AI Watermarks Explained in Plain English
Impact on U.S. Lawyers
AI has helped us get past writer’s block, reduce procrastination, write better (if used properly), and generally, increase productivity. In fact, I used Claude to help me produce this summary in plain English. In most cases, our clients, readers, and the courts don’t mind at all … especially if we are transparent about it’s use. Some clients even demand the use of AI to reduce time/legal fees. The problem arises when we aren’t transparent or when a client demands that we do NOT use AI. Below is a quick summary about how watermarks work so you don’t “step in it.”
How it works (briefly)
There’s no visible tag or metadata saying “AI-generated.” Instead, generated text carries embedded watermarks. Claude nudges its low-stakes word choices (like “overcast” vs. “grey”) according to a hidden key, in a way invisible to a reader but statistically detectable by someone with the key. A detected mark shows only that content “may have been processed by Claude,” and the absence of a mark does not rule out that content was AI-generated, since edited or older-model text may carry nothing.
When it went into effect
This isn’t a future change. It’s already live. Anthropic began applying the watermarking on August 2, 2026, timed to the EU AI Act’s Article 50 transparency rules, which became enforceable that same day. Claude models launched on or after August 2, 2026 support marking at launch, and Anthropic is working to add marking support to models released before that date.
Is it US-only or EU-only?
Not EU-only. Anthropic says the markings apply worldwide, not only to users in Europe. So yes, it applies in the U.S. too, even though the U.S. has no equivalent watermarking mandate. The rule is driven entirely by EU law, but Anthropic chose to apply the same technical behavior globally rather than build a geo-fenced version. It covers all the ways people use Claude … the consumer app, the API/Claude Platform, Claude Code, Claude Cowork, and Claude Tag, as well as Claude accessed through AWS, Google Cloud, and Microsoft Foundry.
How do readers discover if something has been written by AI?
Right now, they can’t. Detection isn’t self-service yet. Anthropic hasn’t released a public detection tool, so today this is mostly a latent risk rather than something opposing counsel can check right now. That will change once the detection API ships.
Impact on U.S. lawyers
There’s no U.S. legal requirement for this (no American equivalent of Article 50 yet). The watermark shows up here purely as a side effect of Anthropic’s global rollout, not because U.S. law demands it.
Because it applies everywhere, U.S. firms using Claude are in the same position as EU firms: if a client, opposing counsel, or a court obtains a detection tool, AI involvement in a document could in principle be confirmed even where firms haven’t disclosed it.
The watermark travels with copy-pasted text and, per the article, seems to persist once embedded so old AI-touched boilerplate reused in a new contract could still carry a detectable signal.
It doesn’t affect authorship, ownership, or malpractice exposure directly. Anthropic states a watermark doesn’t determine legal responsibility for the output, but firms with strict “no AI without disclosure” client policies, or who bill on the assumption work was fully human-drafted, may want to revisit those practices given the marking is now persistent and hard to strip from raw text.
Is Open AI doing this too?
OpenAI has not publicly deployed a text watermark in ChatGPT as of August 2026.
It already watermarks other formats though. OpenAI attaches C2PA metadata and SynthID watermarks to images and SynthID to audio, and runs a public verification tool at openai.com/verify. Text is not yet marked. Interestingly, OpenAI uses Google’s SynthID scheme for images/audio rather than its own.
OpenAI has reportedly had the technical capability to watermark ChatGPT text for years but held off. Reporting attributed this to concerns about false positives and competitive risk.
But it’s committed on paper:
OpenAI, along with Anthropic, Google, Meta, Microsoft, and Mistral, signed the EU’s Code of Practice on Transparency of AI-Generated Content, which requires signatories to invisibly watermark AI-generated text under Article 50 of the EU AI Act.
OpenAI has said its goal is to “expand provenance signals to all modalities including text,” giving customers and developers “clear ways to meet their own transparency obligations as standards and tooling continue to” develop, but no firm launch date has been announced.
xAI (Grok) is notably the one major holdout that hasn’t signed the code at all.
What about Microsoft Copilot?
Copilot is actually further along than OpenAI’s own consumer product on this, though the coverage is uneven by content type:
Microsoft rolled out Copilot watermarking through native watermark controls and digital provenance signatures for AI-generated audio, video, and images across Microsoft 365 Copilot.
Microsoft adds invisible metadata to every file regardless of whether a visible watermark toggle is on, and major companies including Microsoft have committed to implementing these standards in their tools.
Text specifically: Copilot runs partly on OpenAI models and partly on Microsoft’s own, and I don’t see confirmation of an invisible text watermark (as opposed to image/video/audio marking and file metadata) shipped yet. It appears to be following the same “images/audio first, text still pending” pattern as OpenAI.
So, Anthropic moved first and fastest, with real watermarking live since August 2. OpenAI has signed the same commitment and updated its support pages to signal it’s coming, but as of today ChatGPT text itself still isn’t watermarked ... only its images and audio are. We’ll see what Grok does, and the big question remaining is what will AI legal software companies do? Many major legal AI tools (Harvey, Thomson Reuters CoCounsel, Claude for Legal itself) are built directly on Claude or on multiple foundation models, so Anthropic's watermark, since it's applied at the model level, would flow through into their outputs whether or not the legal tech vendor itself has said anything about it.
See also : https://www.artificiallawyer.com/2026/08/17/claudes-watermarks-and-their-legal-sector-impact/?jetpack_skip_subscription_popup.