Why AI gets historical gear wrong

Take a Squier Stratocaster with E812345 on it. Under Fender's chart for American-made instruments, a serial of E8 followed by five digits means 1988 or 1989.¹ Under Fender's chart for Japanese-made instruments, a serial of E followed by six digits means 1984 to 1987.² The same seven characters are valid in both schemes and point to different years in different countries. What decides it is the country printed on the decal. To anyone reading the number, the guitar dates from 1988 or from the mid-1980s depending on a word they may not have mentioned.

That is the mechanism this piece proposes for why AI gets old gear wrong, and it is less about bad models than about what models are given to learn from.

A string and its conditions

A language model learns from text. In the text about guitars, a serial prefix appears constantly and the condition that governs it appears only sometimes: the country, the color of the decal, the factory, the model line. So the model learns the string well and the condition loosely. Asked about a particular instrument, it can recall that E8 means 1988, which is true in one country, and apply it to another.

The same pressure works on the brand name itself. "Squier" covered instruments made in Japan, Korea and the United States within its first decade,³⁴⁵ and the name has been used on many more since. Most writing about the name is probably about whatever Squier is now, simply because there is more of it. A question about one instrument from one early era has to be answered against that background. The prediction that follows: the more specific the question, the more the brand's common meaning leaks into the answer.

There is evidence that thinly represented facts are where models are weakest. Nikhil Kandpal and colleagues found that a language model's accuracy on factual questions rose with the number of relevant documents in its training data, and that removing relevant documents lowered it.⁶ Their questions were general knowledge, not gear, so applying the result here is an inference. But a 1983 serial scheme from one Japanese factory is about as thinly documented as a fact gets.

Three tables, no referee

The story usually told about this problem is that models repeat internet myths. The record is stranger than that.

Fender's own pages document only part of it. Its Korean guide says production began in 1988 but its digital records go back only to mid-1993, and that serial dating for Korean instruments made before 1993 is "highly uncertain."³ Its Japanese guide says early records "are not complete."² For amplifiers made before 1994, it says dating by serial number "is all but impossible."⁷

Where Fender's records stop, enthusiasts' records take over, and they do not agree. Fender's Korean guide does not list an E prefix at all.³ One dating site attributes Korean E serials to the Young Chang and Sung-Eum factories and gives a sequence that does not follow the calendar: E1 for 1987 to 1988, E2 for 1988 to 1989, then E0 for 1989 to 1990, with the plant that began in 1987 using neither a 7 nor an 8 as the first number.⁸ A wiki of Squier serials gives an American example in which E9 means 1988.⁴ Fender's own American chart lists E9 under 1989 and 1990.¹

Even the official digits can mislead. Fender's American chart explains that serial decals were printed far in advance, and some N9 decals, which denote 1999, were put on instruments in 1990 by mistake, so some 1990 guitars carry 1999 serial numbers.¹ A reader, or a model, that trusts the digit is wrong by a decade, and only the manufacturer's footnote says so.

So there is no clean division between authoritative charts and myths. For the instruments where questions are hardest, the manufacturer's own account is incomplete or missing, and the tables that fill the gap are enthusiast-made, presented as settled, and inconsistent with each other. A model trained on all of it has learned several mutually exclusive tables and nothing that says which one governs a given guitar.

Why the answer comes out sure

None of Fender's charts gives one year. They give ranges, overlapping ones, with warnings: its American guide says that even its dates for the E-prefix era are "rather loose."¹ A careful human answer to "when was this made" is a range, plus what else to check: the neck date, the pots, the decal.

A chatbot is shaped by different pressures. It is trained toward direct, helpful answers, and "1988" reads as more helpful than "somewhere between 1984 and 1989, depending on where it was made, and the charts disagree." When the record is a set of conflicting tables, the most fluent answer is to pick one and drop the conditions. That turns a fragmented record into a confident sentence, which is the failure the original question described.

What a test would show

This piece set out to test these predictions directly by asking current Claude models, with no web access, a ladder of questions running from the brand ("are Squiers good?") down to specific serials and amplifier date stamps, and scoring the answers against Fender's charts. The test could not be run in the session that produced this article, so what follows are predictions, not results.

If the mechanism is right, answers at the brand level should be reasonable. Answers about specific serials should give single years where Fender gives ranges, and should miss the country condition on strings like E8. Answers about early Japanese instruments should carry the brand's current reputation, calling a guitar "budget" because the name is now a budget name. Where Fender and the enthusiast tables disagree, the answer should follow one without saying there is a disagreement. If models instead give ranges, ask about the decal, and flag conflicting charts, the argument here is wrong about them, even if it is right about the record.

Where this breaks

The record problem is not unique to AI. People misread these charts too, and the tables circulate by being copied from one page to the next. What a model adds is scale and a confident tone, not the error itself.

An enthusiast table may also be closer to the truth than Fender's page. The Korean serial schemes exist in public at all only because collectors pieced them together; when a model repeats one, it may be repeating the best knowledge available. The real fault would then be presenting it without its source, not using it.

And the clearest examples here are serial numbers, the most structured kind of gear history. Questions about pickups, woods or which factory built which model are messier, and they may fail in other ways.

How to ask

The practical consequence follows from the mechanism. A question that carries its conditions gets a better answer from anyone, human or model: the full serial, where it appears, the country on the decal, the decal's color, the model name as printed. An answer worth trusting gives a range, names the chart it used, and says when the charts disagree. When one gives a single year without asking where the guitar was made, it has answered a question about a string, not about the guitar.