AI Watermark Status: Who Marks What, Today
Which AI tools currently watermark content, how those marks work, and whether they can be detected or removed.
As of September 14, 2026, 7 provider surfaces publish a machine-readable mark on their output. This page tracks which providers watermark what, whether a detector exists, and when each verdict was last checked.
| PROVIDER | CONTENT | WATERMARK | DETECTION | REMOVAL | CHECKED |
|---|---|---|---|---|---|
| Claude Anthropic | Text | Statistical text watermark | Partial: Partial Private preview, by access request | Partial: Partial Full rewrite only; unverifiable | Sep 14 |
| Claude Anthropic | Generated files | C2PA metadata | Available: Available Readable locally; Claude Content Checker is free | Available: Available Metadata only; wording unaffected | Sep 14 |
| ChatGPT OpenAI | Text | None confirmed | Unavailable: Unavailable No confirmed statistical detector | Partial: Partial Hidden Unicode removable where present | Sep 14 |
| GPT image OpenAI | Image | Pixel watermark + C2PA metadata | Available: Available Verifiable via OpenAI’s Verify tool and API | Partial: Partial Metadata only; pixel mark stays | Sep 14 |
| ChatGPT audio OpenAI | Audio | Embedded audio watermark | Available: Available Verifiable via OpenAI’s Verify tool and API | Unavailable: Unavailable Embedded in the signal | Sep 14 |
| Gemini Google | Text | Statistical text watermark | Partial: Partial Detector portal is waitlist-only | Unavailable: Unavailable Rewrite only; not a character strip | Sep 14 |
| Gemini Google | Image | Pixel watermark + C2PA metadata | Partial: Partial Checks in the Gemini app and Google Search | Partial: Partial Metadata strips; pixels stay | Sep 14 |
| Firefly Adobe | Image | Content Credentials (C2PA) | Available: Available C2PA manifest readable locally | Available: Available Manifest removable; provenance breaks | Sep 14 |
| Any image file EXIF · XMP · PNG text chunks | Image | File metadata | Available: Available Parsed locally in the browser | Available: Available Removed and verified by rescan | Sep 14 |
| Other text models Copilot · Grok · Meta AI · Perplexity · Mistral · DeepSeek | Text | None announced | Unavailable: Unavailable No text watermark announced | Partial: Partial Hidden Unicode removable where present | Sep 14 |
Capability labels: Available · Partial · Announced · Unavailable. Each provider name links to its page.
Cite this page- As of September 14, 2026, no public Claude watermark detector exists. Anthropic's watermark detection is in private preview for eligible organisations (regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations, EU civil society groups, and enterprises with their own AI Act obligations), by access request. Anthropic's free Claude Content Checker reads Content Credentials on generated files, not the text watermark.
- Anthropic's rule is the model's launch date: Claude models launched on or after August 2, 2026 are marked at launch, and earlier models are being retrofitted. Fable 5.1 and Mythos 5.1, launched September 1, 2026, are the first models marked at launch. As of September 14, 2026, no public detector exists, so every per-model row still reads Unknown.
- Google's SynthID Detector portal is waitlist-only for journalists, media professionals, and researchers as of September 14, 2026.
- OpenAI has not deployed a text watermark for ChatGPT or the API as of September 14, 2026; its help centre says the goal is to expand provenance signals to all modalities including text, with no method or date announced. Images from ChatGPT carry SynthID in the pixels and C2PA Content Credentials in the file, and as of September 7, 2026 both signals are verifiable through OpenAI's public Verify tool and Content Provenance API, for OpenAI-supported signals only; a not-detected result from that tool says nothing about other providers or about a file whose marks were stripped. Run the OpenAI check
- The EU AI Act Article 50 deadline for retrofitting pre-existing systems is December 2, 2026.
DetectMark, "AI Watermark Status", last checked September 14, 2026. https://detectmark.com/ai-watermark-status
The underlying data is published under CC BY 4.0. Reuse it with attribution.
About this tracker
It is reviewed weekly, and out of band whenever a provider ships a change or a detector opens.
Anthropic (Claude), OpenAI (ChatGPT and GPT image), Google (Gemini and SynthID), Adobe (Firefly and Content Credentials), and the text models that have announced nothing: Copilot, Grok, Meta AI, Perplexity, Mistral and DeepSeek.
Two small sets, one for results and one for capabilities. A scan result is always one of four labels: Clean means we ran the check and found nothing. Found means we ran the check and a known signal is present. Unknown means the check cannot be verified, usually because no open detector exists. Unsupported means the check does not apply to this content. A capability cell, such as detection availability or removal, is always one of four labels too: Available means a check or removal can be performed today through any public means, not necessarily DetectMark itself. Partial means only some of it can be performed, or access is restricted, such as a waitlist portal. Announced means a provider or standard committed to it but nothing verifiable ships yet. Unavailable means nothing exists to perform. Per-Claude-model rows carry a third label, Unknown, meaning no public detector has confirmed the mark either way; the row’s note says whether the model is marked at launch per Anthropic or predates the cutoff. Never a percentage, and never an AI likelihood score.
Yes. Each provider page carries the source list behind its row, and the machine-readable export includes a source URL per record.
It is the date the verdict was last verified against the provider's own documentation, an independent test, or a published policy. It is never a generated or estimated date.