An AI image can look convincing while its caption makes a claim nobody checked. Before publishing, review the picture and the words together: what would a reader reasonably believe happened, exists, or is being offered? Then connect each important assertion to evidence outside the generated image.
Reviewed October 23, 2026. Hero artwork is an original AI-generated illustration. Worked cases are teaching examples; they are not customer results or model benchmarks.
1. Decide what the picture is supposed to establish
A mood illustration and a record of an event have different jobs. A generated workshop scene can illustrate a proposed class. It cannot establish that the class happened, who attended, or what the venue looks like. Make that decision before choosing an attractive candidate. Otherwise a realistic picture can silently become evidence for a story it was never qualified to tell.
Write one publication sentence: “This illustration accompanies an announcement about a planned pottery session.” Now list what must come from another source: the organizer's approved plan, booking conditions, date, capacity and location. If the piece reports an actual past event, obtain actual reporting material. Changing the prompt cannot supply missing reporting.
For a product listing, use the product-photo approval checklist to compare the exact item with its reference. This article handles the broader decision about what an image and caption jointly claim.
2. Work through a caption that overstates the evidence
Here is a fictional exercise. Maple Yard is planning an introductory pottery session. Its approved planning note says twelve seats are planned and bookings have not opened. It contains no accessibility assessment and no record of a previous session. These are invented facts for practice, not an advertisement for a real event.
Draft caption to reject: “Twenty guests enjoyed yesterday's fully accessible Maple Yard pottery workshop. Book your seat now.”
The sentence contains several different problems. Do not reduce them to a single “AI image looks wrong” flag. Split the caption into assertions, including the action it asks a reader to take.
| Claim | Evidence in the exercise | Decision |
|---|---|---|
| A session is planned | The approved brief describes a future introductory session. | Supported as a plan, not a completed event. |
| Twenty guests attended yesterday | No past-event record; the brief concerns twelve planned seats. | Remove the invented attendance and date. |
| The venue is fully accessible | No access information was supplied. | Hold this claim and request specific verified details. |
| Book now | The brief says bookings have not opened. | Remove the unavailable action. |
A supportable exercise caption is: “Illustration of a proposed pottery session at fictional Maple Yard. The planning brief provides for twelve seats; booking details are not yet available.” For an actual organization, replace the fictional premise with its approved announcement and verified details. Do not publish our practice copy as a real offer.
Download the worked claim ledger and blank review row. The useful result is a set of decisions someone can act on, not an impressive-looking confidence score.
3. Check what the image implies without saying
Read the image once with the caption hidden. Count people and objects only as observations about the picture, not as facts about an event. Look for recognizable locations, uniforms, logos, equipment, entrances and apparent demonstrations. Each can imply something beyond the written copy. A ramp in a generated scene does not establish a real entrance route; a person wearing a staff shirt does not establish an endorsement.
Then read the caption while covering the image. Notice whether it uses reporting language such as “shows,” “at yesterday's event,” or “our team.” Bring the two back together and ask what a reader would infer. A small illustration label may not resolve a much stronger misleading headline or image choice.
When an irrelevant invented detail carries a factual implication, choose another composition or use a clearly conceptual treatment. Do not erase a real uncertainty by generating a more plausible version. For example, an empty symbolic worktable is a better fit for a planning article than a realistic crowd presented as the organizer's actual attendees.
4. Match the evidence to the claim
Record the source, its date, the exact statement it supports and who can resolve a question. A current approved event brief may support planned capacity. It does not support actual attendance. A dated original photograph may document a scene, but the location and circumstances still need verification. A search snippet or another site's repetition of the same caption is a lead, not independent corroboration.
Keep supported, contradicted and unresolved separate. “Unresolved” means you have insufficient evidence; it does not mean the claim is false. Specify the next action: request the venue's access information, ask the organizer for the corrected date, or replace a past-event formulation with an accurate description of the plan. Assign that action before the publication deadline.
For a source that changes, retain the relevant dated record in the project. Link the record to the candidate filename and caption version. If the organizer changes capacity after approval, the old ledger tells you which placements need another review.
5. Use provenance for the question it can answer
C2PA's FAQ explains how recorded provenance can help describe an asset's origin and history. A valid credential is not a verdict that every depicted event or attached statement is true. Missing credentials are also not proof that an image is genuine or generated.
If the file has Content Credentials, preserve the original and inspect the record using a compatible verifier. Note what was actually reported. Our Content Credentials verification guide covers that file-level process. Here, put the finding beside the claim ledger rather than letting it replace the ledger.
An automated detector's score should not become your only reason to accuse a source or approve an image. The publication decision still needs evidence for the particular claim. “The creation record identifies an illustration” is a narrower finding than “everything in this image is false.”
6. Review the delivered image, caption and action
Check the downloaded export in the real article, social card or listing preview. Is the illustration label adjacent to the image? Does a crop remove context? Does the share caption reintroduce “yesterday” after the article was corrected? Is a booking or purchase button actually appropriate for the approved announcement?
Save the approved image filename, caption, destination, reviewer and open questions. If an important question remains, hold that version or narrow the publication to what is supported. A deadline does not convert an unknown into an approved fact. If a mistake is already public, correct every affected placement and keep a clear record of what changed.
You can prepare a clearly labeled illustration in Image Lab after defining this boundary. The generator supplies artwork; the source record and editorial review establish which statements you can publish. Check permissions separately using the commercial-use evidence checklist.
Frequently asked questions
Can an AI-generated image prove an event happened?
No. A generated illustration does not document an event. Use appropriate reporting evidence for dates, attendance, location and participants, and make the illustrative role clear in the placement and caption.
Does a valid Content Credential prove the caption is true?
No. A credential can provide recorded provenance information. The caption may still describe the wrong date, place or circumstances. Verify the claim separately against an appropriate source.
Should I label an unsupported claim false?
Not automatically. Use unresolved when evidence is insufficient, and contradicted when reliable evidence conflicts with it. Record the distinction and the next verification step instead of turning uncertainty into certainty.
What should the claim ledger contain?
Record the exact claim, source and source date, supported or unresolved finding, candidate file, caption version, reviewer, destination and next action. One row per claim makes corrections easier to trace.
Is the workshop example a real event investigation?
No. Maple Yard, its planning brief and the rejected caption are fictional teaching material. The hero is an AI-generated editorial illustration, not a photograph documenting a workshop or a tested detector result.
For a focused next step, read Content Credentials for AI Images: Verify the File, Then the Claim.