Claude AI Watermark: Why Anthropic’s Invisible Marks Still Can’t Meet India’s Deepfake Law

Anthropic just did something no other major AI lab has managed to ship at full scale: it started hiding an invisible signal inside almost everything Claude writes. The Claude AI watermark, rolled out this week, is designed to survive copy-pasting, ride along through edits, and quietly answer one question — did AI write this? — long after the text has left the chat window.

It’s a genuinely hard technical problem that Anthropic appears to have cracked before OpenAI, Google, or Meta. But here in India, the more interesting question isn’t whether the watermark works. It’s whether it’s even the right kind of watermark for the law India already has on its books — a law that, as of February 2026, does not want AI marks to be invisible at all.

What Anthropic Actually Announced

According to Anthropic’s own Help Center, updated this week, every Claude model released on or after August 2, 2026 now weaves what the company calls “an imperceptible watermark directly into the text itself” whenever it generates a response. Anthropic says this doesn’t change the meaning, quality, or readability of the output — it’s a statistical pattern in word choice, not a visible tag — and it’s designed to travel with the text “when it’s copied and pasted elsewhere,” potentially surviving light editing.

For images — JPG, PNG, and SVG files Claude generates — the approach is different. Anthropic attaches C2PA provenance metadata, an industry-standard signed manifest (also used by Adobe, OpenAI, and Google) that records which model created a file and when, and flags if it’s been tampered with.

Both systems apply everywhere Claude is used — the chat app, the API, Claude Code, Claude Cowork, Claude Tag, and even when Claude is accessed through AWS, Google Cloud, or Microsoft Foundry — and Anthropic says the marking applies globally, not just to users in Europe.

The trigger is Europe, not India. Anthropic has signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content, a voluntary code that entered into force on August 2, 2026. That’s the real reason this exists: comply with Brussels, apply it worldwide because it’s cheaper than building region-locked versions.

The Catch Anthropic Itself Admits

To Anthropic’s credit, its Help Center is unusually candid about what the Claude AI watermark cannot do. By the company’s own account:

In short: a mark means “maybe processed by Claude.” No mark means nothing. That asymmetry is going to matter a lot once you look at what Indian law actually asks for.

Where This Collides With India’s Deepfake Law

Here’s the part almost no one covering this story internationally has flagged, because it requires knowing Indian law, not just Anthropic’s blog post.

On February 10, 2026, India’s Ministry of Electronics and Information Technology (MeitY) notified sweeping amendments to the IT (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 — brought into force from February 20, 2026. For the first time, these rules formally define and regulate “Synthetically Generated Information” (SGI): content that is artificially created or altered to appear real, covering deepfakes, voice clones, and AI-manipulated images and video.

The labelling standard India’s rules set is the opposite of “invisible.” Under the amended rules, visual synthetic content must carry a label that is prominent, easily noticeable, and adequately perceivable — not hidden in a corner, and not something only a detection tool can find. Synthetic audio needs a clearly prefixed spoken disclosure. An early draft had floated a rule requiring the label to cover 10% of the screen; that specific figure was dropped from the final notification, but the underlying requirement — a visible, human-perceivable label — stayed.

That’s structurally different from what Anthropic just built. C2PA metadata on a Claude-generated image is invisible to a viewer scrolling past it — you’d need a detector tool to find it, and Anthropic itself warns that re-saving, screenshotting, or converting the file format strips it out entirely. For platforms that count as a “Significant Social Media Intermediary” under India’s rules (roughly, more than 50 lakh registered Indian users), simply attaching invisible C2PA metadata to an AI image likely wouldn’t discharge their legal obligation to show users a visible tag.

There’s a second, subtler gap: India’s SGI framework, as currently written and analysed by law firms including Lexology, Freshfields, and S.S. Rana & Co., is built around audio, visual, and audio-visual content — deepfake videos, cloned voices, manipulated photos. Plain AI-written text, which is where Anthropic’s watermark is arguably strongest, doesn’t fit neatly into a definition built for things that “appear real” the way a fake video or voice clone does. A blog post written entirely by Claude may carry an invisible watermark under Anthropic’s new system while sitting in a regulatory grey zone under India’s current SGI rules — not banned, not clearly required to be labelled either.

Why This Matters for Indian Creators, Founders, and Publishers

If you run a media site, an ed-tech platform, or any product built on the Claude API in India, three things are worth acting on now, not after a compliance notice arrives:

This isn’t a hypothetical risk. Police in Ahmedabad recently arrested a man for using an AI-cloned identity linked to Aadhaar to commit fraud — exactly the kind of harm India’s visible-labelling rule is designed to catch.

The Bigger Picture: Invisible Marks Solve a Different Problem Than Visible Labels

It’s worth being precise about what Anthropic actually built, because the discourse around it — including a fair amount of user backlash on Hacker News and X over the past two days — has conflated two different goals.

Anthropic’s watermark is a forensic tool: something a platform, a school, or a journalist can run after the fact to check whether a specific piece of text likely passed through Claude. India’s IT Rules are chasing a consumer-facing goal: making sure an ordinary person scrolling their phone can tell, in the moment, that what they’re looking at was made by a machine — without needing a detector.

Those are both legitimate transparency goals. They’re just not the same goal, and a watermark built for one doesn’t automatically satisfy the other. Anthropic has, arguably, built the more technically impressive system. India’s law is asking for the more practically useful one for an ordinary voter or reader. The next real test — for Anthropic, for OpenAI (which has signed the same EU code but hasn’t shipped equivalent text watermarking yet), and for Indian regulators — is whether these two approaches get reconciled, or whether platforms end up needing both a hidden forensic mark and a loud visible label stapled on top of it.

For now, if you’re building in India, assume you need both.

FAQs