Skip to main content

Systems and Workflows

AI Media Verification: Build a Review Beyond One Detector

A proposed media review workflow that keeps the file, the verification evidence and the delivery decision connected.

Delivered media travels with its attached context notes; the authentic Google symbol identifies SynthID’s publisher.
On this page
  1. Begin with the file you were actually given
  2. Record the asset before changing it
  3. Choose a check that covers the source
  4. Preserve what each result actually says
  5. Compare origin evidence with the promised content
  6. Keep the review connected during edits
  7. Make the decision understandable to the next person
  8. An unresolved result needs a proportionate response
  9. Check the workflow on one permitted asset
  10. Deliver the media and its usable context

Keep the review attached to the media you deliver. A detector result belongs to a particular file, and the next person needs to know which one.

Keep three things connected: the original media, the specific evidence collected about it and the decision about its intended use. That gives an editor something more useful than a checkbox saying “AI checked.” They can see the permitted use, the open question or the reason to choose a different asset.

Google's October 7, 2026 SynthID Detector announcement makes this a timely job. Google says the portal is available globally in English for image, video and audio checks involving Google and participating partners. Access to a check creates an opportunity to improve the handoff around it.

The workflow below is a proposed agency process, not a report of detector testing or a client deployment. It follows a hypothetical campaign image from receipt to delivery. You can adapt the same record structure to supported audio or video while checking the chosen service's current conditions. No supplied file has been uploaded for this article.

Begin with the file you were actually given

The first review question is practical: what arrived, and what is the team being asked to do with it? Answer that before opening an image editor.

Suppose a client sends a landscape image for a launch campaign. They describe it as a concept for a product announcement. The designer needs to crop it for a square post, add a headline and export a smaller copy. The account manager needs to know whether the finished creative can go into the launch calendar.

Each person starts with different information. The designer has pixels. The account manager has a brief. A reviewer might receive only a screenshot pasted into a chat. Put the original attachment and the client's description in the same project record so those versions do not become the only account of what was supplied.

Ask what the image is meant to represent. Is it decorative artwork, a proposed product concept or a photograph of the released product? Write that distinction beside the asset. It will determine what evidence matters later, even if the detector response is identical.

Also establish whether the file may be uploaded to the selected verification service. Use the client's instructions and your team's existing handling rules to make that decision. This article does not establish a service's retention policy or authorize uploading confidential material. If permission is unclear, preserve the file and ask the person responsible rather than improvising a public upload.

Finish intake with a named asset, a stated use and an owner for the next question.

A supplied media sheet keeps its source-note tag while a separate working copy is made. View image detail

Choose Actual size to read the graphic closely.

Record the asset before changing it

Preserve the supplied file before producing a working copy. Record its received filename, receipt date, sender and source description under one project asset identifier. Later exports can use different filenames while remaining linked to that record.

For the hypothetical campaign, call the original asset CAM-01. The working crop becomes CAM-01-square-v1, and the delivery export becomes CAM-01-square-final. These are proposed naming examples, not fields in SynthID or a required standard. Their purpose is to show which copies belong together.

Store the original separately from the editable project and the delivery folder. A teammate should be able to open the exact received attachment without deciding which file called “final” is newest. Do not overwrite that attachment when you resize it. If the client replaces it, retain the relationship between the replacement and the earlier version.

An optional file hash can help identify whether two stored files have the same bytes. A hash is a file comparison aid; it does not establish the creator or prove that the depicted product exists. Keep it if your team can produce and understand it consistently. Keep the hash beside the file reference and explain what it is being used to compare. It should not become a mysterious approval code pasted into every task.

Separate what the sender declared from what the reviewer established. “Client says generated in a named tool” belongs in the source-description field. It does not become a verified finding merely because it sits in a database. Preserve a relevant export record or message when available, and record a missing source detail as a question.

When a later export arrives, identify which reviewed version it replaces. The source conversation stays connected to that asset.

