|
How the watermark works
Anthropic says future Claude models will use a version of Google DeepMind’s SynthID-Text method. A language model usually has several reasonable words it could choose next. The watermark changes the source of randomness used to make those choices, creating a pattern that a detector can measure across a long passage.
No hidden characters or extra tokens are added. The signal carries no person, organization or chat identifier, and Anthropic says it has negligible effects on speed, quality and cost. Translated output will be watermarked. Claude-created images and files will use signed C2PA metadata, a separate provenance system attached to the file.
What a detector can prove
The detector will estimate the likelihood that some text came from Claude. It cannot identify the user, establish ownership, certify human authorship or rule out another AI system. Short samples are harder to detect. Code, proofreading and factual passages also leave a weaker signal because the model has fewer equally good word choices.
Anthropic expects the pattern to survive light editing. A complete rewrite can remove it. That limitation matters because watermark results should be treated as one piece of evidence, especially in schools, hiring, publishing and legal work. Anthropic plans to release a detector API.
What could change next
Claude’s rollout is tied to the European Union’s AI transparency rules, and other major providers have signed the same industry code. If model makers adopt compatible methods, publishers and platforms could check provenance automatically. Organizations may also add watermark checks to existing review systems for student work, marketing claims, public records and regulated documents.
The larger risk is false certainty. A statistical signal can support a review; it cannot settle who supplied the ideas, checked the facts or made the consequential decisions. Strong policies will need to define acceptable assistance and require process evidence when authorship matters.
How to keep AI-assisted work human
Start with your material. Give the model interview notes, customer language, examples, constraints and the decision the reader needs to make. Thin inputs usually produce generic output.
Keep the judgment calls. Use AI to find gaps, compare structures, challenge an argument and tighten a draft. Choose the claim, order, examples and conclusion yourself.
Run a specificity pass. Replace broad claims with names, numbers, dates, constraints, firsthand observations and concrete consequences.
Break the machine rhythm. Delete canned contrast formulas, repeated three-item lists, empty scene-setting and polished sentences that never say anything definite. Read the draft aloud and vary the pace where it feels mechanical.
Leave a responsible trail. Verify every factual claim, keep source notes and preserve meaningful revisions. Follow disclosure rules for school, client, legal, financial and public-facing work.
A five-step working pass
1. Write five points the reader should remember and gather the evidence for each one.
2. Ask Claude for three possible structures, missing questions and weak claims.
3. Choose the structure and write the opening, examples and point of view in your own words.
4. Let Claude flag repetition, unclear sentences and unsupported leaps, then make the edits yourself.
Watermarking and robotic voice are separate issues. The detector reads probability patterns. Readers notice generic ideas and predictable cadence. Your best defense against forgettable work is real material, clear judgment and a final human edit.
| 
