AI detection and AI watermarking are not the same thing
An AI detector such as Pangram, GPTZero or Turnitin makes a statistical guess about authorship by measuring how machine-like your prose reads. It has no access to any signal the model left behind. A watermark is the opposite: a pattern embedded deliberately by the generator as the text is written, verified with a key rather than inferred from style.
That difference decides how each one ages. Detection depends on humans and models writing differently, and that gap is closing as people absorb the vocabulary and cadence of the models they read, so detector false positives rise over time. Watermarking never depended on a style difference, so the same convergence does not weaken it at all.
The two therefore fail in opposite directions. A detector can flag writing no model ever touched. A watermark cannot appear in text a model did not generate, but it also cannot show who wrote the argument, and no third party currently holds a key to read Anthropic's.
Frequently asked questions
What is the difference between an AI detector and an AI watermark?
A detector infers authorship from writing style after the fact, with no access to anything the model left behind. A watermark is a signal deliberately embedded during generation and verified with a key. One is a guess; the other is a lookup.
Are AI detectors becoming less accurate over time?
Structurally, yes. Detection relies on human and machine writing being distinguishable, and that distance is shrinking as people adopt model phrasing. Pangram benchmarks its false positive rate on pre-2022 documents because later human writing can no longer be assumed uncontaminated, and states it is not known how long that will remain sufficient.
If my writing sounds like AI, will it have a watermark?
No. A watermark is not a style measurement. It is a pattern in token selection that only the generating model can introduce, so writing that merely reads like AI cannot acquire one. This is the clearest practical difference between the two.
Can an AI detector find Claude's watermark?
No. Reading Anthropic's watermark requires their private key, which no third party holds. Any tool reporting a percentage is running style analysis, whatever it calls itself.
Which is stronger evidence, a detector score or a watermark?
Neither settles authorship. A watermark is the stronger indicator that a model was involved, but says nothing about who wrote the argument or whether the assistance was allowed. A detector score is an inference, and its own vendors' terms generally forbid using it as the sole basis for action against a person.