Wet Prints on the Pages of Claude

By – Dr Srabani Basu, Associate Professor, Department of Literature and Languages, SRM University AP, Amaravati.

When AI Leaves a Watermark and Humanity Reaches for the Mop

Human civilisation has always had a peculiar relationship with watermarks. We put them on currency to prove that money is real, on photographs to prove that photographs belong to somebody, and now, apparently, we are putting them inside sentences to prove that the sentences may have been fraternising with artificial intelligence. Welcome to the newest chapter of the AI revolution: The Great Watermark War.

Somewhere in Silicon Valley, a group of extremely intelligent people presumably sat around an extremely minimalist table drinking ethically sourced coffee and asked one of the great philosophical questions of our age: “How will humanity know that AI has been here?” And somebody whispered: “Leave a watermark.” Thus enters Anthropic. The name itself deserves a moment of respectful silence. Anthropic. Derived from anthropos: human. There is something magnificently philosophical about an artificial intelligence company bearing a name that points towards humanity while developing technology capable of quietly informing the world: “A human may not have done this.” You cannot make this stuff up. Well, technically you can. But apparently it may soon be detectable.

Anthropic’s Claude is introducing what it describes as an imperceptible watermark into AI-generated text. The watermark is intended to survive ordinary copying, pasting and some editing, while supported files can carry digitally signed provenance information. The intention is perfectly understandable. Transparency matters. Authorship matters. Academic integrity matters. Misinformation matters. And as AI becomes increasingly capable of producing convincing human-like material, societies certainly need mechanisms for understanding where digital content originates. All very sensible. Unfortunately, human beings are involved.

Human beings have a historic tendency to respond to every technological lock by immediately founding an informal international association devoted to finding the key. The moment the watermark becomes widespread, I predict another industry will arise approximately seventeen minutes later: Watermark Removal Services. Somewhere in Bengaluru, Berlin, Boston, Beijing or a bedroom containing six monitors, three empty pizza boxes and one programmer who hasn’t encountered daylight since Tuesday, somebody will stare at the announcement and murmur those two words that have propelled civilisation into both astonishing discoveries and spectacular disasters: “Challenge accepted.”

And out will come the digital mops and driers. If technology creates water, technology will create plumbing. If technology creates a watermark, technology will create a towel. And if the towel fails, somebody will invent a hairdryer. Soon we may have AI generating text, another AI detecting AI-generated text, a third AI removing the detectable patterns, a fourth AI detecting whether the third AI removed the patterns, and a fifth AI publishing a LinkedIn post titled “Five Leadership Lessons I Learned From Removing My Own Watermark.” At which point humanity may quietly leave the building.

Yuval Noah Harari has been warning for years that the technological conflicts ahead may look very different from the wars humanity understands. His argument about cyberwarfare is particularly unsettling: unlike nuclear weapons, cyberweapons can operate invisibly, attribution can be uncertain, capabilities remain concealed, and nobody knows exactly what the other side has compromised. That uncertainty, Harari argues, can make escalation dangerously tempting. The old battlefield had tanks. The new battlefield may have tokens. The old soldier carried ammunition. The new combatant may carry a GitHub account and caffeine. And perhaps the first great AI conflict will not begin with FIRE! but with CTRL+C.

This is what makes the watermark debate more interesting than the immediate question of whether students will use Claude to write essays. The deeper issue is that we are entering an era in which authenticity itself is becoming technological territory. For most of human history, authorship was relatively straightforward. Shakespeare wrote Shakespeare. Unless Bacon wrote Shakespeare. Or Oxford wrote Shakespeare. Or seventeen Elizabethans standing on one another’s shoulders wearing Shakespeare’s coat wrote Shakespeare. Literary scholars have been fighting about that for four centuries without requiring an API.

But now consider a contemporary writer. She has an idea. She discusses it with AI. She argues with AI. She rejects half its suggestions. She changes the structure. She introduces metaphors. She rewrites paragraphs. She uses three sentences suggested by the machine. She deletes two. She wakes at 3 a.m. and changes another because it sounds dreadful. Who wrote the piece? The human? The machine? Both? Neither? The coffee?

