Images generated by ChatGPT, Gemini, Midjourney, or Adobe Firefly can contain provenance markers: C2PA, EXIF, XMP, and sometimes an invisible signal embedded in the pixels. Since August 2, 2026, Article 50 of the EU AI Act has made machine-readable marking a standard requirement for major providers.

This guide explains what those files contain, how an AI watermark detector finds the signals, what an AI watermark remover can remove, and where the legal boundaries sit.

What an AI-generated file contains

Two types of markers can coexist:

  • C2PA metadata / Content Credentials: a signed manifest stored in a JUMBF box inside JPEG, PNG, WebP, MP4, and other files. It records the producer, date, and actions such as c2pa.created.
  • A latent watermark: Google SynthID, also used by OpenAI, modulates a signal into pixels or audio. It is invisible to the eye and designed to survive crops, compression, and light re-encoding.

In 2025–2026, OpenAI began including both on images generated with ChatGPT, Codex, and the API (source: help.openai.com, accessed August 2026). Google uses the same layered approach across Gemini, Imagen, and Veo.

Why metadata disappears on its own

The C2PA specification states its limitation directly: provenance metadata can be removed (C2PA explainer 7.2.3). Social platforms re-encode uploads and often strip metadata, a behavior documented since 2025.

Your file can therefore lose its C2PA manifest while retaining a latent signal embedded in its pixels.

How to detect AI watermarks

An AI watermark detector reads the file structure directly:

  1. JUMBF boxes and jumb chunks: the C2PA manifest signature in JPEG APP1 segments, PNG chunks, WebP RIFF chunks, and MP4 boxes.
  2. EXIF and XMP: Software, Description, digitalSourceType, trainedAlgorithmicMedia, and ai-generated fields.
  3. Open-source pipeline fields: Stable Diffusion and ComfyUI parameters blocks containing prompts, samplers, and model names.
  4. Generator signatures: OpenAI, Google, Meta, Adobe, Stability, Midjourney, and others in those fields.

OutWatermark AI also decodes the CBOR manifest so you can inspect the producer, date, and actions before removing anything.

How to remove markers from your own content

1. Remove C2PA metadata without quality loss

C2PA, EXIF, and XMP live in blocks separate from the image data. Removing these blocks does not resample the image. A C2PA remover can delete them byte-for-byte while leaving the pixels unchanged.

2. Treat latent signals as purification, not a guarantee

SynthID is designed to resist mild transformations. Published approaches such as WAVES, MarkNull, and diffusion purification (arXiv, 2024–2026) use controlled transformations, re-encoding, or GPU-based regeneration.

Purification can weaken the signal, but it cannot guarantee removal. Claims of guaranteed SynthID removal should be treated with caution.

3. Understand text watermark detection

Text watermarks such as SynthID-Text and OpenAI’s “gamma” method distribute statistical patterns across tokens. Full paraphrasing can disrupt those patterns; this is documented in arXiv 2508.20228, which reports vulnerability to paraphrasing, copy-paste, and back-translation.

What the law says

  • Your own content: removing metadata from files you generated concerns your control over your data. The EU AI Act regulates marking and disclosure; it does not authorize deceptive use.
  • Other people’s content: removing markers from third-party work can violate copyright. In the United States, 17 USC 1202 prohibits removal of copyright management information, with penalties under 17 USC 1204 of up to $500,000 or five years.
  • Deceptive use: presenting synthetic media as authentic can trigger fraud, deepfake, or disclosure laws.

Summary

  1. Detect C2PA boxes, EXIF, XMP, and generator signatures.
  2. Remove metadata surgically and without quality loss.
  3. Treat latent-signal purification as weakening, not a promise.
  4. Work only on content you created or have the rights to modify.

OutWatermark AI provides free local detection and metadata cleanup with no account. Detect an AI watermark →