A few days ago, I saw a long thread on X. It was about how writing generated by AI could contain a watermark invisible to the human eye. Can it really not be removed? What happens when it is translated? If a person changes it a little, whose writing is it? As the replies grew, the debate shifted toward the question, “Can you prove that a human wrote it?”
When we hear “watermark,” we imagine an invisible stamp pressed into the writing. The actual method is closer to a casino than a stamp. When the model chooses the next token, it divides the candidates into two groups according to a secret rule and slightly increases the probability of one group. Once is a natural choice. After hundreds of choices, one group appears more often than chance would allow. A detector with the same secret key counts the accumulation of this small tilt, not the meaning of the sentence. It is as if AI secretly highlighted every sentence with a fluorescent marker, except that the fluorescent color is invisible to the human eye.
This method has been studied in practice, and Google DeepMind applied SynthID-Text to Gemini outputs. The EU AI Act also points toward making synthetic text machine-readable and requiring it to be detectable. But the law does not mandate any particular watermark, and public research alone cannot establish that every commercial model contains the same device.
The side making the window plants a trace at the moment of generation. The side raising the shield changes the sentence. If you copy it as is or make small edits, the signal may remain. But if you translate it, have another model rewrite it, or take only the meaning and create new sentences yourself, the original token arrangement breaks down. Instead of erasing the ink, you are copying the same content onto a new sheet of paper. It is less like changing your face to get past immigration and more like having an entirely different person enter the country in your place.
When methods that withstand editing appear, attackers estimate the words a model prefers through repeated queries. They may not stop at erasing the trace. They may also plant false signals in human writing. Leaving and erasing, stealing and forging, each create the other’s next move. Short texts do not even provide enough of a sample. Apply a small false-positive rate to hundreds of millions of texts, and it comes back bearing the names of real people. In trying to catch machine-written text, we may end up summoning humans first simply because their sentences are neat.
Will there come a moment when we can no longer distinguish AI writing from human writing? If we look only at the text, it is possible. The same sentence could have been written by a human, generated by AI, or copied out from memory by a human after reading an AI sentence. There are cases where the result contains no information that can reveal its cause.
So verification moves outside the sentence. We record drafts and intermediate versions, pasting and model calls, and the time of each edit, then attach a digital signature to every step. But a record showing that AI was not called from the editor is not proof that you did not look at an answer on the phone beside you. A signature guarantees only that a particular key signed the file. It does not guarantee where the thought began.
Then let’s build a cell with no windows or radio signals. We install 360-degree cameras on the walls and ceiling, with no blind spots. Before entering, we inspect the person and their clothes, and leave only fresh paper and a pen with ink. Let’s also attach a certificate confirming that, fortunately, the pen still cannot think. We film every movement as one person writes for several hours. This would be the closest thing to certifying, “I wrote this by hand without outside technological help during this time.”
But the person could have memorized AI-generated sentences before entering the room. They might have turned someone else’s thoughts into their own sentences. We can check the person holding the pen, but not verify who first invented the thought. To inspect memory as well, the camera would have to look inside the brain. In trying to certify human writing, we end up demanding that people submit their inner selves.
Would extending the surveillance time solve it? Cut off the internet for several days before admission, and if that is still not enough, isolate them for several months. To erase even the news AI selected and the expressions it translated, we would have to raise people who had never seen a digital medium from birth. At that point, we would not be certifying human authors. We would be breeding humans who can be certified. We would be putting an organic label on a person in an attempt to verify the origin of a single piece of writing.
Even so, the market will not stop. Book covers may carry labels such as “Written solely by a human,” “AI research, human writing,” or “Human first draft, AI editing.” Publishers may ask for writing records before the manuscript, schools may require students to submit their thought process along with their answers, and literary prizes may define which tools are allowed. Is brainstorming okay but sentence generation not? Is searching allowed but summarizing not? Which side do translation and spell-checking belong on?
In the end, “written by a human” becomes not a boundary in nature but a production protocol set by society.
Manufactured goods cannot avoid the same question. How long will we continue to trust the “handmade” tag on a luxury handbag as evidence of quality? Surgical robots already turn a doctor’s hand movements into smaller, more precise motions. There is also less and less reason for a robotic arm to be worse than a master craftsperson at cutting, tanning, and sewing leather. The stitches may be straighter and the defects fewer. Then the value of handmade comes not from the quality of the finished product but from the record that a human hand actually touched it. Bags and sentences alike become goods we buy not only for how well they are made, but for who made them and how. One day, a craftsperson’s crooked stitch may become not a defect but a watermark certifying its humanity.
Once a protocol appears, a certification industry follows. Writing apps record screens and keystrokes, insurers assign higher risk to manuscripts with gaps in their records, and platforms may place writing with a complete provenance at the top of search results. At first, the mark was meant to stop fake writing. At some point, thoughts that were not recorded become a disadvantage. A sentence written on a café napkin, a metaphor that came to you while walking, a hesitation shown to no one may become writing that is considered less trustworthy simply because it has no birth certificate.
Writers may then split into two kinds. Writers who disclose the entire creation process to earn trust, and writers who defend the right to leave no record at all. The former accept more surveillance to prove they are human. The latter protect their privacy but accept suspicion that they are AI. A system created to protect human writing ends up demanding that only humans produce an alibi for their own thoughts.
A strange reversal happens here. AI produces thousands of pieces of writing with the push of a button and quietly goes home. A human submits a record of their life to have a single sentence recognized as their own. The cost of machine production approaches zero, while the cost of proving humanity keeps rising. In the end, “written solely by a human” may become a luxury label earned by those who can bear slowness, isolation, and the cost of surveillance rather than a guarantee of quality.
There is no guarantee that writing produced in an analog cell is truer or more beautiful than writing made with AI. But the act of choosing alone before a blank page remains.
Perhaps what we are trying to protect is not a pure sentence but a place for choice. Who left the final sentence, and in whose name will they take responsibility when it is wrong?
