Claude's invisible watermarks meet the redline: what SynthID-Text means for legal drafting

What it is

Claude's text watermarking is an integration of Google DeepMind's SynthID-Text into Claude's output pipeline. It biases token selection in a statistically detectable pattern so that a downstream detector can tell, with some confidence, that a passage was produced by Claude — without changing what the text says to a human reader. According to the piece that surfaced this for the legal sector, the watermark is invisible to reviewers but survives light editing, and its arrival has practical consequences for firms drafting with Claude (Artificial Lawyer coverage).

The next-work-session test

Concrete scenario: a senior associate asks Claude to draft an indemnity clause and a limitation-of-liability schedule for a vendor MSA, pastes the output into the negotiated draft, redlines against the counterparty's markup, and sends it to the client for sign-off.

What changes: the clauses the associate pasted in likely carry a statistical watermark. If counterparty counsel — or a regulator, or opposing counsel in a later dispute — runs a detector, portions of the "firm work product" become identifiable as machine-drafted. That does not change the enforceability of the contract, but it changes three things in the associate's next session: (1) what the engagement letter says about AI use, (2) whether the firm's AI-use disclosure to the client matches what's on the page, and (3) whether privilege and work-product characterizations still hold up when the drafting provenance is externally verifiable.

The immediate action is small: decide before you draft whether watermarked output is acceptable for this matter, and note it in the file.

Pricing

Pricing: unverified. Anthropic has not, to my knowledge, published a separate SKU for watermarking — historically SynthID-Text has been described by DeepMind as a technique layered onto model outputs rather than a paid feature. Whether watermarking is on by default across all Claude tiers (Free, Pro, Team, Enterprise, API), whether Enterprise customers can disable it, and whether detector access is gated or paid are the questions a GC should send to their Anthropic account manager in writing. I was not able to verify current pricing or the toggle behavior via search in this session; treat any specific dollar figures you see elsewhere as claims until confirmed on Anthropic's own pricing page.

What we'd actually use it for

Honestly, lawyers don't "use" a watermark — it uses them. The narrower, realistic use is defensive:

• Internal audit trail. If your firm wants to prove which passages in a brief or memo came from Claude (for training-data review, malpractice defense, or a client audit), a detector run against your own archive is more reliable than trusting attorneys to log prompts.
• Counterparty due diligence. Running suspicious counterparty submissions through a detector to flag AI-drafted sections you should scrutinize harder — particularly boilerplate reps, warranties, and indemnities where a hallucinated cross-reference is easy to miss.
• Ghostwriting checks on expert reports and declarations, where authorship representations matter.

Limits

• It's a probabilistic signal, not proof. SynthID-Text returns a confidence score. Short passages, heavy edits, translation, and paraphrase all degrade detection. A "not detected" result does not mean "not AI-written."
• Only Claude's outputs. The detector doesn't identify GPT-, Gemini-, or Llama-generated text unless those vendors ship compatible watermarks. Mixed-model workflows produce mixed signals.
• No privilege protection. A watermark is a provenance marker, not a confidentiality control. It doesn't prevent prompts or outputs from leaving your tenant, and it doesn't answer whether feeding a privileged document to Claude waived privilege — that's still your engagement terms and your Anthropic contract.
• Detector access is the real question. If detectors are widely available, so is adverse discovery. If they're gated, your firm can't self-audit. Neither state is fully comfortable.
• Still manual: classification of which matters permit Claude at all, engagement-letter language, client disclosures, and CLE-competence documentation.

Try it if

• You run a firm-wide AI-use policy and want a technical basis for the "we can detect it" claim.
• You handle matters (M&A diligence, expert reports, regulatory submissions) where AI authorship disclosure is contested.
• You already log Claude usage and want a second, output-side check against your prompt logs.
• You're advising clients on AI governance and need a working example of provenance tooling to point to.

Skip it if

• You expect the watermark to answer privilege, work-product, or unauthorized-practice questions. It doesn't.
• Your workflow relies on Claude output being indistinguishable from attorney drafting for reasons your engagement letter doesn't cover — fix the letter first, not the tool.
• You need cross-vendor detection today. One-vendor watermarks won't give you that.
• You're hoping to detect counterparty AI use reliably at the paragraph level after human editing. Current watermarking degrades under exactly the kind of light rewrite a junior associate would do.

The honest read: this is a governance and disclosure tool, not a drafting tool. The next-session change for legal isn't a new button in Claude — it's a paragraph in your engagement letter and a line in your matter-intake checklist. Do those before your next Claude-assisted draft, not after. See the Artificial Lawyer piece for the sector framing that prompted this write-up.

---

This content is for informational purposes only and is not legal advice. Confirm confidentiality, privilege, and jurisdictional rules before using any AI tool with client matters.