Is DeepSeek Watermarked?

DeepSeek's Open Platform Terms of Service spell out a labeling obligation for anyone building on its API. Chinese law classifies those developers as "network information service providers," and that status brings a duty to label AI-generated content before it reaches end users.

Note where that duty sits. DeepSeek's terms place it on the developer integrating the API, not only on DeepSeek itself.

The rule behind it is specific. China's Cyberspace Administration finalized the Measures for Labeling of AI-Generated Content, alongside mandatory technical standard GB 45438-2025, on March 14, 2025, both effective September 1, 2025.

On DeepSeek's own consumer product, the company's model and algorithm disclosure page describes what that looks like in practice:

• A reminder shown on the welcome page when a chat starts
• A label appended to the end of generated text

Both are visible to whoever reads the output, so the detector Google's autocomplete nudges people toward has nothing to look for. The disclosure page describes nothing sitting underneath that tag, just the plain-text label itself.

DeepSeek issued its own announcement around September 1, 2025 confirming the rollout. It isn't findable self-hosted on a deepseek.com or deepseek.cn domain, but it was reproduced verbatim across multiple independent outlets, including Sina, 快科技 (mydrivers), and cctime.

China's rule doesn't stop at a visible tag. GB 45438-2025 requires explicit labeling, meaning on-screen text, audio, or graphics a person can see or hear, and implicit labeling, meaning machine-readable metadata embedded in file headers.

The implicit half has named fields:

• The provider's name or code
• A content reference number
• In the regulation's own words, "watermarks where feasible"

It applies across text, images, audio, video, and virtual scenes. It also binds both the platforms that host content and the creators who generate it.

DeepSeek's disclosure page, the one document that actually describes chat.deepseek.com's behavior, accounts for the explicit half and stops. It says nothing about embedding provider codes, reference numbers, or a watermark into file metadata.

So there's a real open question here that DeepSeek hasn't answered publicly. Does the company consider implicit metadata labeling infeasible for its chat product, or has it built that layer without documenting it?

Both are plausible. Neither has been confirmed.

Nothing in DeepSeek's own documentation describes an invisible watermark. No statistical signal embedded in token choices, no hidden pattern a detector could pull back out of the text later.

A report from phandroid.com published around September 3, 2025 complicates that slightly. It describes DeepSeek's approach as combining the visible label with "implicit" hidden digital fingerprints baked into the content itself.

That claim doesn't appear in DeepSeek's Terms of Service or disclosure page. It reads much more like a paraphrase of the Chinese standard's general implicit-labeling requirement than a DeepSeek-specific technical disclosure.

A useful contrast sits one company over. Alibaba's Tongyi Qianwen, the chat product built on the Qwen model family, was reported by Chinese tech outlets in 2023 to use "Orange Shield" technology for screenshot-resistant invisible watermarking, tied to an early draft of the same national labeling push.

That's what makes "DeepSeek has an invisible watermark" look less like a documented feature and more like a borrowed description of what the regulation generally asks of everyone.

Two related gaps stay open. There's no confirmation of any watermark, visible or invisible, in output from DeepSeek's Janus-Pro image generator, and no DeepSeek statement addresses EU AI Act Article 50 compliance at all.