The watermark is statistical, not stamped into the text
Invisible watermarking for generated text works by steering next-token sampling with a secret key and the preceding words. The output still reads normally. There are no hidden characters, no extra tokens, and nothing a user can spot by eye.
That also means detection depends on statistical evidence. A detector checks whether the sequence of words lines up with the keyed pattern well enough to call it watermarked. If the output is too short, there may not be enough signal to separate a real watermark from ordinary randomness.
Why short outputs leave too little signal
A few lines of text give the system very little to work with. A short answer can look clean, plausible, and still produce too few keyed token choices for reliable detection. The watermark may be present in principle, but there is not much to measure.
That is a real weakness for chat-style use. One-line replies, brief summaries, and short completions do not accumulate much evidence. Detection gets shakier fast once the text stops giving the model room to make varied token choices.
Where exact wording rules force the watermark out
The watermark only has room to operate where the model has options. If the output needs to be exact, the mark gets dropped or weakened. That includes factual completions that would become wrong if a different term were chosen, and code where changing a token could break the result.
That boundary matters more than it sounds. A system can watermark ordinary prose quite well and still leave holes in the very outputs people often want to check. Claude’s watermarking plan follows that line by avoiding cases where exact wording is required.
Detection gets weaker once entropy drops
Watermarking performs better when the model has more than one sensible token to pick from. In low-entropy output, the sampling space narrows and the keyed bias has less room to move. SynthID-Text uses a form of tournament sampling, and the underlying point is blunt: more entropy gives detection more to hold on to.
Low-entropy text is common in technical content. Code comments, factual answers, boilerplate explanations, and tightly constrained completions all reduce the number of real choices. The watermark can still exist in some of those regions, but it is less dependable.
Code, factual answers, and other low-choice outputs
Code is awkward for watermarking because correctness is fragile. If a token must stay fixed for the snippet to run, the watermark has nowhere safe to nudge the output. The same issue appears in factual answers where wording is constrained by accuracy.
That leaves provenance checks with uneven coverage. Some parts of a response may carry a detectable pattern, while the exact bit people care about does not. In arbitrary code regions such as comments, watermarking can still appear, but that is not the same as having a strong signal across the whole file.
Why longer passages are easier to flag
Longer text gives the detector more evidence. Each additional token is another chance for the keyed pattern to show up, and the statistical gap becomes easier to measure. The watermark does not get stronger in a magical sense, it just gets easier to spot.
That is why long-form generated content is the easiest place to use it. A full article, extended answer, or sustained summary offers enough sampling events for the detector to build confidence. Short replies rarely do.
Practical limits for provenance checks in the real world
Content provenance checks live or die on scope. Watermarking can help flag some generated text, but it is not a universal marker for every output shape. If the text is short, exact, low entropy, or heavily constrained, the detector has less to work with.
There is also the question of coverage. Anthropic has said it is applying watermarking globally at launch because it does not yet have a durable way to scope it by region. That matters for compliance and for anyone trying to use the watermark as a neat jurisdictional switch. The system is being added across Claude-generated text, including older models over time, but that does not remove the technical limits at the output level.
A watermarking scheme can support provenance checks, yet it still sits inside the same messy reality as the text model itself. Some generated text will be easy to flag. Some will not. The awkward bits are usually the ones that matter most.
