Claude now leaves an invisible mark on anything it touches — including your RFI responses and claim narratives

Anthropic started embedding an invisible watermark into every piece of text Claude generates, to comply with a European transparency law. If your RFI responses, daily reports, submittal cover letters, or delay-claim narratives get drafted or cleaned up with Claude, that mark is now riding along in your project documents — and it proves a lot less than it sounds like it should.

What actually changed on August 11?

Anthropic confirmed it is watermarking Claude's text output worldwide, not just for users in the EU. The mechanism is a statistical bias in the model's word-choice pattern, applied at the moment the text is generated — invisible to a reader, doesn't alter meaning, and detectable later with the right tooling. It covers Claude.ai, the Claude API, and Claude Code, and Anthropic says image and file outputs get separate, standards-based C2PA provenance metadata instead. The trigger is Article 50(2) of the EU AI Act, which became enforceable August 2 and requires generative AI systems to mark their outputs in a machine-readable, detectable format or face fines up to €15 million or 3% of global revenue.

Where does this show up on a construction project?

Anywhere Claude touches project text that later becomes a document of record. That's a longer list than it first sounds like on a job that's already using AI for back-office work: RFI responses, meeting minutes, transmittal language, submittal cover letters, change order narratives, and — the one that matters most — delay-claim and REA (request for equitable adjustment) writeups, which routinely get drafted or polished with an LLM before they go out. Those documents are also exactly the kind of paper that ends up as evidence in an arbitration or a dispute over who caused a schedule slip. A watermark that survives copy-paste and travels into your document management system means that paper trail is now, in principle, forensically checkable for AI involvement — something it wasn't three months ago.

Does that mean opposing counsel can prove your claim narrative was AI-written?

No. Anthropic has been explicit that the watermark proves processing, not authorship. A superintendent who writes a daily report from scratch and only asks Claude to fix grammar can still trigger the mark. There's no way, from the watermark alone, to distinguish "Claude wrote this" from "a human wrote this and Claude touched it once." Treating a detected watermark as proof a claims consultant didn't do the underlying work would be a mistake — and treating its absence as proof a document is untouched by AI is an equally bad read, since ordinary editing can strip the mark out entirely.

How fragile is it, really?

Fragile enough that it shouldn't anchor a legal argument either way:

That last row matters for risk planning: this isn't a tamper-evident seal. It's a soft signal that can be defeated by anyone who wants to defeat it, and destroyed by accident through ordinary editing.

Is coverage consistent across the AI tools construction firms actually use?

No, and that inconsistency is its own problem if your firm standardizes on more than one assistant. As of mid-August 2026, Google's Gemini is watermarking text at scale, OpenAI has said it intends to add text watermarking under the same EU rule but hadn't shipped it yet, and xAI's Grok has signaled it may skip the commitment altogether. A firm using Claude for claims drafting and ChatGPT for RFI triage is watermarking one workflow and not the other, with no way to unify how those documents would later be evaluated.

What should a GC or claims team actually do with this?

• Don't rely on watermark detection as evidence, and don't assume the other side will either. It's not reliable enough to build a discovery strategy around, in either direction.
• Keep a separate, honest record of how a document was produced — who drafted it, what AI assistance was used, and what a human reviewed and signed off on — rather than depending on an invisible mark to reconstruct that later.
• Flag it for whoever owns your document retention and e-discovery process. If your firm hasn't asked "does our AI-assisted drafting policy account for detectable AI fingerprints in dispute documents," this is the prompt to ask it before it comes up in a claim.

Related: once a dispute is already underway, the more useful AI application is pulling the record together fast — a 1M-token context window now lets a claims analyst load an entire spec book, RFI log, and change order history into one session instead of hunting through folders.

Forward this to whoever manages document retention or claims prep at your firm — this is a policy question, not just a tech curiosity.

Friday one chart. Every week, one piece of data that should change a decision on your project. Subscribe at constructionaibrief.com.