AI Image Generators Inject Hidden Tags. Most of Them Anyway

BySarah Fraser
fal.ai fast-sdxl c2pa: Laptop screen showing image file metadata and provenance details

fal.ai’s fast-sdxl endpoint was the one image-generation endpoint in this 4 August 2026 spot check that did not deliver machine-readable AI-origin markers. Five comparable commercial endpoints did: OpenAI gpt-image-1, fal-ai/ideogram/v3, fal-ai/flux/dev, fal-ai/flux/schnell, and fal-ai/recraft-v3 each returned a file with a C2PA manifest and the IPTC DigitalSourceType value for AI-generated media.

That matters because the result was per endpoint, not per platform. fal.ai exposes multiple distinct endpoints, and on the same platform, same API-style delivery path, and same inspection method, some outputs arrived marked while fast-sdxl arrived as a bare JPEG.

One image was generated per endpoint using the same prompt, “a plain red ceramic coffee mug on a white table, soft daylight”, then fetched from the returned HTTPS URL and inspected with ExifTool 13.55. The negative fast-sdxl result was repeated four times. This was a spot check of the API delivery path on 4 August 2026, not a broad audit, and endpoints can change without notice.

Endpoint-by-endpoint marking results on 4 August 2026

The head-to-head result is simple:

Endpoint Marker result
OpenAI gpt-image-1 C2PA manifest present; DigitalSourceType = trainedAlgorithmicMedia
fal-ai/ideogram/v3 C2PA manifest present; DigitalSourceType = trainedAlgorithmicMedia
fal-ai/flux/dev C2PA manifest present; DigitalSourceType = trainedAlgorithmicMedia
fal-ai/flux/schnell C2PA manifest present; DigitalSourceType = trainedAlgorithmicMedia
fal-ai/recraft-v3 C2PA manifest present; DigitalSourceType = trainedAlgorithmicMedia
fal-ai/fast-sdxl No C2PA/JUMBF manifest; no DigitalSourceType value

The C2PA 2.2 specification defines the manifest structure, and the IPTC URI used here, trainedAlgorithmicMedia, is the standard machine-readable label for AI-generated media. Five endpoints shipped both; one shipped neither.

The file formats varied, which is useful because it rules out an easy excuse. OpenAI gpt-image-1 returned PNG, fal-ai/recraft-v3 returned WebP, and both fal-ai/flux/dev and fal-ai/flux/schnell returned JPEG. JPEG itself was not the reason fast-sdxl came back unmarked, because other JPEG outputs in the same spot check carried the expected metadata.

The endpoint-level split is the real finding here. A platform can truthfully say some of its image endpoints emit machine-readable markers while another endpoint on the same service does not. A single spot test of one model is not enough to conclude that “this provider marks its output.”

Under the European Commission’s guidance on Article 50 transparency duties, those obligations began applying from 2 August 2026. The Commission’s separate fact page says a limited grace period to 2 December 2026 is envisaged only for generative systems placed on the market before 2 August 2026 and only for the marking and detection duty under Article 50(2). That timing is context, not the measurement: the delivered files either contained markers or they did not.

What the manifests identify and what they do not prove

The cryptographic claim on the marked fal.ai outputs identified the hosting intermediary, not the upstream model maker. In the delivered files, the claim generator was fal-ai/..., with endpoint-specific assertions attached by fal.ai, not by Black Forest Labs, Ideogram, or Recraft.

That is a practical point, not a legal conclusion. The AI Act overview does not settle from file inspection alone whether the legally responsible party for Article 50(2) in a reseller setup is the host, the model developer, or the downstream deployer. But the bytes do show who is making the signed claim in the delivered artifact: on these marked reseller-hosted outputs, it was fal.ai.

A machine-readable marker and a third-party-verifiable provenance chain are also not the same thing. A C2PA manifest can exist without offering the same level of outside verification across providers. The C2PA technical specification supports cryptographically verifiable provenance structures, but implementations can differ in what supporting material they embed.

In this spot check, OpenAI gpt-image-1 produced the strongest version of that stack: a manifest with the sort of certificate and timestamp material that supports later third-party verification. The marked fal.ai outputs still counted as machine-readable because they carried a C2PA manifest plus the IPTC AI-origin label, but they did not present the same full verification chain in the delivered file. That is the difference between “detectable by software” and “independently auditable later.”

“Providers of AI systems generating synthetic audio, image, video or text content should ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.”, European Commission guidance on transparency obligations

A machine-readable marker and a third-party-verifiable provenance chain are also not the same thing.

A provider’s consumer web app may embed provenance markers that its API path does not, and the reverse can also be true. This article is about the API delivery path only.

What this spot check rules out

The negative fast-sdxl result held up better than a one-off miss. Four separate fast-sdxl generations came back unmarked. That rules out simple generation flukes.

It also rules out a tooling problem in the inspection path. The same session, same code path, same HTTPS fetch method, and the same ExifTool 13.55 build detected markers on positive controls including fal-ai/flux/dev and fal-ai/ideogram/v3. If the parser or download path had stripped or missed the metadata, the controls should have failed too.

It was not just a format conversion issue either. fast-sdxl delivered JPEG without a C2PA/JUMBF manifest and without the IPTC DigitalSourceType value, while other JPEG endpoints on the same platform delivered both. That narrows the plausible explanation to endpoint-specific output handling.

The strongest claim the measurements support is narrow: on 4 August 2026, fast-sdxl stood out as unmarked in delivered API files while five comparable endpoints did ship machine-readable AI-origin markers. The strongest claim they do not support is a broader legal judgment. Whether any given setup satisfies Article 50(2), and who exactly bears that duty in a hosted-model chain, is not settled by file inspection alone.

One image was tested per endpoint, except the negative fast-sdxl result, which was repeated four times. The next useful measurement is obvious: rerun the same endpoint matrix after providers update their delivery paths, especially around the 2 December 2026 transitional date described by the Commission.

Key Takeaways

  • fal-ai/fast-sdxl was the only tested endpoint on 4 August 2026 that delivered no C2PA manifest and no IPTC DigitalSourceType AI-origin label.
  • Five other tested endpoints, OpenAI gpt-image-1, fal-ai/ideogram/v3, fal-ai/flux/dev, fal-ai/flux/schnell, and fal-ai/recraft-v3, did deliver both machine-readable markers.
  • The mixed results on fal.ai show that AI-origin marking is per endpoint rather than per platform.
  • On the marked fal.ai files, the cryptographic claim identified fal.ai as the asserting party, not the upstream model developer.
  • Machine-readable marking and independently verifiable provenance are different bars, and the delivered files in this spot check did not all meet the stronger one equally.

Further Reading