AI Watermarking Explained: C2PA, SynthID, and Who Owns AI Content

May 8, 2026 · Editorial Team · 9 min read · ai-tools copyright legal

Two questions come up repeatedly in conversations about AI-generated content: how do we know something was made by AI, and who owns what gets made? These questions have technical answers (watermarking standards) and legal answers (copyright status), and in 2026, neither is fully settled. But both are far clearer than they were two years ago.

This covers the watermarking systems that exist, how they work, and the current copyright status of AI-generated content in the major jurisdictions, in plain terms.

What watermarking means for AI content

"Watermarking" in AI content means embedding a signal that identifies the content as AI-generated. There are three main types:

Visible watermarks are graphical overlays, logos, text, or patterns added on top of an image. Every tool that watermarks free-tier output this way (like Pika or Runway on free plans) is using visible watermarking. It's the most obvious form, and it's trivially removed with image editing tools. Visible watermarks tell you what tool made something, not whether something is authentic.

Metadata watermarks embed information about AI generation in a file's metadata (EXIF, XMP, or other header formats). This survives normal file handling but is stripped by basic metadata removal tools, photo resaving without metadata, or screenshot. ElevenLabs' watermark on generated audio works this way. Useful for audit trails in compliant systems; not reliable for adversarial detection.

Signal watermarks embed information in the content itself, the pixel values of an image, the audio waveform of a voice clip, in ways that are imperceptible to human perception but detectable by a corresponding detector. These survive format conversion, compression, and most casual processing. SynthID is the leading example. They can be defeated by determined adversaries but represent a meaningfully more durable signal than metadata.

C2PA: the content provenance standard

C2PA (Coalition for Content Provenance and Authenticity) is the most important standard to understand for anyone working with AI-generated media. It's not a single watermark; it's a provenance framework.

A C2PA manifest is a cryptographically signed record attached to a file that describes the content's history: what created it, what tools modified it, and when each step happened. The manifest is verified by cryptographic signature, you can check that the provenance record hasn't been tampered with.

The C2PA coalition includes Adobe, Google, Microsoft, Sony, the BBC, and a large number of camera manufacturers and platform companies. Support is being built into:

Adobe Creative Cloud (Photoshop, Premiere, Firefly, C2PA Content Credentials are natively generated)
Camera manufacturers including Canon, Nikon, and Leica (hardware-level C2PA signing for original photos)
Major AI tools including Midjourney and Adobe Firefly
Platforms including LinkedIn and TikTok (for displaying provenance information)

When you generate an image with Adobe Firefly, the resulting file carries a C2PA manifest identifying it as AI-generated by Firefly. If you then open that image in Photoshop and make changes, a new manifest entry records the editing step. The chain of provenance is preserved.

What C2PA solves: it creates a verifiable record of content origin for tools and platforms that implement it. When a news organization uses an image with C2PA credentials, they can verify whether it's original camera output or AI-modified.

What C2PA doesn't solve: tools that don't implement it generate content with no manifest. Stripping or ignoring the manifest is trivial. It's a trustworthy signal when present, not a universal detection system. Bad actors don't use tools that generate C2PA manifests for the content they're trying to pass as authentic.

The long-term value of C2PA is in building a baseline of provenance tracking across legitimate uses, journalism, commercial photography, creative work, where verification adds value. It's not primarily a detection tool for synthetic media used maliciously.

SynthID: signal watermarking from DeepMind

Google DeepMind's SynthID is a signal watermarking system that works differently from C2PA. Instead of attaching provenance metadata, it modifies the content itself to carry a watermark signal.

For images: SynthID subtly adjusts pixel values in patterns that human vision doesn't detect but a trained detector identifies. The changes are designed to survive JPEG compression, resizing, color adjustments, and cropping.

For audio: SynthID embeds a signal in the audio waveform. The modifications are inaudible but survive re-encoding, pitch adjustment, and mixing with background audio at moderate levels.

For text: SynthID's text watermarking works at the generation level, shaping token selection probabilities to embed a detectable pattern. This is the least durable, paraphrasing the output removes the signal.

SynthID is integrated into Google's AI tools (Imagen, Lyria for music, Gemini) and licensed to third-party providers. The detector is available through Google's API for platforms that want to run detection.

Durability limitations: SynthID's image watermark is designed to survive typical processing, but adversarial attacks, specifically designed to remove the watermark while preserving image quality, can degrade detection rates. Research from 2025 showed that while casual processing preserves SynthID signals well, targeted attacks can reduce detection accuracy. Google has continuously updated the watermark in response.

Practical use cases: SynthID is most useful for platforms distributing AI-generated content who want to maintain provenance (Google's own tools, licensed partners) and for organizations running detection workflows over content at scale (news verification, platform moderation).

Licensing: what AI tools actually give you

The licensing terms for AI-generated content vary substantially by tool and plan tier, and many creators misunderstand what they've agreed to.

Midjourney

Midjourney's commercial rights terms depend on your subscription:

Basic and Standard plans: personal use only; commercial use requires the Pro plan ($60/month) or annual Standard.
Pro and Mega plans: commercial use rights included.

