Anthropic has a new scheme for watermarking AI-generated text, and it hinges on something most readers would never notice: inconsequential words. The idea is to subtly shift word choices — the kind of filler and function words that don’t change meaning — in ways that create a detectable statistical pattern. Think swapping ‘a’ for ‘an’ in rare cases, or favoring ‘shall’ over ‘will’ in a way that reads naturally but leaves a fingerprint. The example floating around, ‘Shall I compare thee to a summer’s afternoon,’ shows the flavor: slightly off-kilter phrasing that a human might not flag but a detector could.
What makes this notable is the timing. Watermarking has been a hot topic since ChatGPT exploded, but most approaches either degrade output quality or are trivially easy to strip. Anthropic’s bet is that by targeting low-information words, the degradation becomes negligible while the watermark survives light editing. The company apparently expects others to adopt similar techniques, which would be a meaningful shift from the current patchwork of unreliable AI-text detectors that mostly guess based on perplexity scores.
Skepticism is warranted, though. Any watermark that relies on word-choice statistics can be defeated by paraphrasing — run the text through another model and the pattern dissolves. That’s the fundamental tension: a watermark strong enough to survive rewriting would also be noticeable enough to annoy users. Anthropic hasn’t published specifics on robustness, and until they do, this remains a clever demo rather than a solved problem.
The bigger picture is that AI companies are feeling real pressure to make generated content identifiable. Regulators in the EU and California have floated disclosure mandates, and educators are desperate for tools to catch AI-written essays. If Anthropic’s approach gains traction, it could turn watermarking from an academic curiosity into a default feature — which would be a genuine inflection point for how we think about provenance in the age of generative text.