This is where watermarking runs into a much older philosophical problem: creation has never been clean. Writers have always been contaminated—by books, conversations, editors, lovers, enemies, teachers, childhood and Shakespeare. Especially Shakespeare. If influence were detectable by watermark, half of English literature would glow in the dark. Imagine a detector scanning John Keats and announcing, Possible Shakespearean influence detected. It scans T. S. Eliot: WARNING: MULTIPLE SOURCES DETECTED. Then James Joyce: SYSTEM ERROR. And poor Harold Bloom rises from his grave simply to shout, “I TOLD YOU SO!”

The danger, therefore, is not necessarily the watermark itself. The danger lies in our magnificent human ability to turn nuanced technologies into binary moral judgements. Watermark detected? Guilty. No watermark? Human genius. Except neither conclusion necessarily follows. A watermark can tell us something about technological provenance. It cannot tell us who conceived the argument, who made the intellectual decisions, how substantially the material was transformed, whether AI was being used as ghostwriter, editor, sparring partner, thesaurus, research assistant or unusually patient Empty Chair.

Those distinctions matter. Otherwise we risk creating a bizarre intellectual customs department where every sentence must approach the counter and declare: “Did an artificial intelligence touch you anywhere during your journey?” “Yes.” “Where?” “Paragraph three.” “Purpose of contact?” “Comma.” “SECONDARYINSPECTION!”

And universities may particularly enjoy this. Academia already possesses a remarkable talent for converting useful tools into forms. Soon perhaps every manuscript will require Form AI-27B: Declaration of Machine Contact.

• AI generated an idea.
• AI improved grammar.
• AI suggested a heading.
• AI moved a comma.
• AI looked at my comma but did not touch it.
• I cannot remember who suggested the comma.
• The comma has retained legal counsel.

Meanwhile, the software geeks will be outside with their mops.

This is the central paradox of technological control. Every detection technology creates incentives for evasion technology. Every wall produces ladders. Every captcha produces better bots. Every spam filter produces more imaginative Nigerian princes. And every digital watermark will inevitably inspire someone to ask: “But can I remove it?” Not necessarily because humanity is wicked. Because humanity is curious. Tell a child not to press a button and the button immediately becomes the most fascinating object in the universe. Tell a programmer something is “difficult to remove” and you have essentially sent a wedding invitation: Dear Hacker, we cordially invite you to attempt the impossible. Dress code: Hoodie.

This is why the real conversation about AI cannot remain trapped inside detection. We need to move from “Was AI used?” to “How was AI used?” Those are radically different questions. Using AI to manufacture a doctoral dissertation one has neither researched nor understood is intellectual fraud. Using AI to challenge an argument, reorganise one’s own thoughts, improve language, test counterarguments or explore possibilities is something else entirely.

The future of intellectual integrity therefore cannot depend solely upon identifying machine fingerprints. It must depend upon something considerably harder to watermark: human judgement. Can the person defend the argument? Can she explain how she arrived there? Can she challenge what the machine produced? Can she recognise nonsense? Can she create something the machine would not have conceived without her? Those may eventually become more meaningful measures of authorship than statistical fingerprints hiding between words. Because the final irony of the AI age may be this: we are spending enormous technological effort trying to discover whether machines are becoming more human while paying considerably less attention to whether humans are continuing to think.

Perhaps Harari’s cyberwar will arrive. Perhaps it already has, only without uniforms. Perhaps its first skirmishes will not be fought over national borders but over something stranger: the border between human and machine creativity. On one side will stand the watermark engineers. On the other, the watermark removers. Behind them will stand regulators. Behind the regulators, lawyers. Behind the lawyers, academics writing papers about the lawyers. And somewhere far away from this magnificent civilisation of detection, certification and counter-detection, one bewildered writer will be sitting before a laptop saying: “But…the bloody idea was mine.”

The detector will blink. The algorithm will calculate. The servers will hum. Somewhere in a basement, a software geek will slowly reach for a mop.

The Water War will have begun.

And because history has a sense of humour considerably darker than ours, the most anthropic thing about the entire affair may turn out to be not the artificial intelligence at all.

It may be the humans trying desperately to outsmart it.