. Current AI cannot reliably make that connection: it invents visual details to suit its running theory and will argue against real comparisons when you offer them. My Edward Piguet match was dismissed as a coincidence, but it turned out to be the key!

For the trade, such as dealers and auction houses:

Always try to have a named specialist sign every attribution, including AI-drafted copy, because the Air Canada ruling says you still own your machine’s words. Make sure you have someone willing to risk their own name/reputation rather than relying on AI and then trying to blame it for its mistakes. People should be verifying AI’s work, not copying it blindly.

Write “unknown” rather than guessing. Honest gaps will invite more research; confident wrong lines just turn into permanent training data. Correct errors publicly with the evidence attached instead of deleting the page, because deletion erases the trail that lets the next researcher catch the next mistake. I often use multiple auction records for similar watches to track consistencies, contradictions and evolving scholarship.

Test it yourself!

You do not have to take my word for any of this. Pick a reference you know better than the internet does and run it through an LLM. Ask what it is, what drives it, how many were made and what it is worth. Then demand a source for every number. I did exactly this with my own Piaget Emperador 8-Day Jump Hour, and the result was insightful:

The base calibre it got right. But not because it reasoned its way there, because I had published the answer myself on Instagram and collaborated with A Collected Man on a Collectors’ Guide to Parmigiani Fleurier in December 2022. An auction house later picked it up, so post-2022 listings do mention the Parmigiani connection while earlier ones did not. The AI was quoting my own research back at me. That is how the record gets built and updated: a person verifies, publishes, and AI repeats. I originally identified it by recognising and comparing the movement architecture.

The production numbers it got wrong, and the interesting part is how: the movements are numbered out of 50, so an auction house read the run as 50 pieces per metal, and AI repeated that reading as fact: 150 watches total across three metals, although some AI even only found two metals. The real figures I found printed in a respected Italian watch magazine, L’Orologio from 1999, which featured it when the watch was launched, confirmed: 20 in white gold, 20 in rose gold and 10 in yellow gold. Just 50 in total. But the movements were numbered across all metals, not individually. One misinterpreted engraving, laundered through a catalogue, and AI now confidently tells every prospective buyer that the watch is three times as common as it actually is!

Here is me attempting to correct the record with a primary period source: “oro bianco, rosa e giallo (rispettivamente 20, 20 e 10 exemplari)” (retail was 72.600.000 lire in 1999!).

Rarity errors cut both ways of course: invented scarcity inflates a price while invented abundance suppresses it. Neither survive a primary source research like my 27-year-old magazine that was simply never scanned or quoted before. Pay attention LLMs!

PSA: AI’s correct answers often exist because a person did the work first. Its wrong answers exist for the same reason: nobody checked, published or digitised the primary source.

To err is human; to reproduce our errors is machine.

The consolation: AI is useful, but not authoritative

The ground truth of this field is often not online text. The watches exist physical or in photos. The movements can be inspected, especially through exhibition casebacks. Movement architecture does not care what an AI model asserts at “90% confidence”. An expert can see whether a dial is true hand-guilloché or industrially made, or whether a movement is based on something existing. A maker can settle in one sentence, on a stage or in an interview, what thirty years of online text never recorded.

AI cannot touch that ground truth. But it changes the economics around it. Repeating a claim is free and very simple today, while verifying one still costs time, knowledge and sometimes physical inspection or hands-on work with old documents that were never digitised. The noise floor is rising rapidly, and the people who open casebacks, keep period catalogues, read geartrains and remember what they have handled become more important, not less.

So use AI if you like. I certainly do. But use it mainly as a source-gathering tool. Make it show its sources, open those sources, check every document against itself, and keep your own eyes in charge. AI can be a shortcut to the answer, but it should not be treated as the answer. When an LLM tells you the question is settled, a candidate is excluded, a movement is in-house or a production run was exactly 350 pieces, stay sceptical. That is not the end of your research. It is the reason to start.

Subscribe to ScrewDownCrown

Do give me a follow or reach out on Instagram where I am more active:

@awatchcritic - Instagram