The debate over labelling synthetic media has spent years in the realm of principle. Should audiences be told when an image, a voice or a video was generated or altered by a machine? Is a label a courtesy, a warning or a form of censorship? Those arguments are not settled, but they are no longer the ones that decide whether labelling works in practice.
As transparency duties move from guidance into enforceable obligations in major markets, the question changes. It becomes operational. Can a mark survive the ordinary life of a media asset — cropped, re-encoded, resized, converted, syndicated, screenshotted and re-uploaded — and still be legible to whoever checks it at the end? That is not a communications problem. It is an engineering and procurement problem, and most organisations are further behind on it than they assume.
From a voluntary courtesy to an expected baseline
The direction of travel is clear. The European Union's rules for artificial intelligence include transparency obligations for providers of generative systems, requiring that outputs be marked in a machine-readable format and that certain synthetic content, including deepfakes, be disclosed. China has introduced labelling requirements for generated content. Other markets are consulting on comparable measures, and platform policies are tightening independently of statute.
What matters for delivery teams is the shape of the duty. It is not satisfied by a sentence in a terms-of-service document or a caption added at the point of publication. The obligation attaches to the output itself and is expected to travel with the asset. A provider that marks a file at generation time has done only the first step; the mark now has to survive everything that happens next. That catches organisations off guard, because labelling feels like a statement of intent. In practice it is closer to a data integrity requirement: it needs an owner, a specification, a test and a fallback when it fails.
Why a single watermark rarely holds
The fragility of provenance signals is the central engineering fact of this field. Metadata embedded in a file is routinely discarded by tools that rebuild the file rather than edit it: re-encoding for web delivery, format conversion, aggressive compression, screenshotting, screen recording and the re-upload pipelines operated by large platforms. None of these steps is malicious. Most are ordinary housekeeping, and each one can quietly strip the evidence a downstream checker depends on.
Imperceptible watermarking helps, but it is not a complete answer. There is a permanent trade-off between robustness and invisibility: the stronger the signal, the more likely it is to become perceptible or to degrade image or audio quality. Robust marks survive some transformations and not others, and determined removal is a research problem that is not standing still.
The more durable approach is layered. A signed provenance credential — the approach embodied in the Content Credentials work around the Coalition for Content Provenance and Authenticity — travels as a manifest asserting who made or changed an asset and how, backed by cryptographic signatures that let a verifier detect tampering. A durable watermark provides a recovery path when the manifest is lost. A visible label serves the audience. Publisher records provide the audit trail when no technical signal arrives intact. The weakness in that stack is the chain of custody: if any handler strips the manifest without preserving or re-signing it, the chain breaks, and a broken chain looks much the same as an asset that was never marked at all.
The duty runs the length of the toolchain
This is where framing labelling as a model-level obligation breaks down. The model provider is one participant. Between generation and audience sit asset management systems, creative editing suites, transcoding and delivery services, content management platforms, advertising systems, syndication partners and the platforms on which content is finally consumed. Each is a custodian of the signal, and each makes a decision, often by default rather than by design: preserve it, strip it or rewrite it.
A marketing team that exports a campaign asset at six aspect ratios may be running it through six separate pipelines, and the provenance outcome may differ across all of them. The practical implication is that provenance has to be tracked through the workflow, not treated as a file that accompanies an asset by luck. Teams that get this right instrument ingest and export points, log what entered and what left, and can say where a signal was lost without reconstructing the history by hand.
Marking and detection are not the same problem
A common conflation is to assume that if content is marked, detection is solved. It is not. Marking is a deterministic act by a producer; detection is an inference made by a third party, frequently using classifiers that are probabilistic and that fail in ways producers cannot control. False positives are the underrated risk. A detector that flags authentic footage as synthetic damages the credibility of the label itself.
The harder judgement calls sit in the grey zone of partial assistance. Background removal, object cleanup, voice enhancement, automated transcription and machine-assisted editing all involve generative components without producing wholly synthetic work. Rules that draw a bright line at generation say little about these cases. Organisations need disclosure thresholds, defined by content type and audience, plus an escalation route for the ambiguous ones. Those decisions are far cheaper to make in a policy workshop than during a live incident.
Design matters too. Labels that appear on almost everything stop functioning as signals, and provenance manifests that expose internal tooling or workflow detail can leak information a company would rather not publish. Usable provenance has to be informative to a verifier without being revealing to a competitor.
What procurement should be asking now
Because the signal is only as strong as the weakest handler, much of the practical work sits in supplier conversations. Buyers of creative, marketing, publishing and media tools should be able to establish, before signing, whether outputs carry provenance metadata, which standard is used, and whether that metadata survives the supplier's own processing steps.
Useful questions are concrete. Does the tool write a signed manifest at export, or only at generation? Which export presets and transformations preserve it, and which discard it? Can records be exported for audit, and for how long are they retained? What happens when an asset is re-ingested after a client has edited it elsewhere? Where does the supplier's own sub-processor chain touch the file?
The most reliable check is not a datasheet but a round trip: take a marked asset, run it through the supplier's actual pipeline, and verify what comes out the other side. Contract terms should reflect the same expectation, with clarity on who is responsible when a mark is lost in processing.
None of this is glamorous, and that is the point. The transparency regime now taking shape will be judged less on the ambition of its disclosure rules than on whether provenance survives contact with production. For a category such as the Global Tech Awards, which tracks work across the technology sector, the interesting engineering sits exactly where policy meets plumbing: systems that keep a signal intact at scale, without degrading the product or exposing the workflow behind it.
Organisations that treat labelling as a statement will find themselves unable to evidence it. Those that treat it as a chain of custody — specified, tested and written into supplier agreements — will be the ones for whom the label means something when a regulator, a platform or an audience actually checks.