Earlier file versions remain visible in a stack marked with a history clock. View image detail

Choose Actual size to read the graphic closely.

Choose a check that covers the source

Choose the verification surface by its documented coverage. A shared technology name does not make every interface interchangeable. As checked on October 8, 2026, these three surfaces have different scopes:

  • Google SynthID Detector portal: Documented scope relevant to this workflow: Google announces checks for images, video and audio from Google and participating partners.
  • Verification in Gemini Apps: Documented scope relevant to this workflow: The current Gemini help page says its SynthID recognition is currently limited to Google AI content. Content Credentials provide a separate origin and history check.
  • OpenAI Content Provenance API: Documented scope relevant to this workflow: The current API guide covers images and audio containing supported OpenAI signals. It is not a detector for every AI provider.

A client saying “made with AI” has not yet supplied enough information to select a provider-specific check. Ask for the originating product and export when that affects the review. If the source remains unknown, record why a particular check was chosen and what its scope leaves unanswered.

Check access and upload conditions at the surface you will actually use. A format accepted in one app does not prove another service will accept it. A limit documented for Gemini Apps does not become the SynthID portal's limit. Use the current help for that service rather than copying a tutorial's settings into the project plan.

For CAM-01, the useful question is whether an appropriate check can add origin evidence for the original image. It is not whether the team can run the largest number of detectors. Write the question before selecting the check: does this service recognize a supported origin signal in this file? Keep the answer alongside the service’s scope.

If the required service is unavailable, write that down. The asset can remain waiting for a check while the designer prepares a permitted working copy. Keep those statuses distinct so progress on the creative does not silently become permission to publish it.

A media file waits beside a paused checking arch while a separate crop copy is edited. View image detail

Choose Actual size to read the graphic closely.

Preserve what each result actually says

Retain the service name, check date, submitted file reference and actual returned evidence. If the service exposes a downloadable result, preserve it. If the result is visible only on screen, record the relevant wording with enough context to identify the file and check.

OpenAI's guide says to read applicable result entries independently. Images include C2PA and SynthID entries, while audio includes SynthID; the response has no combined top-level outcome. An omitted check is different from a negative result. A not_detected result does not establish human origin. A reported signal also has a narrower meaning than a complete account of the file's history.

In the project record, keep the returned evidence separate from your interpretation. For a proposed example, a reviewer could record a supported signal and then write, “Source description is consistent with an AI-assisted campaign concept; product appearance remains unconfirmed.” That sentence describes a review decision. It must never be presented as wording returned by the detector.

Do not simplify an unavailable check into “no AI found.” Likewise, do not convert a missing value into a guessed creator, prompt or date. OpenAI explicitly cautions against inferring a prompt, account or individual creator from its result. A record with a blank field and a clear explanation gives the next person a better starting point than an invented answer.

Avoid a generic “verified” badge that hides the question asked. “Origin check completed” and “approved for this campaign use” describe different events. Link the completed origin check to its returned evidence. Link the use decision to the reviewer and the creative brief.

Keep discrepancies visible in the record. If the sender's description and the returned evidence point in different directions, retain both and ask the source owner to explain the discrepancy. Do not delete the inconvenient detail to make the task status look tidy.

Two different evidence slips stay together on a ring beside a separate reviewer note. View image detail

Choose Actual size to read the graphic closely.

Compare origin evidence with the promised content

Review the image beside the actual creative brief. An origin signal can help explain how media was produced. Check what the image and accompanying copy tell the audience before deciding its use.

In the CAM-01 example, a clearly presented product concept can be a suitable use of generated artwork if it matches the agreed brief. The same image beside copy claiming “the product shipping today” raises a different question. The reviewer needs evidence of the actual product's appearance, not merely evidence about the file's production method.

Separate those questions explicitly. Ask the product owner to confirm the depicted version. Ask the content owner to examine the caption and surrounding copy. If the image shows a person apparently endorsing the product, resolve that representation through appropriate source and permission evidence. A detector response cannot supply an endorsement agreement.

