AI Watermarks: SynthID, textGrain, and C2PA—what are the differences?

An AI watermark is a signal embedded in content to allow a compatible tool to verify its artificial origin. SynthID and textGrain embed an invisible pattern in generated media, while C2PA attaches a signed history to the file. None of the three proves that content is authentic, belongs to someone, or was entirely produced by a artificial intelligence.

What are the differences between SynthID, textGrain, and C2PA?

SynthID and textGrain are AI watermarking mechanisms embedded during generation. Their detectors then look for a specific statistical signal. C2PA, by contrast, is an open provenance standard: a signed manifest describes a file’s origin and modifications. C2PA is neither a watermark nor a universal artificial-intelligence detector.

The difference may seem technical, but it changes how evidence is handled. An invisible watermark remains attached to the content’s characteristics. A C2PA manifest relies on structured data linked to the file and protected by a cryptographic signature—that is, a mechanism that makes it possible to detect an unauthorized modification.

These approaches can coexist in the same file. In 2026, OpenAI uses textGrain for certain European texts, SynthID for supported images and audio files, and C2PA-based Content Credentials for certain images. Therefore, one mechanism should not systematically be chosen over another.

How does the SynthID AI watermark work?

Google launched SynthID in 2023 to insert invisible labeling directly into generated content. In 2024, SynthID covered text, images, video, and audio, with availability depending on the generator and partner. The signal is designed to remain detectable after certain transformations, unlike simple metadata.

For text, SynthID-Text modifies the selection of tokens, that is, the words or word fragments chosen successively by the model. The detector then looks for the statistical pattern produced with a key. The scientific publication released in Nature in 2024 specifies that it is not a classifier capable of recognizing all text produced by all models.

Since October 7, 2026, Google has offered a public SynthID detector worldwide in English for images, video, and audio. Google said it currently supports its own media and that of partners such as OpenAI, NVIDIA, and Kakao. Public text verification was not included in that announcement.

Google also stated on October 7, 2026, that it had marked more than 180 billion images and videos, as well as the equivalent of 240,000 years of audio. These figures come from Google and describe deployment volume, not the detector’s success rate on transformed content.

Can textGrain recognize all text written by AI?

textGrain cannot recognize all text written by artificial intelligence. Announced by OpenAI on October 5, 2026, this AI watermark modifies the statistical pattern of words or tokens during generation. The detector checks only this specific signal using OpenAI’s secret key.

textGrain adds neither hidden characters, nor special spaces, nor artificial punctuation. Its verification does not require the original prompt or the probabilities calculated by the model. It does, however, require the text to come from a compatible system on which marking was enabled.

The results published by OpenAI in October 2026 provide a useful order of magnitude. For psychology passages, the company reported approximately 80% detection at 200 tokens and 95% at 400 tokens, with a target false-positive rate of 1%. These measurements had not yet undergone independent comparative replication shortly after launch.

Resistance to rewriting remains the weak point. In OpenAI’s English-language evaluation on 400 tokens, replacing 10% of the words with synonyms reduced detection from approximately 92% to 66%. With 25% of the words replaced, it fell to 17%.

Visit performances also vary by language. Before adjustments specific to each language, OpenAI reported in 2026 rates ranging from 42.2% for Romanian to 69.0% for Spanish among the 24 official languages of the European Union, at a 1% false-positive threshold. Honestly, a negative result should never be used alone to accuse or exonerate an author.

Does a C2PA signature prove that content is authentic?

A valid C2PA signature proves that the signed provenance information has remained intact and is attributable to a signer under the applicable trust model. It does not prove that the image, video, or represented claim is accurate, honest, or free from manipulation.

The C2PA standard associates a manifest with the file, including assertions about its creation and editing history. A cryptographic signature links this information to the digital asset. If someone makes an unauthorized modification to a retained manifest, validation can flag the alteration.

The little-known pitfall is the complete disappearance of provenance. A screenshot, a service that removes metadata, or a file conversion can remove the C2PA manifest. Its absence therefore does not demonstrate fraud. It simply means that the provenance chain is no longer available in the file being examined.

This issue directly affects publishing workflows. An automatic conversion to WebP or AVIF to speed up a website, a resizing tool, or a social network may modify the delivered file. From an agency’s perspective, the standard practice is to test whether provenance information is preserved throughout the chain, from export to public download.

What value should be assigned to a detection result?

A positive AI watermark result indicates in 2026 that a compatible system probably generated or processed at least part of the content. This result does not identify the user, measure human contribution, determine intellectual-property ownership, or by itself assign any legal responsibility.

A negative result is even less conclusive. The content may come from a model without a watermark, may have been generated before the mechanism was activated, or may have undergone translation, paraphrasing, or substantial editing. C2PA, for its part, may have been removed during an otherwise routine technical process.

A sound procedure should therefore bring together several elements:

• retain the original file rather than a screenshot or recompressed copy;
• identify the generator, creation date, and provenance features available on that date;
• verify the embedded watermark and the C2PA manifest separately when they coexist;
• record the detection tool, its access, its threshold, and the date of the check;
• have the human and contractual context reviewed before taking any disciplinary or legal decision.

In the projects we carry out, we often see provenance added at the end of production. That is too late. If images from an image generator integrated into a platform or sequences created with generative video tools are used on a website, the originals and generation logs must be preserved from the outset.

What does the European AI Act change for businesses in 2026?

Article 50 of the AI Act requires providers of systems generating synthetic text, images, video, or audio, as of August 2, 2026, to make the orputs machine-readable and detectable as artificial or manipulated, insofar as this remains technically feasible.

The European Commission provides for a limited transition period until December 2, 2026, for certain systems previously placed on the market. This obligation explains the European rollout of textGrain and the expansion of provenance tools, but it does not transform every detection into automatic judicial evidence.

For an executive using artificial intelligence, the right approach is to organize traceability rather than purchase a hypothetical universal detector. Preserve source files, published versions, usage rights, human approvals, and the names of the tools used. An AI usage policy understood by the teams reduces risk more effectively than an isolated review after publication.

Establishing provenance from the design stage prevents the loss of subtle signals during conversions, optimizations, or distribution. An outside perspective can alor help test the technical chain and distinguish a regulatory requirement from a mere marketing promise.

FAQ on watermarks and the provenance of AI content

Can an invisible watermark be removed from an AI-generated image?

An invisible watermark can be weakened by certain transformations, but SynthID is designed to withstand some of them. Resistance depends on the generator, the medium, and the modification applied; the absence of detection does not prove human origin.

Does a screenshot preserve Content Credentials?

A screenshot can completely remove the C2PA manifest linked to the original file. The visual content remains, but its signed provenance chain generally no longer accompanies the new image.

Is the textGrain detector accessible to everyone?

The textGrain detector was limited in October 2026 to researchers and approved expert organizations by OpenAI. The company justified this restriction by citing the risks of false positives and false negatives.

Does a translated text retain its statistical watermark?

Translation weakens statistical text watermarks such as textGrain and SynthID-Text because it replaces many of the original token choices. Therefore, a failure to detect the watermark after translation does not allow one to conclude that the text was written by a person.