AI watermarks explained

What is an AI Watermark?

A plain-English guide to the hidden signals AI companies now embed in generated text, images, audio and video, and who started doing it when.

Try the humanizer

AI watermarks in one paragraph

An AI watermark is a signal placed into AI-generated content at the moment it is created, so that software can later recognize the content as machine-made. In images, video and audio it is usually an imperceptible pattern hidden in the pixels or sound wave. In text it is a subtle statistical pattern in which words the model picks. You can't see it by reading, but a detector holding the provider's secret key can measure it. Watermarks normally identify the AI provider or model only. They do not contain your name, account or prompt.

Three ways AI content gets marked

👁️

Visible watermarks

Logos or labels you can see, like the moving Sora logo on OpenAI videos or the 'AI info' tags on Instagram and YouTube.

🔏

Invisible watermarks

Hidden signals inside the content itself, such as Google's SynthID or Claude's text watermark. They survive copy-paste and only a detector with the key can read them.

🏷️

Provenance metadata (C2PA)

Signed 'Content Credentials' attached to a file that record which tool made it. They are easy to read but disappear after a screenshot or re-upload.

AI watermarks vs AI detectors

They are often confused, but they work very differently. A watermark detector looks for a signal the AI provider deliberately planted, so it only works on that provider's content and requires its secret key. AI detectors like GPTZero, Turnitin or Originality.ai are classifiers that guess from writing style (perplexity, burstiness, word choice). They need no cooperation from the AI company, and they make more mistakes, including false positives on human writing. Text can carry a watermark and still pass a style-based detector, or be flagged by a detector and carry no watermark at all.

Timeline: when AI companies introduced watermarks

  1. July 21, 2023

    White House voluntary commitments

    Amazon, Anthropic, Google, Inflection, Meta, Microsoft and OpenAI pledged to develop provenance and watermarking for AI-generated audio and visual content. Text was not included.

  2. August 2023

    Google DeepMind launches SynthID for images

    Invisible watermarks were added to images generated by Google's Imagen models.

  3. October 2023

    Microsoft and Adobe roll out Content Credentials

    Bing Image Creator began attaching C2PA Content Credentials, and Adobe introduced its 'cr' Content Credentials icon.

  4. November 29, 2023

    Amazon Titan watermarks images by default

    Amazon's Titan Image Generator shipped with an invisible watermark on every image. Detection was added to Bedrock in April 2024.

  5. February 2024

    OpenAI adds C2PA to DALL-E 3; Meta labels AI images

    OpenAI began attaching C2PA metadata to DALL-E 3 images. Meta announced 'Imagined with AI' labels across Facebook, Instagram and Threads.

  6. May 2024

    SynthID expands to Gemini text and Veo video

    At Google I/O, Google extended SynthID to text generated in the Gemini app and to video.

  7. August 2024

    OpenAI shelves its text watermark

    OpenAI confirmed it had built a ChatGPT text watermarking method but chose not to ship it. It cited easy circumvention via translation or rewording and possible stigma for non-native English speakers.

  8. October 23, 2024

    SynthID Text is open-sourced

    Google DeepMind published SynthID Text in Nature and released it open-source through Hugging Face Transformers.

  9. May 2025

    Google announces SynthID Detector

    A portal for checking images, audio, video and text for SynthID. Google said more than 10 billion pieces of content had been watermarked.

  10. September 1, 2025

    China's AI labeling rules take effect

    China began requiring visible and hidden (metadata) labels on AI-generated content, including text, and prohibited removing them.

  11. May 2026

    OpenAI adds SynthID to ChatGPT images

    OpenAI began embedding Google's SynthID watermark, alongside C2PA metadata, in images generated through ChatGPT and its API.

  12. August 2, 2026

    EU AI Act transparency rules apply

    Article 50 requires generative AI providers to mark outputs in a machine-readable, detectable way. Systems already on the market have until December 2, 2026. California's AI Transparency Act (SB 942) also took effect the same day for images, video and audio.

  13. August 2026

    Anthropic starts watermarking Claude's text

    Anthropic announced that text generated by Claude carries an invisible statistical watermark, and that files Claude creates carry C2PA Content Credentials. Models launched from August 2, 2026 are marked from day one; older models are being covered by December 2, 2026.

AI Watermark FAQ

As of September 2026, ChatGPT text is not known to carry a statistical watermark, although OpenAI has said it plans to extend provenance signals to text to meet EU obligations. ChatGPT images do carry C2PA metadata and, since May 2026, a SynthID watermark.

Yes. Since August 2026 Anthropic embeds an invisible statistical watermark in Claude's text output across the API, Claude apps and Claude Code. Anthropic says the watermark contains no user or organization information.

Yes. Google has applied SynthID to Gemini app text since May 2024, and to images, audio and video through its other models.

No. Text watermarks are a statistical bias in word choice, not hidden characters. Reading the text, or searching it for special characters, won't reveal them. Only the provider's detector can measure them.

Not conclusively. Detection is probabilistic and needs a reasonable amount of text. Short or heavily edited passages may produce weak or no signal, and providers themselves describe a detected mark as a signal rather than proof.