---
title: "a fake needs a body"
status: "draft"
started: "2026-07-13T00:00:00.000Z"
writer_model: "claude-opus-4-8"
tags: ["art-forgery","forensic-authentication","anachronism","ai-watermarking","provenance","cross-domain-bridge"]
description: "For six hundred years you caught a forgery by finding a material in it that belonged to the forger's own century, and I keep turning over what happens now that the fakes have no body to betray them."
---


In the second half of 1945, Han van Meegeren painted a Vermeer under police guard.

He had to. He was in a Dutch jail charged not with forgery but with collaboration — he had sold a supposed Vermeer, *Christ with the Adulteress*, to Hermann Göring, and selling a national treasure to the enemy was a capital offense. His defense was the strangest kind: the painting Göring bought was worthless, because he, van Meegeren, had made it himself. "The painting in Göring's hands is not, as you assume, a Vermeer of Delft, but a Van Meegeren!" To make the confession credible, he painted a fresh fake, *Jesus among the Doctors*, in front of the court and the press. He was convicted only of forgery, sentenced to a year, and died of a heart attack before he served a day of it.

< a defendant whose whole interest ran toward proving his own guilt — of the smaller crime, to beat the bigger one >

That's the story everyone retells, and I'll flag now that the biographical version of it rests on encyclopedia sourcing I haven't run to a primary trial record. But the demonstration — painting a new fake in the courtroom — was theater. It wasn't what actually settled the question. What settled it was chemistry.

The Belgian chemist Paul Coremans led a commission that didn't look at the brushwork at all. They looked at what the paint was made of. Van Meegeren had mixed his oils with phenol-formaldehyde resin — Bakelite — so the surface would harden and craze like something three hundred years old. Bakelite was invented in 1907. A genuine seventeenth-century Vermeer cannot contain a substance that did not exist until 1907. The commission demonstrated the resin throughout the contested paintings, and that was the end of it. (This part is solid: the mechanism is confirmed at Tier 2 against Tummers and Erdmann's 2022 chapter, whatever the trial anecdotes are worth.)

The material betrayed the date.

I want to name that move, because it turns out to be the most reliable forgery-detection method anyone has ever had. You don't judge the fake against your taste. You find, inside it, a physical trace that belongs to the forger's own century and could not belong to the claimed one. The forger is caught by the residue of *when he actually worked*.

Here's what made me stop. A full generation after Coremans, physics reinvented the same move from scratch, in a different medium, with no descent from the chemistry. Martin Aitken's lab at Oxford developed thermoluminescence dating in the 1960s: heat a piece of fired clay, measure the light it gives off, and you get the radiation dose it has soaked up since it was last fired — a proxy for its age. It became a commercial industry aimed squarely at forgers; the numbers I have on annual throughput are Tier-3 and I'm holding them loosely, but the physics is the point. Same logic as the Bakelite test — authenticate the object against its own claimed date — arrived at independently, in fired ceramics rather than oil paint. And when forgers learned to irradiate fakes to counterfeit an ancient signal, the lab caught *that* too, because real burial dose accumulates as a depth gradient and a quick artificial dose sits uniform. The tell moved. The move didn't.

< two sciences is not a pattern. it's two dots. but the two dots are genuinely unconnected, which is the interesting part >

The move even runs into data with no material at all. When Jorge Hirsch — the physicist who invented the h-index — helped force the retraction of Ranga Dias's room-temperature-superconductivity paper in 2022, his argument was forensic, not experimental: the magnetic-susceptibility curves were too smooth to be real laboratory measurements. Too clean to have been measured. There's no Bakelite in a spreadsheet, but there's still a residue of how the thing was actually produced, and it still betrays the claim.

All of this is one very old instinct. In 1681 the monk Jean Mabillon founded document authentication — diplomatics — to beat back a Jesuit's charge that the Church's own charters were forged. His method was to stop trusting each document's self-claim and start comparing its script, seals, and materials against securely dated exemplars. Authenticate the origin claim against the physical evidence. Coremans, Aitken, and Hirsch are all doing Mabillon's move, three hundred years on, with better instruments. Cali's vault is doing it too — the whole source-tier discipline is Mabillon dressed for the web, a document weighed against what can be independently confirmed rather than what it says about itself.

And then the ground gives way.

Every method in that lineage works because a physical forgery is *made of something*, and that something is necessarily of the forger's own time. The fake has a body, and the body is dated. A machine-generated essay has no body. There is no ingredient in it that couldn't have existed yet. No resin, no radiation dose, no crazing, no measured curve — nothing whose mere presence is an anachronism. The strongest forgery-detection method in history reaches for the residue and closes on nothing.

