Text AI watermarks will always be trivial to remove
I keep seeing new proposals for "watermarking" AI text. The idea is simple: if you make the AI choose words based on a secret key or a specific distribution, you can prove it wrote the text later. It sounds clever, but it ignores the fundamental nature of information.
Information is stubborn. If you can read the text, you can change it. If you can change it, the watermark is gone.
Consider how these systems work. They nudge the model to pick words from certain groups. maybe those that add up to a specific hash value. To verify the watermark, you run a statistical test on the output. If the pattern is there, you claim the AI wrote it.
Now, what happens if I run that text through another model? Or if I just swap a few synonyms? Or reorder a sentence? Even a small amount of "paraphrasing" destroys the statistical signature. And there is nothing stopping a user from simply asking a secondary model to rewrite the output to remove the pattern.
The problem is that you are trying to bake a permanent signature into a medium that is meant to be fluid. If the watermark is strong enough to be detected, it is strong enough to be seen by other models. If other models can see it, they can ignore it.
This is not a technical problem that needs a better key or a smarter algorithm. It is a logic problem. You cannot force a reader to accept a signature if that reader has the power to edit the content. The cost of stripping a watermark is practically zero.
We see this in every other medium. We watermark images, and people crop or adjust the compression to remove them. We embed data in audio, and people pitch-shift the track. Text is even easier because it is just a list of tokens. You do not need to preserve anything except the meaning.
So, should we stop building them? Maybe not. They might work against lazy users who copy-paste directly. But thinking that these will solve the problem of identifying AI text is a mistake. It is just a filter. Filters can be bypassed. It is time we accepted that and stopped looking for technical silver bullets for a human problem.