Since August 2026, new Claude models have marked their text output at the model level; Google has watermarked Gemini text since 2024. What changes the picture now is that Anthropic has announced plans to release a watermark detection API.
Once one major provider makes detection self-service, the others are unlikely to hold out for long — checking whether a document passed through a model may soon stop being something only the provider can do, and become something a client, a court, or opposing counsel can run themselves.
Fee review is about to get an AI dimension
Clients will soon review law firm work and invoices for whether the work was worth the hours billed. A watermark detector adds a second layer to that review: whether — and how much — of the delivered work was AI-generated.
Courts or opponents sit in a similar position, only with a different concern. A judge or the opponent who runs a submission through Anthropic’s future detector and gets a high AI signal invites exactly the kind of scrutiny most practitioners would rather avoid, particularly where a court is already wary of AI-assisted filings. A detection result becomes a data point that is easy to obtain and hard to explain away in a sentence.
What editing does, and why “just rewrite it” is bad advice
Once a watermark can be checked, the natural instinct is to try to edit it away. That mostly does not work, and understanding why requires knowing how a watermark gets into the text in the first place.
Fixing typos leaves the mark almost fully in place, tightening sentences only dilutes it, and even a heavy edit still leaves a surprising amount standing. What actually erases marks is recomposing the whole document - without ai.
Very short passages, quotes, or text with only one sensible wording — a statute number, a party name, a list of figures — don’t offer enough to leave a reliable mark in the first place, regardless of editing.
Drafting with AI and without watermarks is already possible
Watermarks attach to model-generated prose. Lawlift Drafting Features are built around your text and your rules. That is an important difference.
Deterministic (rule-based) templates are the clearest case. Every sentence in the output came from your template. So there is no AI-generated text, no watermarks inserted.
That also applies to AI-supported features like Drafting with Guardrail templates, Re-Drafting, Reviewing, Anonymizing, working with the Clause Library, and Formatting.
For readers who want to dig deeper into the mechanics: Anthropic’s own account of how Claude’s text watermarking works, the research it builds on, Google DeepMind’s SynthID for text, and the academic paper that started this line of work, Kirchenbauer et al., ICML 2023.