Your Deepfake Label Has to Survive the Repost

A disclosure label is only useful while a viewer can see it. That sounds obvious, but the file leaving an editor is rarely the file reaching the audience: social platforms decode it, resize it, change its aspect ratio, lower its bitrate, generate previews, and sometimes let another user publish a clipped section without the original caption.

The concrete test case here is a vertical campaign video with an AI-generated speaker. Its master export carries a disclosure near the bottom edge, a disclosure in the post caption, and provenance metadata in the file. On the editing timeline, all three appear to cover the same obligation. After ordinary publishing transformations, they do not.

The metadata may disappear during upload. A caption can become detached from a downloaded copy. The burned-in label, meaning text rendered into the video pixels rather than added by the player, travels with the picture but can still be cropped away or reduced until its letterforms break down.

This is a testing workflow for campaigns, newsrooms, and other publishers. It is not legal advice, and it does not replace the exact wording or placement required by a law, regulator, election authority, contract, or platform rule.

Turn the policy sentence into a pass-or-fail test

Start with the requirement, not the graphics template. For this workflow, the internal acceptance criterion is:

“The disclosure must remain visible and legible in every published derivative, including cropped, clipped, recompressed, downloaded, and reposted copies.”

That sentence is stricter than “add an AI label.” It names the condition that must survive and the transformations the team must test. A legal or platform requirement may define “conspicuous,” prescribe particular words, or apply only to certain synthetic media; if so, quote that language in the test plan and do not replace it with a creative-team summary.

Also record who enforces each rule. A platform may enforce its own label through an upload interface, while a publisher enforces a burned-in disclosure through export review. A proposed regulation is not an enforced requirement, and an internal standard is not a statement of law. Keep those columns separate in the approval record.

For the vertical campaign clip, make the pass condition observable: a reviewer must be able to read the complete disclosure during normal playback on a phone-sized display, after each derivative has been created. “The label layer exists” is not enough. Neither is “the metadata was present in the master.”

Build one test master with deliberate weak points

Create a clean master containing the same visual conditions as the real publication: bright and dark shots, motion behind the label, cuts, and any subtitles or lower-third graphics competing for space. Use the real disclosure wording. Placeholder text often has a different length and produces a misleading fit.

A practical phrase identifies both the mechanism and its scope, such as “AI-generated video” or “AI-altered video and audio.” “Synthetic media” may suit a policy document, but some viewers will not know what it covers. “Deepfake” is familiar yet imprecise, especially when only a voice, face, or background was changed. If an applicable rule supplies exact words, that requirement controls the editorial preference.

Keep the disclosure on screen for the full duration. An opening card fails as soon as someone clips the middle, while an end card fails in autoplay previews and shortened exports. A persistent label costs visual space, but that is the tradeoff: the disclosure can travel with an arbitrary segment only when it appears in that segment.

For the first export, place the label away from every outer edge and outside areas commonly covered by interface controls. Do not assume one universal safe-zone template. Platforms place captions, buttons, account names, and progress bars differently, and those interfaces change. Instead, generate several placement variants, including one near the bottom and another farther toward the visual center, then subject both to the same crop test.

Test more than one text size. Relative sizing is more useful than a fixed point value because the output will be resized; prepare small, medium, and large versions based on frame height, then choose the smallest version that remains readable after the harshest accepted derivative. Preserve strong contrast with a solid or near-opaque backing panel. A thin outline around white text is vulnerable when recompression introduces block noise around moving detail.

The vertical campaign clip now has a real test object: exact wording, persistent timing, competing graphics, and placement variants. Save it as the reference master and retain its file hash so reviewers can identify which export entered the test.

Put the label through the transformations users create

Upload the reference master through the same account type and publishing path intended for release. Draft tools, advertising systems, scheduled-post services, and native mobile uploads may produce different derivatives, so testing an unrelated route does not clear the production workflow.

Wait until processing finishes, then inspect the streamed version rather than relying on the local upload preview. Download a copy where the service permits it. Make a screen recording as a fallback observation, clearly marking it as a screen capture rather than treating it as the platform’s source file.

