Reposted videos lose AI provenance

AI labels are becoming routine on major social platforms, but the provenance trail that lets those labels travel with a video is still the weak link. Researchers audited 10,722 public posts across Instagram, TikTok, X, and YouTube Shorts and found that only about one in three expert-identified deepfakes had a platform-applied label. If platforms preserve and pass on the records, a family watching an alleged flood video could see the same AI-origin marker wherever the clip appears.

A person scrolling past fabricated war footage or a realistic fake of a politician usually has no reliable way to tell whether it came from a camera or a generator. The information that could answer that question may disappear as a file is uploaded, downloaded, edited, and reposted. A video can cross several platforms carrying no visible clue about where it came from.

The trail behind a video

That missing information is called provenance. It is a record of a file’s origin or of the use of artificial intelligence to make it. Cryptographic provenance metadata is designed to work like a signed note attached to the file. Watermarks try to make the same fact visible to people.

Neither is hard to damage. Converting a file, taking a screenshot, or letting a platform process it can remove metadata. A visible watermark can be cropped out or covered. A detector that judges only the pixels has to guess, and that gets harder as generated images and videos become more realistic.

The label also answers a narrower question than many people think. It can say that content was made or changed with AI. It cannot say whether the claim in the content is true. Authentic footage can mislead, while a synthetic illustration can be harmless.

Under the EU AI Act, providers of generative-AI systems must make their outputs machine-readable. The Digital Services Act asks the largest platforms to address systemic risks, but it does not prescribe one label design or one detection method.

Labels that stop at the platform door

An audit of Instagram Reels, TikTok, X, and YouTube Shorts found that automated labels are now operating on all four services. Creators can add labels themselves, but platforms also apply some labels without waiting for a confession.

The weak point is what happens after a file moves. The researchers found that platforms generally stripped the original AI-provenance signals from downloaded uploads. TikTok appeared to consistently attach signals to labels that TikTok itself had applied, but that did not create a shared trail across services.

The audit collected 10,722 public image and video posts using keywords linked to AI and systemic risks, then manually examined a sample of them. Among 43 deepfakes identified by experts through the risk-related searches, only about one in three had a platform-applied AI label. Those posts had a median of 160,000 views, and the average was 8 million among posts for which view data was available.

The researchers also uploaded files from ten generative-AI tools. Of the source files that carried detectable provenance signals before upload, 61% received an automatic label within 24 hours. Instagram labelled the most of the tested uploads, followed by TikTok, X, and YouTube. The result is a system that can recognise some synthetic media, but often breaks the evidence that another service would need to recognise it again.

When the marker travels with the clip

If platforms preserve provenance signals, check them consistently, and put the same visible label on every re-upload, a person watching an alleged flood video in a family kitchen could see the same plain AI-origin marker whether the clip arrived through Instagram, TikTok, X, or YouTube. They would not have to open captions or hunt for visual glitches before deciding how much trust to place in it.

That would make provenance useful as shared infrastructure rather than a voluntary badge. It could give platforms a cheaper and more defensible way to label files than judging every image from its pixels. It could also give auditors and regulators comparable information about what platforms label and why, while giving creators a clearer way to distinguish disclosed synthetic work from deceptively presented material.

For that future to arrive, the records must survive editing and distribution, platforms must pass them between services, and labels must be prominent and reliable. Otherwise an unlabelled clip will still be mistaken for proof of authenticity. The paper is an unpublished, non-peer-reviewed preprint, and its authors describe the results as a time-specific, context-specific audit rather than an assessment of legal compliance.

Sources

Drowning in AI Slop: How Social Media Platforms (Do Not) Label AI and Deepfake Content under EU law