Synthetic media: check provenance and consent before you post

Before an AI-generated or AI-edited image or video of a real person goes anywhere public, two separate questions need a yes: did the person consent, and is the synthetic origin disclosed. A practical checklist for both.

What you should be able to do

A generated or edited image of a real, identifiable person needs two separate answers before it goes anywhere public: did that person agree, and does the post disclose that AI touched it. Neither question is answered by the image looking convincing.

Saved only in this browser.
In this article

Someone generates a birthday video of a colleague giving a heartfelt speech they never gave. A parent uses an AI tool to “improve” a family photo by swapping in a better smile from a different shot. A creator turns a friend’s holiday snapshot into a stylized illustration and posts it. None of these examples are malicious. All of them raise two separate questions: did the person in the image agree to this, and does anyone who sees it know AI was involved?

Here is a practical way to answer both questions before you post, not after someone asks why they were not told.

Consent, lawful basis, disclosure, biometric data, publicity rights, copyright, defamation, and platform duties vary by jurisdiction and use. Treat the checklist as a minimum editorial control, not a legal compliance opinion.

Two questions, not one

It is tempting to collapse “is this okay to post” into a single gut check, but consent and disclosure are genuinely separate failure modes.

Consent is about the person depicted. Did they agree to being generated, edited, or synthesized in this way, for this purpose, in this context? A person can be completely fine with a fun filter for a private chat and genuinely upset about the same technique used in a public post, a dating profile, or a work presentation. Consent is contextual, not a one-time checkbox.

Disclosure is about the audience. Does a reasonable viewer understand that what they are looking at was AI-generated or AI-edited in a way that changes its meaning? A viewer can be fully aware that filters exist in general and still be misled by a specific piece of content that does not say so.

You can satisfy one and fail the other. A fully consented AI video of a real person can still mislead an audience if it looks like an unedited recording and nothing marks it otherwise. A clearly labeled “AI-generated” image can still violate someone’s consent if they never agreed to be depicted that way at all.

Consent for the original photo is not consent for a synthetic transformation of it. Someone who let you take their picture at a party did not automatically agree to appear in an AI-generated video, a stylized deepfake, or an edited version depicting them somewhere they never were.

What “identifiable” means here

The checklist in this article applies whenever a real, identifiable person could be recognized in the result - by face, voice, name, distinctive setting, or context that narrows it to one person even without a clear face. The consent step differs for fully synthetic people or your own likeness. Audience disclosure still needs checking, including when you depict yourself. When in doubt about whether someone is identifiable, assume they are; the cost of asking is a short message, and the cost of skipping the question is a much harder conversation later.

Before generating or posting synthetic media of someone else, work through these in order:

  1. Does the person know the proposed transformation and use? Ask before generating it; do not substitute an assumption that they would be fine with it.
  2. Did they agree to this specific use? A person consenting to an internal team photo did not consent to a public marketing video built from it.
  3. Would the context surprise them? A humorous private-chat edit and the same edit posted publicly are different requests, even with identical output.
  4. Can they ask you to take it down? If the honest answer is “not really, once it’s posted,” that is a signal to slow down before posting, not after.

If any answer is no or unclear, ask the person directly before you generate or share anything. Describe the source material, transformation, audience, and whether the service stores or trains on uploads; agreement to one of those is not agreement to all of them.

The disclosure check

Separately, before anything goes anywhere a stranger might see it:

  1. Would a reasonable viewer assume this is an unedited photo or recording, without a label? If yes, and it is not, label it.
  2. Does the platform provide an AI-content label for this use, and is it turned on? Check the platform’s current policy on the day you post; do not assume the feature exists or survives resharing.
  3. Is there a human-readable disclosure near the content, not just embedded metadata? Metadata can be stripped by re-uploads, screenshots, and format conversions; a caption may not travel with the file, and a visible label can be cropped. Inspect the copy you publish.
  4. Does the disclosure describe what actually changed? “AI-enhanced” covers a lot of ground; if the change is material - a different setting, words the person never said, an edited expression - say so specifically.

The EU AI Act’s transparency provisions became applicable on 2 August 2026. Article 50 distinguishes provider duties around machine-readable marking from deployer duties to disclose certain deepfake and synthetic content (European Commission transparency overview). The Commission’s current Article 50 FAQ describes a limited transition to 2 December 2026 for providers of systems placed on the market before 2 August 2026, only for Article 50(2) marking and detection. This is not a general delay of deployer disclosure duties. Definitions, exceptions, actors, and specific facts determine whether a duty applies; obtain legal review rather than treating this checklist as compliance.

A fictional invitation request separately records the photo, service, transformation and event-group audience, while disclosure explains that the illustrated scene did not occur.
Ask about the exact use before upload and creation, then inspect the disclosure on the published copy. Open the full-size diagram.

Provenance signals help, but do not replace either check

Tools like Content Credentials (the C2PA standard) can attach a cryptographically signed record of an asset’s creation and editing history, which is a genuine improvement over an unlabeled file (C2PA technical explainer). But a valid credential establishes the integrity and signer of recorded assertions; C2PA explicitly says it does not make a value judgment that every provenance assertion is true. It also does not answer “did the depicted person agree” or “does the viewer actually see the disclosure.” A perfectly signed, fully provenance-tracked video can still be posted without the subject’s consent, and a credential can be stripped by an ordinary screenshot or re-upload before it ever reaches your audience. For more on what provenance can and cannot prove, see Content Credentials and watermark basics for the individual-reader version, or provenance, watermarking, and Content Credentials for the organizational publishing-policy deep dive.

NIST’s synthetic-content transparency report explains why provenance, watermarking, labelling, and detection each have different limitations. The C2PA Content Credentials specification defines what validation can establish. Neither source makes a consent determination.

If you are the one generating the media, keep your own record of what you made, from what source, with whose agreement, and when - even a one-line note. If a question comes up later, “here is what I generated and here is the message where they agreed” documents the permission you relied on. It does not settle every rights question or guarantee that all later uses are authorized.

What this does not cover

This checklist is for content you are creating or sharing about someone else with generally good intent. It is not the process for what to do if you discover a deepfake made of you without your consent - that is a different, harder situation, covered in the adult first-response plan. Before upload, check the provider, account, retention and sharing settings as well as the person’s permission; the photo-consent workflow brings those decisions together. And if the concern is specifically about whether an image or video looks real versus AI-made, rather than about consent, AI image limits, not detectors covers what you can and cannot reliably tell.

Record the permission and the label separately

Invented request: “May I use this photo to create a clearly fictional illustrated invitation, upload it to the named editing service, and share the result with our event group? I will show you the draft first.” A yes to that request does not authorize a public advertisement or a voice clone.

For the agreed invitation, a disclosure could read “AI-assisted illustration based on an authorized photo; this scene did not occur.” Keep that visible on the asset where practical and inspect the published version. Captions, metadata and visible labels can all be lost or cropped in later copies; do not promise permanent attribution or complete takedown.

Read next

Continue through the same learning path with the next practical articles.