Midjourney retains a license to use your prompts and generated images to train future models. You own the outputs in the sense of commercial use rights, but Midjourney can use them.

Flux (Black Forest Labs)

Flux models have different terms by version:

Flux Pro and Flux Dev: commercial use permitted; see the specific license file for each version. Flux Dev (the open-weights version) is licensed for non-commercial use; Flux Pro (API access) includes commercial rights.
Flux Schnell: Apache 2.0 license, commercial use permitted.

Stable Diffusion

Stable Diffusion base models (v1.5, SDXL, SD3) are released under open-source licenses. SDXL and SD3 use the CreativeML OpenRAIL+M license, which permits commercial use with restrictions on harmful content applications.

Community fine-tuned models on Civitai have varied licenses, some allow commercial use, some don't. Always check the specific model page.

Suno and Udio

Suno and Udio free plans include personal use only. Paid plans (Pro and above) include commercial rights to generated music. However, both services retain rights to use generated content to improve their models.

ElevenLabs

ElevenLabs free plan restricts commercial use. Creator plan and above include commercial rights for generated audio, including voice content. The specific terms of what "commercial use" covers are defined in their current terms of service, worth reading if you're using audio in anything that generates revenue.

Copyright status: who owns AI-generated content?

This is the big question, and the legal answers differ by jurisdiction and are still evolving.

United States

The U.S. Copyright Office has stated clearly and repeatedly that copyright protection requires human authorship. Content generated solely by AI cannot be registered for copyright.

The nuance is in "solely." The Copyright Office has approved copyright for:

AI-assisted works where the human made substantial creative selection and arrangement decisions
Works where AI was a tool, with humans directing specific creative choices

The 2023 Zarya of the Dawn case (Kristina Kashtanova) resulted in copyright for the textual elements and the selection/arrangement of AI images, but not the AI-generated images themselves. This is the current framework: copyright attaches to human creative decisions, not to AI output.

Practical implication: a creator who carefully curates, arranges, and modifies AI outputs may have copyright in the resulting work. A creator who types a prompt and publishes the output unchanged does not have copyright under current U.S. law.

This creates a real gap for content creators who rely on AI-generated images. You may not be able to stop others from reproducing your AI-generated images under U.S. law.

European Union

The EU does not have a uniform answer on AI copyright as of mid-2026. EU copyright law requires "the author's own intellectual creation," which courts have interpreted as requiring a human author who makes creative choices.

The EU AI Act does not address copyright directly, it regulates AI systems by risk level, not content ownership.

Individual EU member state courts are beginning to develop caselaw. The general direction aligns with the U.S.: purely AI-generated output without substantial human creative input is not protected. The threshold for what counts as "substantial human creative input" is not uniform across EU jurisdictions.

United Kingdom

The UK Copyright, Designs and Patents Act 1988 includes Section 9(3), a unique provision that grants copyright for "computer-generated works" to "the person by whom the arrangements necessary for the creation of the work are undertaken." This was written for rule-based software, not AI, but some legal analysis suggests it could extend to AI-generated content.

The UK Intellectual Property Office acknowledged the ambiguity in 2022 guidance and indicated a review of whether Section 9(3) should apply to generative AI outputs. As of mid-2026, no legislative change has been made, leaving the application of Section 9(3) to AI outputs to court interpretation.

The practical result: UK law may offer more protection for AI-generated content than U.S. law, but the specific cases haven't been fully tested.

China

China has the most notable court decision explicitly granting copyright to AI-generated content. A Beijing court ruled in 2023 that a specific AI-generated image had copyright protection because of the human creative input in directing the AI. This decision does not establish universal precedent across all AI output, but China's approach is more open to AI content copyright than Western jurisdictions.

What this means for creators in practice

The copyright gap for AI-generated content matters in specific ways:

Licensing to clients: if you sell AI-generated images, you can't guarantee copyright protection in the U.S. or EU. Be transparent with clients about this. Some commercial agreements require copyright warranties that AI-generated content may not satisfy.
Protecting your work: under current law, someone who copies your AI-generated images may not be infringing your copyright in most jurisdictions. Building in sufficient human creative modification (significant post-processing, selection and arrangement, substantial editing) is the practical path to stronger legal protection.
Using AI images commercially: using AI-generated images in advertising, products, or commercial content is generally fine from a copyright perspective, the output isn't protected by anyone else's copyright (absent training data disputes). The tool license is what governs, not copyright law.
Training data exposure: there are ongoing cases about whether AI models trained on copyrighted works create liability for outputs that are substantially similar to training data. This is a separate issue from output copyright, watch for developments if your work involves recognizable styles or characters.

For the latest terms on specific tools, the agent pages for Midjourney, Stable Diffusion, Suno, and ElevenLabs link to current pricing and terms information. The legal landscape in this area changes faster than most, building a relationship with an IP attorney familiar with AI is worth it if your work depends on AI content at commercial scale.