SynthID

In plain English

It is a specific implementation of watermarking rather than a detector. It does not examine finished text and guess; it marks output as the model produces it.

That means it can only ever speak about text produced by a model running SynthID. It has nothing to say about anything else, which is a much narrower claim than "AI detection" usually implies.

A worked example

Reading the documentation for what it says and for what it does not. Both quotes below are from DeepMind's own SynthID page.

  STATED — the mechanism
  "Large language models generate text one word (token) at a
   time. Each word is assigned a probability score, based on
   how likely it is to be generated next... SynthID adjusts
   these probability scores to generate a watermark."

  STATED — the user-facing property
  "not noticeable to the human eye, and doesn't affect the
   quality of the output"

  NOT STATED — whether the watermark survives paraphrasing
  NOT STATED — limitations of any kind

The mechanism is described precisely, and the absence is just as informative. On the page read for this entry, nothing addresses whether the signal survives editing, translation or a paraphrasing pass — the exact conditions under which every style-based detector is known to fail.

That is not an accusation of concealment. It is a reason to be careful about what a SynthID result licenses you to conclude, because the published material does not establish the thing most people would want to know.

The safe reading: a watermark found is meaningful evidence about origin. A watermark not found is not evidence of anything, and the documentation does not currently let you say more than that.

Why it matters for AI detection

It is the most prominent real deployment of the one detection approach that does not infer authorship from style, so it is the concrete version of a debate usually held in the abstract.

Its coverage boundary is the thing to understand. A watermark exists only where the model's operator planted one, which means text from models that do not participate — including open-weight models anyone can run locally — carries nothing to find. No amount of improvement to SynthID changes that.

For anyone facing an accusation, the practical consequence is that this technology does not help you and does not threaten you unless the specific model involved was watermarking. It is not the system producing the score on your report.

Commonly confused with

Watermarking

Watermarking is the technique; SynthID is one implementation of it by one company. Other schemes exist and are not interoperable with it.

C2PA

SynthID hides a signal inside the content itself. C2PA attaches signed provenance data alongside it. One survives copy-paste; the other survives editing history but can be stripped.

Read next

• Watermarking
• Token
• How AI detectors work