Which AI Companies Use Watermarking? A Guide for Copywriters and Students
AI watermarking explained for copywriters and students: which companies use SynthID, what detectors can tell you, and what matters for SEO and assignments.
9 min readBranislav Mateas
Copy post as Markdown
On this page
If you write for clients or use AI while studying, you may want to know whether that help leaves a detectable mark.
Does ChatGPT watermark its text? Can a teacher check for it? Could Google treat an AI-assisted article differently in Search?
I want to understand what these tools reveal before I use their output in work that carries my name.
Google and Anthropic use SynthID-based text watermarking. OpenAI uses SynthID for supported images and audio, but its text method is textGrain.
Other companies use different marks or have not publicly confirmed SynthID use. These distinctions matter more than keeping a list of individual model versions.
A company’s policy also does not guarantee that every response carries a detectable watermark. The service, output type and deployment settings still matter.
For copywriters, useful content and the client’s requirements remain central. For students, the assignment’s AI rules determine what is permitted.
AI watermarking by company
This comparison covers the companies behind the tools discussed here. Provenance means information about how content was created.
An API lets another application use a provider’s AI service. Its marking settings can differ from those of the provider’s own chat app.
Company policies reviewed on 17 October 2026. Coverage varies by service and output type.
Gemini app and web text is marked. A Google representative also confirms API coverage.
Anthropic
A version of SynthID-Text.
Claude uses embedded text watermarks. Supported deployments include products, API and cloud partners worldwide.
OpenAI
Confirmed for supported images and audio.
Text uses textGrain, rather than SynthID-Text. Coverage is EU-first in ChatGPT; API customers can opt in worldwide.
Moonshot AI
No public confirmation found.
Kimi’s service terms describe visible AI labels and metadata identifiers. This does not establish a SynthID text watermark.
DeepSeek
No public confirmation found.
DeepSeek describes visible AI notices after generated text and in the interface. The checked sources do not establish SynthID or a statistical text watermark.
Z.ai
No public confirmation found.
Service terms address preserving AI identifiers and require users to label outputs. They do not establish a statistical text watermark.
xAI
No public confirmation found.
The Grok FAQ confirms watermarks on images and video. It does not establish SynthID or statistical text marking.
Cursor
No public confirmation found.
Marking for Cursor’s own models is unconfirmed. Cursor also offers external providers’ models; their policies and service routes matter.
Meta
Supported in a research toolkit; product deployment unconfirmed.
Text Seal provides text-watermarking tools, including SynthID support. This does not confirm text marking in Meta AI. Content Seal separately marks supported images.
“No public confirmation found” means the checked sources did not establish SynthID deployment in the company’s services. It does not mean the outputs are unmarked.
The source check included company announcements, social posts, service terms and research code. Research support and a deployed product feature are separate findings.
OpenAI illustrates the distinction. Its media provenance system uses SynthID, while textGrain handles text. Calling both systems “Google watermarking” would be inaccurate.
API opt-in is also different from marking by default. OpenAI’s API text watermarking is off by default and available for supported models.
For a specific workflow, check the provider’s current coverage documentation. This company overview describes the approach; it cannot certify an individual response.
SynthID, metadata and visible labels are different things
SynthID embeds a signal in generated content. For text, it changes statistical patterns in word choices during generation.
The watermark is not a visible label. It also does not require hidden characters that disappear when you paste text into a website.
C2PA Content Credentials serve a different purpose. They attach signed metadata describing a file’s origin or processing history.
A service can also display a visible AI label, use its own watermark, or combine several methods.
That is why the company table needs more detail than a yes-or-no column. A visible label or media watermark says little about statistical text marking.
Why regulation is accelerating adoption
The EU AI Act helps explain why providers are expanding provenance features. It does not require every provider to use Google’s technology.
Article 50 requires machine-readable marking and detectability for covered synthetic outputs. It also recognises technical limitations and exceptions, including qualifying standard assistive editing.
Provider-side marking and publisher-side disclosure are separate obligations. A watermark does not automatically mean every AI-assisted blog post needs a visible AI banner.
For text informing the public on matters of public interest, Article 50 includes an exception for human review or editorial control, with editorial responsibility.