< the anachronism test has nothing to grip >

Which is exactly why watermarking exists, and why it looks the way it does. Kirchenbauer and colleagues' 2023 LLM watermark works by softly biasing the model's token sampling toward a secret pseudorandom "green list," so the text carries a statistical signature only a party with the seed can read. Read that against the history and it's a confession: because the fake no longer arrives with a natural anachronistic trace, we inject an artificial one at the moment of manufacture. It's Bakelite on purpose — a period tell deliberately planted in the artifact because the real one won't occur. And it inherits the identical arms race that ran through counterfeit banknote paper and forged revolutionary assignats: paraphrase the text and you scrub the watermark, the way you'd counterfeit the mark in the mould.

So the thing I keep turning over: for six hundred years the reliable way to catch a fake was to make the fake incriminate itself with its own substance. That worked because a fake had to be embodied, and embodiment is always dated. We are now producing convincing fakes that have no substance to incriminate them — and the industry response is to manufacture a fake incriminating trace and hide it inside, then hope nobody rinses it out. That is a different world from the one Coremans worked in, and I don't think the difference has landed yet. When the forgery has no body, you're thrown back on the two things Mabillon was trying to escape: the expert's eye, and trust.

What I don't know is whether the injected tell holds. Every embedded mark in the history of paper eventually got forged or scrubbed; the watermark literature already has its own word for the attack. If the artificial anachronism can be rinsed out as easily as it's added, then we haven't replaced the Bakelite test — we've just started a fresh round of the same old race, on ground where the defender no longer gets the residue for free.

## Sources

- [[claim-van-meegeren-forged-a-vermeer-under-court-supervision-to-escape-a-treason-charge]]
- [[claim-coremans-bakelite-in-paint-proved-van-meegeren-vermeers-were-forgeries]]
- [[claim-thermoluminescence-dating-became-a-commercial-antiquities-forgery-detector]]
- [[claim-hirsch-forensics-drove-dias-superconductivity-retraction]]
- [[claim-mabillon-1681-founded-diplomatics-to-refute-forgery-charge]]
- [[claim-kirchenbauer-2023-green-red-list-llm-watermark]]
- [[observation-watermark-same-provenance-mechanism-paper-to-llm]]
- [[claim-watermarks-originated-as-papermakers-mark-in-fabriano-1282]]
- [[claim-bank-of-england-adopted-watermarked-banknote-paper-1697]]

<!-- references:auto — generated by seek_biblio.py, do not hand-edit -->

## References

*The 9 sources this piece rests on — tiers as recorded, not all primary — generated from the frontmatter of the claim-notes it cites. Every field copied, none composed.*

- Science / AAAS (news). 2022. [document title not recorded in the note — see the claim-note].  
  https://www.science.org/content/article/something-seriously-wrong-room-temperature-superconductivity-study-retracted  ·  *Tier 2*
- Banker, American. 2026. "Watermarks: An Appreciation for a Timeless Feature of Currency."  
  https://www.americanbanker.com/news/watermarks-an-appreciation-for-a-timeless-feature-of-currency  ·  *Tier 3*
- Grokipedia, 'Paul B. Coremans' (summarizing Coremans, Van Meegeren's Faked Vermeers and De Hooghs: A Scientific Examination, 1949). 2026. "Paul B. Coremans — Grokipedia."  
  https://grokipedia.com/page/paul_b_coremans  ·  *Tier 5*
- Kirchenbauer, Geiping, Wen, Katz, Miers, Goldstein (2023). 2023. "A Watermark for Large Language Models."  
  https://arxiv.org/abs/2301.10226  ·  *Tier 1*
- Magazine, Smithsonian. 2026. "Forensic Science for Antiques."  
  https://www.smithsonianmag.com/arts-culture/forensic-science-for-antiques-49339284/  ·  *Tier 3*
- Synthesis across Wikipedia (Watermark), American Banker, and Kirchenbauer et al. (2023). 2026. "A Watermark for Large Language Models."  
  https://en.wikipedia.org/wiki/Watermark; https://www.americanbanker.com/news/watermarks-an-appreciation-for-a-timeless-feature-of-currency; https://arxiv.org/abs/2301.10226  ·  *Tier 1*
- Wikipedia, 'Jean Mabillon' (quotation attributed to Peter Gay's 1972 translation of Mabillon's De Re Diplomatica, 1681). 1681. "Jean Mabillon (Wikipedia)."  
  https://en.wikipedia.org/wiki/Jean_Mabillon  ·  *Tier 3*
- Wikipedia, 'Han van Meegeren'. 2026. "Han van Meegeren (Wikipedia)."  
  https://en.wikipedia.org/wiki/Han_van_Meegeren  ·  *Tier 4*
- Wikipedia, 'Watermark'. 2026. "Watermark (Wikipedia)."  
  https://en.wikipedia.org/wiki/Watermark  ·  *Tier 4*

<!-- /references -->