What Anthropic’s new AI watermark means for tax practitioners
Anthropic’s introduction of a digital watermark for AI-generated content through Claude Sonnet 4.6 significantly impacts tax practitioners, enhancing the visibility of AI involvement in their work.
14 August 2026 · 4 min read
Part of our AI in accounting coverage. See the full AI for accountants guide →
Every piece of AI-assisted content your firm produces through Claude is about to become identifiable. On 12 August 2026, Anthropic announced that content generated through Claude Sonnet 4.6, via the Claude API, will carry a digital watermark. Imperceptible to human readers. Detectable by algorithms. Designed to be robust to some modifications of the text.
Three weeks earlier, on 22 July 2026, the Tax Practitioners Board released TPB(GS) 55/2026, its final guidance on AI use in tax agent services. The Code obligation was already clear. Practitioners remain accountable for anything AI touched, and firms need to document how professional judgement was applied. What changed on 12 August is that AI-assisted work will now be identifiable in the finished output, whether the firm has documented it properly or not.
What Anthropic actually announced
The watermarking system is called SynthID, jointly developed with Google DeepMind. The initial rollout covers content generated in the Claude API through Claude Sonnet 4.6. Anthropic has flagged this as the first step, with more products and services expected in the coming months.
A testing environment has also been released, letting practitioners check text, image and video content against the SynthID system. Anthropic has said it will monitor the rollout and adapt as needed.
The technical claim is that the watermark is designed to be robust to some modifications, including formatting changes and minor edits. The exact threshold at which modifications defeat the watermark has not been made public.
Processed does not mean authored
Practitioners are drawing a distinction between content that Claude processed at some point during creation and content that Claude authored in the sense of producing finished work with no human input. A document watermarked by Claude does not, on its own, tell a reader how much human judgement was applied.
A practitioner who runs a rough draft through Claude to tighten the phrasing, or uses it to translate technical wording into plain English for a client, produces a document that carries the watermark. A reviewer looking at the watermark alone cannot tell that from a document Claude produced unaided.
Anthropic has not confirmed how the watermark will be interpreted by third-party detection tools, and other AI providers have not yet indicated whether they will adopt equivalent watermarking. Both remain open.
TPB(GS) 55/2026 already set the standard
TPB(GS) 55/2026 already required practitioners to remain accountable for AI-assisted work and to document how professional judgement was applied. The watermark does not change the standard. It changes the visibility of AI use.
A practitioner who correctly applied professional judgement to AI-assisted advice remains a practitioner who correctly applied professional judgement. If the polished output carries a watermark and the firm cannot show the human process behind it, the firm is exposed.
AI-assisted work now needs a much clearer audit trail. Notes taken during the drafting process. Version history showing the human edits. Records of what was reviewed, changed and approved. The engagement letter clause covering AI use matters more, not less.
Two open issues to work through
Two additional issues are being raised in professional discussions.
Consistency across providers is the first. Anthropic is watermarking Claude output. Whether other providers will follow, and how quickly, is not yet known. If detection tools are built around SynthID before other providers catch up, Claude-using firms could appear to have used AI more heavily than firms using competitors, regardless of the underlying practice.
Removal is the second. Practitioners have raised concerns about what happens if the watermark is stripped from a document before it reaches the client. Anthropic has not published guidance on the compliance implications of removal. Until it does, the safer position is to leave the watermark in place and document the human oversight applied to the work.
What firms should be doing this week
Three practical steps.
• Review the AI clause in your engagement letters. If it is not there, add it. If it is there but does not name the specific tools your firm uses, update it. Clients need to know AI is part of the service, and consent needs to be documented.
• Set up a documented review process for AI-assisted work. Every piece of client-facing content that has passed through an AI tool needs a clear record of the human judgement applied to it. Chat logs, version history, review notes. All of it. The record needs to sit somewhere the firm can produce if a regulator or client asks.
• Talk to your team about the watermark. Staff need to understand that formatting an email through Claude will produce a watermarked output. The response is not to hide the AI involvement. The response is to keep clear records of the human oversight and to be transparent with clients about how the firm uses AI.
Well-run firms have nothing to fear
Firms that have been quietly using AI without documenting the use are running out of time to catch up. Firms that have documented properly, have engagement letter language in place, and are keeping records of the human review step will be in a much stronger position.
A well-run firm has nothing to fear from the watermark. Firms with no evidence behind their AI use should be moving now.
Frequently asked questions
What did Anthropic announce about watermarking Claude output?
On 12 August 2026 Anthropic said content generated through Claude Sonnet 4.6 via the Claude API will carry a digital watermark called SynthID, developed jointly with Google DeepMind. It is imperceptible to readers, detectable by algorithms, and designed to survive some edits. Anthropic described it as a first step, with more products expected to follow.
Does a watermark prove that AI wrote the advice?
No. A watermark shows Claude processed the content at some point, not how much human judgement went into it. A draft tightened for phrasing, or translated into plain English for a client, carries the same mark as a document Claude produced unaided, and a reviewer looking only at the watermark cannot tell the two apart.
Does the watermark change a tax practitioner’s obligations?
Not the standard. TPB(GS) 55/2026, the Tax Practitioners Board’s final guidance on AI use, released on 22 July 2026, already requires practitioners to stay accountable for AI-assisted work and to document how professional judgement was applied. What the watermark changes is visibility: AI involvement can now show in the finished output whether or not the firm kept records.
Should a firm strip the watermark before work goes to a client?
The article argues against it. Anthropic has not published guidance on the compliance implications of removing the watermark, so the safer position is to leave it in place and keep a record of the human oversight applied. The aim is to be open with clients about how the firm uses AI, not to hide it.
Do other AI providers watermark their output?
That was still unknown when the article was written. Other providers had not said whether they would adopt equivalent watermarking. If detection tools are built around SynthID first, firms using Claude could look like heavier AI users than firms on other tools, regardless of how carefully either applies professional judgement.
Sponsored
