Which marking methods AI providers use, how platforms handle them, and what tools exist to check content — for images, video, and audio, not text. The framework of marking techniques follows the EU Code of Practice on the transparency of AI-generated content, grounded here in the technologies actually deployed today. Based on public information, with sources linked throughout.
None of this tells you whether the content is true — whether what it shows actually happened, or represents reality accurately. These are provenance measures: when they work, they show how a piece of content was produced — and, sometimes, by whom.
Methods
Each technique marks or records content in a different way — with different reliability, and different ways to defeat it.
A record attached to the file: who made it, with which tool, and when. It can be cryptographically signed or unsigned.
A pattern embedded in the pixels or audio itself, imperceptible to people. It can be secret-key (only the provider's tool can read it) or open (anyone can read it with a published decoder).
The provider computes a fingerprint — a compact perceptual hash of the content, not a copy of it — and keeps a record, so content can be matched back later even without metadata.
A mark placed on the content itself at generation — a logo burned into the corner of an image or video, or an audible tone or spoken disclaimer at the start of a clip.
Estimating whether content is AI-generated from statistical patterns that the generation process leaves in the content — no embedded signal involved.
Criteria
EU AI Act · Reliability
What a result actually tells you. A marking is reliable if finding it means the content really is AI-generated, and if the check rarely errs in either direction. Absence is a separate question: it usually means nothing at all.
EU AI Act · Adversarial robustness
Robustness assumes ordinary handling; security assumes an adversary. A marking that survives everyday use can still be stripped out, or forged onto content that was never AI-generated.
EU AI Act · Robustness
Whether the marking survives ordinary handling. Content is screenshotted, cropped, re-compressed and passed through apps that rewrite files. A marking that does not survive those steps is gone by the time anyone thinks to check.
Reliability + Security
Reliability and security together — and it splits in two. If a mark is present, does that prove the content is AI-generated? If a mark is absent, does that prove anything at all? Each answer only holds as far as the marking resists a deliberate attacker: presence is worth only as much as its resistance to forging, and absence only as much as its resistance to removal.
Robustness
Whether the marking survives ordinary handling. Content is screenshotted, cropped, re-compressed and passed through apps that rewrite files. A marking that does not survive those steps is gone by the time anyone thinks to check.
Verification tools
EU AI Act · Interoperability
Each tool is labeled with the marking it works with. A positive check is reliable; a negative one means little, since signals are often removed.
Reads and displays a file's Content Credentials. For content registered with Adobe's Content Authenticity app, it can also recover credentials after stripping or screenshots — a watermark/fingerprint lookup in Adobe's cloud, not a signature check. Cryptographic verification is Verify's job.
Checks whether an image or video came from Amazon's models (Titan, Nova). Runs in the Amazon Bedrock console or the DetectGeneratedContent API.
c2patoolReads and verifies C2PA from the command line. To confirm the signer is trusted, load the official trust list once, then verify:
c2patool --settings c2pa.toml init trust
c2patool --settings c2pa.toml file.pngReads and cryptographically verifies Content Credentials: who signed the file, which AI tool was used, and the edit history. Its open-source code validates in the browser (WebAssembly), so the file need not be uploaded — though that is what the published source shows, not something we confirmed on the live site. It's open source.
ExifToolReads metadata, without verification:
exiftool -a -G1 file.jpgAvailable via the Gemini chatbot interface.
Suggested prompt:
Analyze this image using your Verify AI tool to determine if it was AI-generated. Provide your detailed findings in two distinct sections: C2PA Results; SynthID Results.Checks whether an image came from Meta's models.
Checks whether an image came from OpenAI's models.
Preview's Inspector and a file's Properties → Details show EXIF/IPTC fields. They do not read C2PA.
Access requires an application; not open to the general public.
Supply chain
Provenance is not a property of a model, but of the whole system that generates AI content — added along the chain that carries it to you, and only at certain links in that chain.
The raw capability: a weights file that turns a prompt into an image, a clip, or a voice. It has no interface and marks nothing on its own.
The product built around the model — interface, filters, policy — and the link where a mark is added, after generation: a metadata record, an invisible watermark, or a perceptible mark.
Where the file is shared — social apps, messaging. It reads, shows, or strips those marks.
Every measure on this page is applied at the system or the platform — after generation, not inside the model. It is also why open-weight models are not listed: a model you download and run on your own hardware or cloud instance passes through no provider, so its output comes out without marking. And a safeguard trained into an open-weight model's weights is no backstop: where safety is baked in during training, fine-tuning has been demonstrated to strip it back out.
Model providers
What each provider applies to its output, broken out by media type. Reported absences are referenced as well.
Platforms
Platforms generally read provenance metadata at upload and remove it from the files they serve, so labels exist only within each platform. This table judges the distribution role only — what a platform does to files that pass through it. Where the same company builds generation features into the app (Veo clips in YouTube Shorts, Meta AI in WhatsApp), the marks on that output are a provider role, covered above.