🏭 Who watermarks what

Most major AI providers now do some form of content provenance. The methods are not the same. Some stack several layers. Some do almost nothing. This is a 2026 snapshot from public announcements, product behavior, and regulatory filings. The details move quickly.

Anthropic (Claude)

Anthropic has talked more than most about text watermarking. In February 2025 it published a technical report on an invisible text watermark that is meant to survive paraphrasing, and said it would test the method in a research preview. It also said it would disclose the marks, unlike some earlier proposals. As of 2026 the technique has not been broadly deployed inside Claude products.

• Text: the technology exists and is tested, but it is not on by default in consumer products.
• Images: Claude-generated images carry C2PA Content Credentials, so platforms and users can check AI origin.
• Detector: Anthropic has published guidance on detecting Claude content, but no public API for verifying the invisible text mark.

Google (Gemini / SynthID)

Google is the only major provider with a deployed, documented watermarking stack across every modality, under the SynthID brand (DeepMind).

• Text: SynthID-Text is applied to Gemini outputs. The detection method is described in a Nature paper (Dathathri et al., 2024) and research code is open-sourced. The production detection key is not public.
• Images / video / audio: SynthID embeds pixel and frequency-domain marks in Imagen, Veo, and audio generation.
• Detection: Vertex AI offers a SynthID detection API for enterprise customers.

OpenAI (ChatGPT / DALL·E)

OpenAI has said text watermarking is a strong idea with fragile deployment economics. It announced a text-watermarking tool in May 2025, initially disclosed, with a slow, low-key rollout that has continued through 2025 and into 2026.

• Text: a statistical watermarking tool exists and is being phased in cautiously. OpenAI has stressed the risk of false positives and evasion.
• Images: DALL·E and GPT-4o image outputs carry C2PA Content Credentials. A provenance classifier for images was released in 2024.
• Detector: no public text-watermark detector. The image provenance classifier is available via API.

Meta

Meta labels AI-generated images on Facebook and Instagram with both visible and invisible markers (2023), and open-sourced AudioSeal (2024) for watermarking AI speech. Its open-weight Llama models ship without mandatory watermarking. Output provenance is left to the deployer, so the Llama ecosystem is unmarked by default.

xAI, Mistral, and the rest

xAI's Grok has published little detail on provenance. Images from its image generator carry platform-level indicators on some surfaces, but there is no widely documented C2PA or invisible-mark scheme. European open-weight providers (Mistral, Aleph Alpha) generally treat watermarking as an opt-in, deployer-side concern.

The open-source ecosystem

Open-weight models ship no watermark by default, but the tooling exists. Kirchenbauer-style red/green-list watermarking has reference implementations. Google open-sourced SynthID-Text research code. Third-party frameworks can add a mark at inference time. The same limit shows up everywhere: without a public detection key, anyone can strip or ignore the mark and nobody can prove it.

What is forcing the pace: regulation

• EU AI Act Article 50 requires machine-readable marking of AI-generated content, with obligations on both providers and deployers.
• California SB 942 (California AI Transparency Act) requires disclosure and provenance mechanisms for generative AI services offered in California.
• China's deep-synthesis rules are among the earliest mandates for labeling AI-generated content.
• C2PA is the industry standard body. Steering members include Adobe, Microsoft, Intel, Google DeepMind, OpenAI, and Anthropic.

At a glance

If you handle AI content, metadata marks (C2PA, XMP) are the common case and are straightforward to inspect and strip. Statistical text marks exist, but only Google's is broadly deployed. Pixel-level marks mainly matter for image and video work. Which layer you are looking at decides whether removal is easy, expensive, or not possible.