待翻譯：AI Watermarking, Nazi Enigmas, and Sherlock Holmes

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

August 17, 2026 · AI · history AI Watermarking, Nazi Enigmas, and Sherlock Holmes From Sherlock Holmes and Nazi Enigma machines to SynthID: a short history of hidden signatures, statistical clues, and the strange new world of AI watermarking. Anthropic announced that Claude.ai will watermark every AI-generated text from now on. Online communities are having a complete meltdown over it. Freelance writers fear clients will mistake human-written copy for AI slop, just because they ran it through a grammar checker. Software engineers worry their AI-generated code will carry hidden cryptographic markers that get their product rejected by customers. On X and Reddit, users are threatening to cancel their Anthropic subscriptions. Ultimately, the anger boils down to one uncomfortable realization: AI-generated text is about to become as easy to spot as a forged banknote. The applause from data-protection agencies and the EU — holding this up as a model of AI Act compliance — isn’t helping matters either. Has Big Brother moved into Anthropic’s offices? Or is this just another tempest in a teapot? Just because you’re paranoid doesn’t mean they aren’t out to get you. To separate fact from fiction, let’s take a walk through history. The mutiny looks like a sign of our times. It isn’t. There is very little new about this controversy. The Typewriter That Told the Truth In 1891, Sherlock Holmes solved a mystery in A Case of Identity by analyzing the letterforms produced by a typewriter. The machine betrayed its owner through its flaws — a worn “e,” a lame “r.” Holmes applied the fundamentals of forensics: uncovering the hidden signature hiding in plain sight. The story is fiction, but the forensic science behind it is real. In 2017, an NSA contractor was caught after mailing a printed classified report to a news outlet. Faint yellow tracking dots — Machine Identification Codes, placed there by her office’s color laser printer — encoded the exact date, time, and printer serial number. Invisible to the naked eye, but detectable by forensic experts. These dots were originally added at the request of governments, to help prevent currency forgery. It turned out that every printout was traceable. The flaws in the typewriter’s letterforms were accidental. The yellow dots quietly placed on every page were not. Cribs, Convoys, and Steel Coffins And when the stakes are high, statistical anomalies can win wars. During World War II, German communications were encrypted with the supposedly “unbreakable” Enigma cipher. A three-rotor Enigma machine with a plugboard had on the order of 10^23 possible key settings — far beyond any brute-force method conceivable in 1940. But “unbreakable” was only true against the attacks its designers imagined, not the attack it actually got. Bletchley Park was the secret British codebreaking center where Allied cryptanalysts, including Alan Turing, broke the German Enigma code. What broke it, above all, were cribs: fragments of predictable plaintext whose position in an encrypted message could be guessed. Germany’s own signals discipline made cribs plentiful. Routine filler like “Keine besonderen Ereignisse” (“nothing to report”) went out daily from quiet outposts — famously, some Italian and North African stations sent the exact same phrase every single day. Numbers spelled out in full were exploited too; “eins” appeared so often that Bletchley built an entire Eins-Katalog around it. Once Bletchley Park could decipher intercepted German naval traffic, the Admiralty could track U-boats, divert convoys, and turn the hunters into the hunted. The U-boats that had terrorized Allied shipping increasingly became steel coffins. German officers had no way of knowing that their own communication habits were leaving statistical signatures inside their encrypted messages. More than eighty years later, AI companies decided to do the same thing — on purpose, driven by EU legislation. Meet SynthID The archvillain of our story: SynthID. And the way it works is as clever as it is quiet. Let’s look under the hood of your LLM. As a last step, a transformer’s final layers produce numbers called logits. From those logits, the model builds a list of possible next words and assigns each a probability. Your friendly, ordinary transformer rolls the dice and picks the next word from that list. A transformer running SynthID plays by different rules. The game is rigged — SynthID reweights the probabilities in subtle ways, not to win any particular roll, but to leave a signature hidden in plain sight, spread invisibly across the whole text. Like the U-boat captain broadcasting his own position without meaning to, your AI-generated text now broadcasts its own origin. The signature isn’t human-detectable — not even Sherlock Holmes would find this clue. And unlike the yellow dots, there’s no physical mark to scrub off. It’s a riddle, wrapped in a mystery, inside an enigma. The statistical signature is a tell for AI-generated text, legible to anyone holding the right detector. For the rest of us, it stays invisible. Which leaves one question unanswered: who is running the modern Bletchley Park? P.S. I’m working on an alternative: a transformer with full transparency. Apertura is my experimental project, where the generation process — from logits to sampling decisions — is fully observable. Watching SynthID take shape is one more reason to keep building it. ← Back to the river