NIST's 2024 technical overview distinguishes approaches to provenance and synthetic-content detection and describes watermark detectors as having nonzero probabilities of false positives and false negatives. It supports using these techniques as parts of a broader review. It does not provide an independent performance result for this newly expanded portal.

Use that context to identify the missing evidence, rather than treating uncertainty as a reason to stop every creative project. For a decorative illustration, the outstanding question might be whether the agreed use permits generated artwork. For a reported event photograph, it might be whether the scene can be connected to the event being described.

This is a proposed editorial decision process, not a legal clearance method. Keep rights, factual representation and origin evidence as separate responsibilities. The account manager can route the missing product confirmation to the product owner. The designer does not need to resolve that question by guessing from the image.

A concept picture, a generic product object and a responsible person remain separate around a representation question. View image detail

Choose Actual size to read the graphic closely.

Keep the review connected during edits

Link the delivery copy to the original and note the meaningful changes. In the hypothetical campaign, that means recording the square crop, added headline and export version. The original review should remain attached to the original file.

OpenAI documents that editing and conversion can remove metadata or weaken signals. Gemini's help also warns that a watermark might become undetectable after many alterations. Neither statement establishes what happened to CAM-01. The practical response is to retain the earlier evidence and consider whether checking the final export would answer a useful new question.

Distinguish a packaging change from a change to the represented content. A smaller export needs a clear link to its source. Replacing the product, adding a person or changing the apparent setting also changes what the reviewer needs to assess. Send that revised representation back to the relevant content owner rather than inheriting the earlier decision.

Keep the transformation record short enough to maintain. “Square crop from original, headline added, final JPEG export” is useful. An automatic history containing dozens of routine saves can bury the change that actually matters. Retain the editable project separately if the team needs to reconstruct the edit.

If a final-file check produces different evidence, preserve the results for both versions. Do not overwrite the original result or decide that the difference proves deliberate manipulation. Compare the checked versions and the transformation record. If those do not explain the difference, keep the question open and ask for the source context that could resolve it.

For a composite, identify its contributing assets separately before reviewing the finished composition. A result associated with the delivered composite should not be assigned to every component without supporting evidence. Name the checked composition as its own version and preserve the references to its component assets.

Two component images feed one composite, whose own note stays attached only to that composite. View image detail

Choose Actual size to read the graphic closely.

Make the decision understandable to the next person

Build a small asset record that answers the questions a delivery reviewer will actually have. The following fields form a proposed structure you can adapt to a shared document or project workspace:

  • File: asset identifier, original location and current delivery version.
  • Declared source: sender's description and any supplied export context.
  • Verification: service, check date, file checked and retained result.
  • Review question: the intended representation and any unresolved evidence.
  • Decision: reviewer, permitted use or hold, and the reason.
  • Handoff: recipient, delivery location and what changes require another review.

Put the record where the creative task can link to it. The scheduler should reach the evidence from the deliverable, and the reviewer should reach the current file from the evidence. If your planning already lives in a workspace such as Rise's Content Creator Workspace, keep a linked review document beside the project. That product page describes content planning and reusable templates; it does not claim a built-in detector integration.

Write the decision so another person can act on it. For the hypothetical campaign: “Use this version as concept artwork in the agreed launch teaser. Product-page photography remains a separate deliverable.” The restriction gives the scheduler a usable boundary. “Looks good” does not.

Assign open questions to people. The designer can explain the edit, the client can clarify the supplied source and the product owner can confirm the represented product. One named owner per question makes the record easier to advance.

Also keep the authority to change the decision clear. A completed design task should not erase a hold recorded by the content reviewer. If the planned channel or caption changes, the person responsible for that new use should revisit the decision. A decision about a teaser should remain attached to the teaser, even when another team requests the same image for a product page.

Two hands pass a folder while a separate open note preserves a question about its use. View image detail

Choose Actual size to read the graphic closely.

