AI provenance · Evidence brief 06

Which AI tools watermark their content?

A current provider-by-provider guide to C2PA Content Credentials, SynthID, invisible watermarks, visible labels, and what can actually be verified.

The short answer

OpenAI, Google, Amazon, Microsoft, Adobe, and Meta document provenance signals for at least some generated media, but they do not all use the same mechanism. C2PA is an open metadata standard; SynthID and Amazon watermarks need compatible provider detection. Anthropic says Claude currently produces text and text-to-speech output and does not claim an equivalent deployed watermark.

01 · What the evidence supports

The strongest reading of the current sources

OpenAI now documents C2PA plus SynthID for supported images and SynthID for supported audio. Google applies SynthID across image, video, audio, and some Gemini text, and exposes verification in Gemini for supported media. Amazon Titan Image Generator and Nova Canvas use an invisible watermark with an AWS detector, while Titan also adds C2PA. Microsoft Copilot and Adobe Firefly add C2PA Content Credentials to supported images. Meta documents visible labels, IPTC/C2PA-compatible metadata detection, and invisible markers for some Meta AI imagery. The C2PA 2.4 specification, published in April 2026, also formalizes durable credentials and soft bindings without turning provenance into a judgment that content is true.

Important distinction

Open C2PA/IPTC declarations can be inspected locally when they remain embedded in a supported file. Pixel-, audio-, video-, and token-level watermarks such as SynthID or Amazon’s signal are provider-specific or proprietary and cannot be inferred from ordinary metadata. A generic AI classifier is a probability estimate, not watermark verification or proof of origin.

02 · Evidence ledger

See the source behind the answer

These are direct source links, not citations to other SEO summaries.

S03
Official documentationAmazon Web Services

Amazon Titan Image Generator G1 model

AWS states that Titan Image Generator adds an invisible watermark and C2PA metadata and documents provider-side watermark detection for Titan and Amazon Nova Canvas.

S04
Official documentationMicrosoft

Transparency Note for Microsoft Copilot

Microsoft states that Copilot-generated images carry C2PA-based Content Credentials that mark their source and can be inspected with compatible provenance tools.

S05
Official documentationAdobe

Content Credentials overview

Adobe states that Firefly automatically applies Content Credentials to assets whose pixels are fully generated with supported Firefly creation features.

S07
Official documentationAnthropic

Anthropic Transparency Hub: Voluntary Commitments

Anthropic says Claude currently has text-based outputs, including text artifacts and text-to-speech voice output, and that it continues to evaluate watermarking developments rather than claiming a deployed output watermark.

S08
Technical standardC2PA

C2PA Technical Specification 2.4

C2PA 2.4 defines cryptographically bound provenance, trust lists, soft bindings and durable credentials, while explicitly separating valid provenance from judgments about whether content itself is true.

03 · Confidence boundary

What remains uncertain

  • Provider coverage varies by product, model, modality, account, export path, region, and release date; a company-level label never guarantees that every output carries the same signal.
  • No signal found does not prove human authorship: metadata may have been stripped, a watermark may have degraded, the file may predate a rollout, or the source may be unsupported.
  • Detection services can make errors, and robust watermarks can still weaken under heavy editing, transcription, translation, short text, recompression, cropping, or format conversion.
  • C2PA confirms the integrity and signer context of assertions only after full cryptographic validation; merely finding a C2PA container is not signature or trust-chain validation.
04 · Safest action

What to do before you share

  1. 01

    Use BeforeShare to inspect supported embedded C2PA, IPTC, prompt, model, and creator-tool metadata without uploading the file.

  2. 02

    Use the relevant provider verifier for pixel-, audio-, video-, or text-level watermark checks when sharing the file with that provider is acceptable.

  3. 03

    Record the exact detector, date, result, file hash, and limitations when provenance matters for journalism, compliance, or evidence.

  4. 04

    Treat a negative result as “no supported signal found,” never as proof that a human created the content.

05 · Questions people ask

Quick clarifications

Can BeforeShare detect SynthID?

Not from ordinary metadata. BeforeShare can report supported C2PA and IPTC evidence embedded in the file, but SynthID is a content-level signal that requires a compatible Google, OpenAI, or partner detector. The file would need to be checked through that separate service.

Can an AI watermark prove which model made an image?

Only within the scope of the specific signal and detector. A trusted C2PA manifest can identify signed provenance details; a provider watermark can support attribution to that provider. Neither proves factual accuracy, ownership, or the absence of later edits outside its supported checks.

Are ChatGPT text answers watermarked?

OpenAI’s current public provenance documentation covers supported images and audio, not a universal ChatGPT text watermark. Google documents SynthID text watermarking for some Gemini-generated text. Do not use writing style or a generic classifier as proof of which model wrote text.

Editorial note

This is a source-labeled synthesis, not a claim that BeforeShare independently tested every platform, client, or version. We update the review date when we re-check the cited evidence or materially revise the conclusion.