Every time someone asks an artificial intelligence tool to draft an email, polish an essay or write a report, an invisible marker gets attached to the result, classifying it as “AI-generated.” It discounts the human who wrote the prompt, steered the revisions and approved the final version. It also sits on top of a deeper problem: these models learned to write by ingesting decades of human writing, much of it without consent or payment. Humans are erased twice: once when their past work trains the model without credit, and again when their present work gets tagged as not quite theirs.
Text watermarking, at least, was not the industry’s idea; regulators forced it. The European Union’s AI Act requires generative AI providers to mark their output in machine-readable form. The goal is reasonable: as AI deepfakes and synthetic media spread, regulators wanted platforms and the public to be able to tell machine-made content from human-made content.
But the tools built to serve that goal do something cruder than advertised. Text watermarks such as Google’s SynthID work by subtly biasing which words a model chooses during generation, embedding a statistical fingerprint in the phrasing. Anthropic’s Claude models now do the same.
