Where an AI check fits in newsroom verification

Source the image first, because provenance settles more cases than pixels. Run the detector as a second opinion, and treat an uncertain result as a reason not to publish rather than a reason to hedge.

The question a detector does not answer

The failure that damages a newsroom is almost never a generated image. It is a genuine photograph from the wrong war, the wrong year or the wrong country, republished with a new caption.

A detector will clear that image, correctly, because it is a real photograph. Everything about the claim attached to it is false. No pixel analysis touches this, and it is by a wide margin the more common error.

The sequence

Establish who took it and when
A name, a device, a time and a place. Where a wire service or a known stringer supplied it, most of this arrives with the file. Where a member of the public supplied it, ask directly and ask early.
Run reverse image search
The earliest appearance is the single most useful fact available. An image indexed three years ago cannot show yesterday's event, and this catches recontextualised photographs that no detector will.
Read the file's own record
Content Credentials, where present, are cryptographic and outrank everything else here. Camera metadata, where it survives, gives you device, lens and time to cross-check against the account.
Score the pixels
Now, and not before. A high score on a photograph you have already sourced means you go back to the source with a specific question rather than abandoning the image.
Read the region map
One hot tile inside a genuine frame is the case that matters most in news: a real photograph with an object added or removed. That is the manipulation that gets published.
Decide, and record why
Write down what you checked and what it returned. A correction six months later needs to show the process, not the conclusion.

Why upload matters for a picture desk

An unpublished image from a source is one of the more sensitive files a newsroom handles. Sending it to a third-party API before publication puts it on somebody else's server, with logs, retention and a jurisdiction attached.

For a photograph from a source in a hostile environment, that is a safety question rather than a policy one. A tool that scores locally removes the issue: there is no request carrying the image, which anyone can confirm in the network tab.

What to do with an uncertain result

The middle band is common on news imagery, because news imagery arrives compressed, cropped and forwarded. An uncertain score is information, and the correct response is to go back to sourcing rather than to hedge in the caption.

Publish on sourcing 0–20
Publish on sourcing 20–45
Go back to the source 45–65
Do not publish yet 65–90
Do not publish 90–100

The five bands and what a picture desk should do with each. The middle band is not a weak yes.

Hedged captions are the worst outcome available. A phrase like unverified imagery gives the picture the reach of a published photograph and the newsroom none of the responsibility, and readers do not distinguish the two.

What to publish alongside a checked image

Who supplied it, at whatever level of detail safety allows.
When and where it was taken, and how that was confirmed.
What verification was done, in one sentence rather than a methodology note.
What remains unconfirmed, stated plainly rather than implied by omission.
Whether the image was altered for publication, including cropping and colour.

Readers have become reasonably good at spotting when a newsroom is hiding uncertainty. Saying what was checked, including that a detector was run, tends to build more confidence than a silent byline.

Building it into the desk, not the individual

Verification that depends on one careful person fails the week that person is on holiday. The newsrooms that handle this well have made it a step in the workflow rather than a habit of the picture editor.

That means a field in the content management system recording what was checked, a rule about which images require it, and a default that an unchecked image does not publish. None of that is technology, and all of it survives staff turnover.

Decide the threshold by risk rather than by volume. A wire photograph from a named agency needs less scrutiny than an image sent to a tip line by an anonymous account during a breaking story, which is precisely when the pressure to publish is highest.

Run the checks before the story is written where you can. A photograph that fails verification after the piece is built creates pressure to publish it anyway, and that pressure is how bad decisions get made on deadline.

Questions people ask

Should a newsroom rely on AI detection?

As a second opinion, yes. As a primary check, no. The error most likely to embarrass a newsroom is a genuine photograph with a false caption, which every detector will clear. Sourcing catches that and pixel analysis does not, so the order of the checks matters more than the tools.

What if a source will not provide the original file?

Treat that as information rather than an obstacle. There are legitimate reasons, including safety, and there are illegitimate ones. Weigh the refusal alongside everything else you know about the source, and be more cautious about publishing than you would be with a file in hand.

Can we cite a detector result in a correction?

Describe the process rather than the number. Saying the image did not survive verification is accurate and defensible. Saying it scored 87 out of 100 invites readers to treat a likelihood as a measurement, and invites a complaint you cannot answer.

How do we handle images from social media?

Assume every signal has been degraded. Platform recompression strips metadata and weakens the pixel signal, so both provenance and detection are working with less. Reverse search becomes the strongest tool available, and finding the original poster matters more than any score.

Do Content Credentials help yet?

Where they exist, substantially. A valid manifest is cryptographic rather than statistical and is the strongest single piece of evidence a file can carry. Coverage is still thin, so treat a credential as a bonus when present and never treat its absence as suspicious.

Who in the newsroom should run these checks?

Whoever already owns picture verification, usually the picture desk rather than the reporter. Splitting the check from the sourcing creates a gap where each side assumes the other did it, and that gap is where a wrong image reaches the page.