AI Watermark Comparison: SynthID vs textGrain vs C2PA vs LLM Watermarking

As AI-generated images, videos and text multiply, the major companies have rolled out technologies for checking "whether this was made with AI." But names such as SynthID, textGrain and C2PA often come up in similar contexts, and it is genuinely hard to tell how they differ. This guide compares the main AI watermark and provenance technologies in terms of how they work, what they cover, who can check them, and their strengths and weaknesses. The information is based on each company's official announcements as of October 9, 2026.

Three approaches: watermarks, provenance and detection tools

Ways to identify AI-generated content fall broadly into three groups.

Digital watermarks: an invisible mark is embedded when the AI creates the content, for example in image pixels or in how words are chosen in text. Examples: SynthID, SynthID Text, textGrain
Provenance records (metadata): signed information about when, by whom and with which tool the content was created, and how it was edited, is attached to the file. Example: C2PA (Content Credentials)
After-the-fact detection tools: estimate how AI-like content is from its characteristics, even without a mark. These are the so-called AI detectors

Digital watermarks and provenance records are methods where the creating side adds a mark in advance. They allow well-grounded judgments, but cannot identify AI content that was not marked. Detection tools can be used on any content, but their results are only estimates, and misjudgments happen.

Comparing the main technologies

Here are the characteristics of the main technologies, item by item.

SynthID (Google)

Approach: digital watermark
Covers: images, video, audio and text
Where it is used: Google's AI, including Gemini, Imagen, Veo, Lyria and NotebookLM. OpenAI, NVIDIA, Kakao and ElevenLabs have also announced adoption
Who can check: anyone. You can check in the Gemini app, SynthID Detector (open to the public since October 2026, in English), Search and Chrome
Strengths: tends to survive even when attached information is removed. Content from multiple companies can be checked in one place
Weaknesses: it tells you little more than whether the content was made with a supporting AI. Not perfect against extreme edits

SynthID Text (Google)

Approach: digital watermark (LLM watermarking)
Covers: text generated by AI
Where it is used: released as open source, so anyone can build it into their own model with Hugging Face Transformers
Who can check: people who hold the watermark key (such as the developers who built it in)
Strengths: large-scale validation with Gemini showed almost no impact on quality. The impact on speed is also minimal
Weaknesses: weak against short factual answers, heavy rewriting and translation

textGrain (OpenAI)

Approach: digital watermark (LLM watermarking)
Covers: text generated in ChatGPT, Codex and the API
Where it is used: being introduced in ChatGPT and Codex in the EU within a few weeks. In the API, it can be turned on optionally worldwide (from October 5, 2026)
Who can check: only vetted researchers and specialist organizations
Strengths: in OpenAI's evaluation, it performs as well as or better than SynthID Text. The impact on quality is also small
Weaknesses: replacing 25% of the words in a 400-token text lowers the detection rate to 17%. Also weak with short text and text such as mathematics

Claude's text watermark (Anthropic)

Approach: digital watermark (LLM watermarking) and C2PA
Covers: text generated by Claude (models released on or after August 2, 2026 support it from launch, and earlier models are already covered on Anthropic's own services) and files generated by newer models
Where it is used: in every region where Claude is offered, including the API, Claude and Claude Code
Who can check: watermark detection is a private preview limited to eligible regulators, researchers and others under EU law. C2PA in files can be checked with the free Claude Content Checker
Notes: the text watermark is based on the method of Google's SynthID Text. Text translated by Claude is also watermarked

C2PA (Content Credentials)

Approach: provenance record (signed metadata)
Covers: images, video, audio, documents and more (version 2.4 of the specification, from April 2026, also supports HTML and Markdown)
Where it is used: OpenAI, Google (Pixel 10, Google Photos, YouTube), Microsoft, Adobe, and cameras from Leica and others. LinkedIn, TikTok and Meta use it for labels
Who can check: anyone. You can check with the Verify site or the labels in each service
Strengths: shows who created the content, with which tool and how it was edited. It is an industry standard with many participating companies
Weaknesses: the information is easily lost through screenshots or re-saving

Video Seal (Meta)

Approach: digital watermark
Covers: video
Where it is used: research released by Meta as open source
Who can check: developers and others who have built in the mechanism

When you want to check images, video or audio

When you want to check whether an image, video or audio file was made with AI, it is efficient to look in the following order.

SynthID Detector (synthid.com): checks whether content was made with AI from Google, OpenAI, NVIDIA or Kakao
Gemini app: checks whether content was made with Google's AI (other companies' AI is not currently supported)
C2PA Verify site (contentcredentials.org/verify): if a provenance record remains, it shows the tool used to create the content and its editing history
openai.com/verify: checks whether an image or audio file was made with OpenAI's AI

In every case, not finding a mark does not mean the content was not made with AI. It may have been made with an AI that is not supported, or the mark may have been removed by editing.

When you want to check text

For text, as of October 2026 there are almost no official watermark detection tools that the general public can use freely. Detection for OpenAI's textGrain and Claude's watermark is vetted or a limited preview, and SynthID Text is for developers who hold the key.

So for now, ordinary users and educators are largely limited to tools such as AI detectors for checking text. Detectors sometimes judge human-written text to be AI-generated, so it is important not to decide based on their results alone. How LLM watermarking works and its limitations are explained in "What is LLM watermarking?".

For developers and companies

For developers and companies offering their own AI, regulatory compliance is also an important consideration. The EU AI Act requires providers of generative AI to make their output detectable by machines as AI-generated.

If you generate text with your own model: you can build in SynthID Text as open source
If you use OpenAI's API: you can turn on textGrain
If you handle images or video: CAI provides open-source tools for attaching C2PA information

Watermarks and C2PA complement each other. Combining both, as OpenAI and Google do, is becoming the mainstream approach.

Limits common to every technology

Every technology shares certain limits.

No mark does not mean a person made it: content made with AI that is not supported, or with AI that does not add marks, cannot be identified
Vulnerable to processing and editing: heavy image edits, or rewriting or translating text, can remove or weaken the mark
They do not tell you whether content is accurate: a mark shows whether AI was involved or the content's provenance, not whether the content is correct
Some technologies can be checked only by a limited group: text watermarks in particular have detectors with limited availability

Summary

Technologies for identifying AI-generated content are advancing on two main tracks: digital watermarks that embed an invisible mark (SynthID, SynthID Text, textGrain and others), and C2PA, which records how content was created. For images, video and audio, anyone can now check using SynthID Detector or the C2PA Verify site, while text watermarks can still be detected only by a limited group.

No technology is all-powerful. Use watermark and provenance results as one piece of evidence, and judge together with checking the source. Each technology is explained in detail in "What is SynthID?", "What is SynthID Text?", "What is textGrain?" and "What is C2PA?".