The Commission’s guidance makes clear that this review must examine the substance. A spelling or grammar check alone does not qualify.
The Commission also provides optional EU icons for AI disclosure. These are visible labels, separate from embedded watermarks.
Using an icon does not establish compliance by itself. The relevant disclosure rules and editorial-review exceptions still apply.
These obligations can apply to providers outside the EU. A company’s headquarters do not settle whether its service needs provenance features.
China also has AI content identification rules. They require visible labels and metadata identifiers in covered services, and encourage digital watermarks.
Those rules do not prescribe Google SynthID. They also do not establish the marking of every Chinese provider’s output worldwide.
Should I use Chinese AI tools to avoid SEO penalties?
I would choose a tool for the quality of its output, rather than an assumed absence of watermarking.
Kimi and DeepSeek already document AI content identifiers. Z.ai’s terms also address identifiers and labelling. Choosing a Chinese provider does not guarantee unmarked content.
More fundamentally, Google does not automatically penalise content because AI helped create it.
Its generative-AI guidance recognises AI’s use in research and content structure. It asks publishers to focus on accuracy, quality, relevance and value for users.
The scaled content abuse policy targets large amounts of unoriginal, low-value content created primarily to manipulate rankings, regardless of how it was produced.
AI assistance and content abuse are different things.
A human can write filler. An AI tool can help create a useful article. I still have to verify the result and decide what deserves publication.
The checked Search policies do not establish an automatic SynthID or textGrain penalty. They also provide no basis for a Chinese-provider ranking advantage.
That conclusion comes from published policy. It is not a claim that we know every internal ranking signal.
Google’s investment in Gemini gives context, but it does not prove a ranking exemption. Its Search guidance is the evidence that matters.
I would assess any writing tool through factual reliability, useful output, privacy, data handling and contractual protection. Those factors affect the work I can responsibly publish.
For client work, agree on permitted AI use and any disclosure requirements before drafting. Verify claims and sources before delivery.
A page can satisfy Google’s content guidance while failing a client’s brief. Search policy does not decide what you promised your client.
What students should check before submitting work
Your assignment rules determine whether AI use is permitted. A watermark detector cannot replace those rules.
Check whether the assignment permits AI for planning, language editing or drafting. Follow its instructions for acknowledging any permitted use.
For example, Oxford gives students written notice of permitted AI use for each assessment. Check your own course’s requirements.
If you use AI to find sources, open and read the originals. Check that each source exists and supports the claim you cite.
Keep your notes, drafts and any required AI-use record. These can help explain how you developed your argument.
A watermark identifies a possible source of generated text. A plagiarism check compares text with other material. Neither determines whether your assignment permits AI assistance.
Google’s approach to AI-assisted web content also does not determine a school’s assessment policy. SEO and academic assessment have different requirements.
What a watermark can actually tell you
A watermark detector looks for an embedded signal. It differs from a classifier that estimates whether prose resembles AI writing.
Even with the correct detector, results have limits. Short or tightly constrained text provides fewer opportunities for a strong signal.
Editing and translation can weaken that signal. Copying the same wording into another document does not automatically remove it.
A detected watermark can indicate involvement by a supported AI system. It cannot establish factual accuracy, ownership or the amount of human contribution.
A negative result cannot prove human authorship either. Unsupported outputs, older generations and detection failures all limit that conclusion.
Detector access is a separate issue from marking. Google’s public SynthID Detector announcement covers supported images, video and audio.
That media service does not provide a public essay-text check. It should not be described as a tool that checks any AI-generated writing.
OpenAI’s text detector requires approved access. Its current access policy explicitly excludes checking student work for academic integrity.
Research on text provenance is a qualifying use case. That does not give teachers general access to check essays or theses.
Anthropic’s text detector remains in private preview. Educational organisations are among its eligible applicants; access is not an unrestricted public service.
A teacher or client who uses a generic AI detector is not necessarily checking SynthID. The result does not certify the presence of that watermark.
For SEO, the practical question remains simple: does the page give the reader something accurate, original and useful?
For an assignment, check the permitted help and acknowledgement requirements. For client work, check the brief and your agreement.
That is how I would choose an AI tool: understand its markings, check the work, and follow the rules that apply to the final text.