An unresolved result needs a proportionate response

Choose the response according to the unresolved question and the consequence of using the asset. There is no useful rule that every uncertain image must be rejected or that every undetected signal permits publication.

Consider two hypothetical files arriving on the same day. One is an abstract background for a webinar. The other is presented as a photograph of a company's new facility. Both have incomplete source descriptions. The background needs an agreed creative-use decision. The facility image also needs evidence that it depicts the actual facility.

For the second file, ask for the original photograph, capture context or a source contact who can substantiate the representation. If that evidence is unavailable, change the creative or hold the factual use. Do not solve the gap by relabeling the image “human-made.” The underlying claim about the facility would still be unresolved.

For the background, a source clarification and the team's ordinary rights and creative review might resolve the practical question. The decision can stay narrow: accepted as decorative artwork for the named webinar. It does not need to establish every step in the file's production history.

Set a concrete next action for a hold. “Waiting for client confirmation of product version” tells the team what can release the task. “Needs more verification” leaves them guessing whether to run another detector, contact the client or replace the asset.

Repeated checks should have a purpose. If the remaining question is who authorized an endorsement, running the same file through another origin check will not obtain that permission. Ask the endorsement owner for the relevant permission evidence, or choose creative that does not make the endorsement claim. That is how the workflow supports delivery rather than becoming another queue of tasks nobody can finish.

A question about promotional media goes to a person holding an open note, with an origin-check lens set aside. View image detail

Choose Actual size to read the graphic closely.

Check the workflow on one permitted asset

Try the process with one asset your team is allowed to handle and submit. This is a proposed exercise to evaluate the handoff, not a detector benchmark. Its success condition is whether the next person can follow the evidence and decision.

Start by preserving the original and creating its review record. Write the intended use before running the selected check. Retain the actual result and explain the specific question it answers. Then make an ordinary delivery edit, save the final copy separately and record what changed.

Ask a teammate who did not perform the review to locate the original, identify the file checked and explain why the delivery version is accepted or held. Let them use the record without a verbal walkthrough. Missing information in this exercise becomes a concrete repair to the process.

For example, they might find the detector response but be unable to tell whether it describes the original or the cropped export. Add that relationship to the record. They might understand the origin evidence but not the approved channel. Add the intended use. These are hypothetical observations to look for, not results from a test performed for this article.

Avoid making the record so elaborate that nobody updates it after a last-minute edit. If a field never affects the handoff, reconsider it. If an omitted detail forces another person to ask the original reviewer, consider adding it.

Use the exercise to decide which clerical steps are stable enough to automate. A file link must resolve, an asset identifier must stay attached to the right record, and an open question must reach its assigned owner. Creating a record, attaching a file reference or routing an open question can be useful automation candidates. Approval of a consequential use still needs a defined decision and responsible owner. No integration or measured time saving is asserted here.

Deliver the media and its usable context

Deliver the final creative with the version reference and a concise explanation of the review. The recipient should know which file to use, which evidence was collected and whether any limitation affects the planned publication.

For CAM-01, the handoff could identify the approved teaser export, link its original and review record, and state that it is concept artwork rather than the product-page photograph. The scheduler gets the file and the instruction together.

Keep detailed evidence available to the people who need it. A public caption does not require a dump of internal notes, and a delivery message should not expose confidential client material. Decide which context belongs with the recipient and which remains in the project record.

The recipient needs a usable decision, not a folder of unexplained screenshots. Put the delivery version and its permitted use first. Keep the original, check evidence and edit record one link away so a changed request can be reviewed from the same starting information.

Choose one upcoming asset and try that complete path. Judge the proposed workflow by whether another person can make the next decision from the record. Start with the version the recipient will publish, then follow its links back to the original and the reviewer’s decision.

Checked for this article

Sources

  1. Global public SynthID Detector announcement
  2. Verify AI-generated images, videos, and audio
  3. Content provenance
  4. Reducing Risks Posed by Synthetic Content: An Overview of Technical Approaches to Digital Content Transparency

Keep reading

All articles