Generative Media, Editing, and Provenance

The same diffusion or autoregressive model may serve all of them. Their obligations still differ. The difference is not a matter of care taken by the operator; it is fixed by what the task promises.

Free generation reconstructs no scene. It owes diversity, prompt adherence, rights, and disclosure, but no fidelity to an original. Source-conditioned editing owes preservation of declared properties: masks, identity and geometry constraints, and a comparison of the regions that were supposed to stay put. Restoration and enhancement carry the heaviest contract of the three, because the detail the model supplies is precisely the detail a reader will treat as evidence. PULSE belongs in that third column. Its model card states the bound of the whole class rather than a defect of one implementation: “It will not help identify or reconstruct the original image.” A restoration pipeline that cannot say that out loud to its users has not understood which task it is running.

Media provenance requires more than a watermark placed at export, and the reason is on the record. On 20 November 2024 the National Institute of Standards and Technology published AI 100-4, an overview of technical approaches to digital content transparency. It describes what happens to a record once the file leaves your systems: “Metadata recorded within a file can similarly be stripped altogether, as it often is when files are shared (e.g., via social media platforms). Metadata may be stripped to deceive recipients of the content about its provenance or for benign reasons such as privacy protection.”

Two steps below follow directly from that sentence. The original is captured and retained because, when the attached record does not survive, the unedited source is the only remaining copy of the evidence. Provenance is published as a layer rather than as a single mark because stripping is routine. And because stripping is often innocent, a missing record cannot by itself be read as a confession.

A signed provenance record can identify the tool, publisher, and sequence of edits associated with an asset. It can help detect missing records or unauthorized modification. It does not prove that the camera captured a truthful scene, that the editor held every necessary right, or that the described event occurred.

That is not a sceptic's complaint from outside. The standard says it first. The C2PA and Content Credentials Explainer, version 2.2, released on 22 April 2025, poses the question as its own heading — "Can provenance information be used to determine whether a digital asset, such as an image or video, depicts the truth?" — and answers it: “Provenance information can help establish the truth about the origin, history and authenticity of digital content, by providing evidence for its creation, discovery, ownership and movement over time; but provenance information alone cannot tell you whether the digital content is true, accurate or factual.” That is the Coalition for Content Provenance and Authenticity describing the limits of its own specification.

NIST AI 100-4 makes the parallel point one level down, about the signature rather than the claim it carries: “A signature cannot verify that the metadata is accurate; on a technical level, it merely serves a notarization function, indicating that the signer attested to the existence of the metadata at a given time”. A valid signature tells you who asserted what, and when. Whether the assertion is true is a separate investigation. Provenance is what makes that investigation possible, not what concludes it.

Evaluation is worth little unless it separates these failures instead of collapsing them into one aesthetic score. One of the six already has a regulator and a date attached to it. The Federal Communications Commission adopted Declaratory Ruling FCC 24-17 on 2 February 2024 and released it on 8 February: “In this Declaratory Ruling, we confirm that the TCPA’s restrictions on the use of “artificial or prerecorded voice” encompass current AI technologies that generate human voices.” A cloned voice placed on a call therefore requires the called party's prior express consent, absent an emergency purpose or an exemption. The audio failure and the rights failure below are, in that setting, the same event seen from two directions.

Build a release set that reflects both creative goals and what has to be preserved. A published one already shows the shape. HEIM, a 2023 benchmark for text-to-image models, scores them against twelve named aspects: text-image alignment, image quality, aesthetics, originality, reasoning, knowledge, bias, toxicity, fairness, robustness, multilinguality and efficiency.

Its abstract gives both the scale and the result: “We curate 62 scenarios encompassing these aspects and evaluate 26 state-of-the-art text-to-image models on this benchmark. Our results reveal that no single model excels in all aspects, with different models demonstrating different strengths.” Twelve aspects, 62 scenarios, 26 models, and no winner across the board. That is the empirical reason a single aesthetic score cannot stand in for the evaluation. It is also why step 2 below needs several rubrics rather than one. A model chosen on image quality alone is being selected without evidence on eleven other axes it was also measured against.

Generated media can support design work, make products accessible, simulate scenes nobody photographed, and carry an idea to someone. The same capabilities can fabricate records or introduce unsupported detail into source-conditioned work. The four documents behind this lesson each mark one boundary of the resulting policy. PULSE's model card marks what a restoration can never be asked to prove. The C2PA Explainer 2.2 marks where provenance stops. Saberi and colleagues mark how far an image watermark degrades under attack. HEIM marks how many aspects a release decision is actually made across. Article 50(2) of the EU AI Act sets a floor under all of it from 2 August 2026, and the SynthID-Text result in Nature suggests the floor is affordable for text.

The next lesson turns to evaluation across generative applications. How good the media looks becomes one part of a broader evidence portfolio that includes task success, safety, latency, cost, and human judgment.