Next, create the transformations that threaten the disclosure. Crop the vertical video into the square and portrait formats used by the campaign. Extract a short segment from the middle, excluding both the opening and closing frames. Recompress a downloaded copy through another ordinary upload.

Finally, view the smallest preview or embedded presentation that the publication expects audiences to encounter.

This sequence isolates different failures. Clipping tests whether the disclosure depends on timing. Cropping tests placement. Recompression tests letterform clarity and contrast, while the download-and-repost path reveals which information traveled inside the pixels and which depended on the original post.

Review at normal playback speed first. Pausing and zooming can diagnose a failure, but they cannot turn unreadable text into a pass. Record the derivative, device orientation, wording shown, visible duration, and result in a simple test log. Attach screenshots of failures because “label too small” leaves the next editor guessing about the conditions.

In this workflow, the edge label is the predictable weak point. Recompression usually leaves its pixels present, but a tighter crop can remove them outright, and interface furniture can obscure text that technically remains in frame. Moving the disclosure inward protects it from cropping at the cost of covering more editorial content. That cost should be decided during layout, not discovered after publication.

Treat metadata as evidence, not the only notice

File metadata is information stored alongside the media rather than drawn into its frames. Inspect the master with a media probe such as ffprobe or a metadata reader such as ExifTool, then inspect each downloaded derivative with the same tool. Record fields that vanished or changed.

Ordinary encoding pipelines can discard descriptive metadata because the platform creates a new file rather than passing the upload through unchanged. A post-level AI label may remain visible on that platform, but it is attached to the post interface, not necessarily to a downloaded video. Neither mechanism should be described as persistent until the derivative test shows that it persists.

Cryptographic provenance systems, including Content Credentials based on the C2PA standard, can provide signed information about a file’s origin and edits. They offer stronger evidence than an editable text field when supported end to end, but a platform can still strip the manifest or publish a derivative that a receiving service does not display. Use provenance as a machine-readable receipt. Do not make it the sole viewer-facing disclosure.

The resilient setup uses separate channels with separate jobs. The burned-in label informs a viewer who encounters the video alone. Platform labeling informs viewers inside that service and may satisfy a platform workflow. Metadata and provenance support later verification when tools preserve and expose them.

Duplication is intentional because no one channel survives every path.

Fix failures at the source, then rerun the chain

If the wording becomes illegible, enlarge the text, strengthen its backing panel, or shorten an editorial phrase only when no controlling requirement prescribes the text. If a crop removes the label, move it into the intersection of all approved crop areas or create format-specific masters. If clipping removes it, extend the label across every frame rather than adding more opening-card time.

Do not patch only the failed repost. Update the reference master, increment the internal revision, and rerun the full publishing chain, because moving a label inward may solve cropping while creating a collision with subtitles or a speaker’s mouth. The final approval should point to tested outputs, screenshots, metadata inspection, and the exact policy sentence used as the pass condition.

For the vertical campaign clip, publication is cleared only when its middle segment still carries readable wording after the approved crops and repost path. The master timeline is evidence of intent. The derivative is evidence that the disclosure reached the viewer.

Questions people ask

Is a disclosure in the post caption enough?

A caption can inform viewers on the original post, but it may not follow a downloaded video, an embedded player, or a clipped repost. Keep the caption when the platform or editorial policy calls for it, and pair it with a burned-in label that remains attached to the video pixels.

Where should a deepfake label appear on vertical video?

Place it inside the area shared by every crop you intend to publish, clear of subtitles and likely interface controls. There is no permanent universal safe zone, so export placement variants and test them through the real account, crop formats, and viewing interfaces before approving one.

Can metadata replace a visible AI disclosure?

No, not for a workflow that requires viewers to notice the disclosure. Metadata and signed provenance can help establish origin or editing history, but platforms may remove them and viewers may never inspect them. Treat those records as supporting evidence while the burned-in wording handles immediate notice.

How often should publishers repeat this test?

Repeat it when the wording, template, export settings, upload route, account type, or target platform changes. A previously approved layout can fail after a platform changes its crop behavior or interface overlays, so retain the reference clip and rerun the derivative chain during periodic publishing checks.