Last year's photo does not prove what happened this afternoon

An account posts a photo of a flooded street with the caption: “Central Hanoi this afternoon. This road has just been closed completely.” The editors constructed this situation. It is not a real incident. The useful question is not whether the picture looks machine-made. It is whether the caption is borrowing an older frame to tell today’s story.

If an older page already used the same frame for a different flood, “this afternoon” no longer holds. The picture can still be a real photograph. What fails is the time and place the caption assigns to it.

This piece is for someone about to tap share: a parent, a student, an office worker. The only result worth writing down is one sentence: share it with the source, correct the post, or hold it. The steps below have not been run on one fixed photo, one computer, and one region.

What an earlier appearance does to this caption

Google describes a way to learn more about an image in Search, and it says the feature is currently available only in some regions. The path in Google’s documentation is: search for the image, open the image viewer, choose About, then choose Learn more about an image. Labels in another language’s interface may differ.

When the data exists, that screen can show when Google has seen the image, or a similar-looking image; other pages that have used the image or a similar one; and pages that appeared much earlier than the rest of the results. The section on how the image was created or edited, when it is present, may say the image was captured by a camera or created with artificial intelligence (AI), along with a line naming who confirmed that detail. Google also says metadata about the source, and about the kind of digital source, can be changed by tools that create or edit images, so it is not necessarily accurate.

For the flood example, you are not looking for a score that says “fake image.” You are looking for an older page that attached the same frame to a different flood. In the constructed example, the early result leads to an article from two years ago about a different city. The signs and the road surface match that older article. “Central Hanoi this afternoon” then contradicts the earlier appearance. You do not need to decide who generated the picture.

Finding nothing does not prove the image is new and accurate either. Google says the information is based on what its systems have seen, including similar images, not on a certified unique copy.

One contradiction is enough to stop the share

Reuters, in an article dated 28 August 2026 on visual verification, describes a group of journalists checking images and video taken from social media. They ask whether the image is an authentic, unaltered version of the event it is said to show, at the stated time and place. The article says they may look at hundreds of images and videos a day and confirm, then publish, only about a dozen. Metadata, when it is still in the file, can help confirm where and when the image was created, and with what device. They compare what the image shows with other public information.

A reader does not have to do that newsroom’s whole job. For the caption in this example, the earlier appearance is already a contradiction. If the event is traffic happening now, the further check is a notice from the public body responsible for that road. It is not a chatbot’s answer to “is this photo real?”

One matching sign, such as the right kind of weather, still does not replace the earlier appearance. Reuters describes verification as a comparison of time, place, and public information, not as a trick of looking at shadows.

A provenance signature describes the file, not the caption

C2PA’s explainer says Content Credentials are a signed provenance record attached to a file: how the image was created, which tools it passed through, and whether any step was recorded as done by AI. The same document says this is not a judgment of whether the image or the caption is true. The signature tells you whether the provenance portion is intact and whether the signer is on a trusted list. It does not prove that the scene is central Hanoi this afternoon.

The provenance chain is not always complete. C2PA gives this example: crop an image with a tool that does not support the standard, and the crop may not be recorded. Missing Content Credentials do not mean the image is fake. Attaching provenance is optional, and the document says not to treat every file without this mark as less trustworthy.

An icon or verification mark, when an app that supports this standard displays it, means the app checked the provenance record attached to the file. It does not turn the poster’s caption into an event that happened. The project publishes this technology at contentcredentials.org. This article does not describe the result of uploading a file there, because that step was not used as evidence.

Send an outside tool only images that are already public, or that you have the right to use. Do not upload photos of children, identity documents, a home address, or health records just to “see the metadata.”

Write one decision, then stop

For the flood example, the fitting sentence is: the current caption does not match an earlier appearance, so do not share it. If you already shared it, correct the post with that older page. Do not correct it with a verdict that “AI drew the photo.”

The three results differ only in the evidence you actually have.

An earlier appearance and an independent notice both match the caption: you may share it, with a link to that source. A feeling that “the photo looks real” is not part of the share.

An earlier appearance contradicts the time or place in the caption: do not share it. A real photograph can still be attached to the wrong story.

All you have is a screenshot, with no older page and no notice from the relevant authority: hold it. That is a result. It is not a failed check.

If the caption is long, split it yourself into the event, the place, and the time before you search. A tool that writes the search sentence may be used only on text that is already public. If it invents a link, or a witness name that is not in the caption, discard that sentence.

For an urgent warning, use the channel of the body responsible for that matter. Do not spread the image to ask strangers when the action is to read the official notice.

These ten minutes are not for winning an argument about whether a machine drew the picture or a camera took it. They are for refusing to let an old frame become evidence for the story the caption just assigned to it.

Sources

1. Google Search Help — Learn more about an image — available only in some regions; can show when Google has seen the image or a similar one, other pages that used it, an early appearance, and creation or edit information when present; metadata can be altered.
2. Reuters — What is visual verification? — 28 August 2026; a social-image desk checks whether a picture matches the stated event, time, and place; metadata still in the file may show place, time, and device; they compare it with public information.
3. Content Credentials — the project’s public site; this article does not treat an upload to that site as evidence.
4. C2PA — Content Credentials Explainer — a provenance signature does not judge whether the image or the caption is true; the chain can be incomplete; no signature does not mean the image is fake.