← All guides Journalism

AI images in election misinformation

Most misleading political images are real photographs with false captions. Where detection helps, where it actively misleads, and what to do in the hours that matter.

· 10 min read · Best AI Image Detector

Check the caption before the pixels. The dominant form of political image misinformation is a genuine, unedited photograph presented with a false date, place or context, and it passes every detector because nothing about it is synthetic.

The three categories, by how common they are

Sorting them first prevents the most common analytical mistake, which is running a detector on something no detector can address and taking the clean result as a verdict.

  1. Real photograph, false context. The dominant category by a wide margin. A genuine image from another country or another year, captioned as today.
  2. Real photograph, altered. A crowd made larger, a sign changed, a person removed or added. Detectable, and the region map is what shows it.
  3. Wholly generated. A scene that never happened. The most discussed and the least common, though it is growing.

The first category is invisible to pixel analysis by construction. The photograph is genuine, was taken by a camera, and carries all the statistical properties of a photograph, because that is exactly what it is.

This is the single most important thing for anyone doing election verification to internalise. A clean detection result on a misleading political image is the expected outcome, and reporting it as evidence of authenticity is how a checker becomes part of the problem.

Reverse search Pixel check Region map
Real photo, false caption Yes Finds the original No Reads as genuine No
Crowd or sign altered Partly May find the unaltered one Partly Score stays low Yes Localised heat
Person inserted or removed Partly Sometimes No Diluted by the frame Yes This is its case
Wholly generated scene Partly No earlier copies Yes Its best case Yes Flat and hot
Which tool reaches which category. The most common one sits in the left column.

Working in the hours that matter

Political image claims move fastest in the period when verification is hardest: the final days of a campaign, the night of a vote, and the hours after a count. Speed pressure is part of the design.

A sequence helps more than any single tool, because it front-loads the checks that are fast and conclusive and leaves the slow ambiguous ones for last.

  1. Reverse image search first

    Ten seconds, and it settles the largest category outright by finding an earlier copy with the true caption.

  2. Check what is in the frame

    Signage, language, vehicle plates, season, architecture. A photograph captioned as one country showing the road markings of another has answered itself.

  3. Run a pixel check and read the map

    This is where an altered crowd or a changed placard appears, and where the headline score will understate it.

  4. Find the earliest publication

    Who posted it first, when, and what did they say it was. Provenance beats analysis whenever it is available.

  5. Publish the uncertainty

    Where the answer is not settled, saying so is better than an implied verdict in either direction.

The fifth step is the one that separates verification from advocacy. An honest inconclusive is more useful to a reader than a confident wrong answer, and considerably easier to correct later.

Why a detector result is risky to publish

A number carries an authority that its error rate does not support. Publishing that an image scored 71 invites readers to treat it as a measurement of truth, and the qualifications get lost in the sharing.

It is also asymmetrically dangerous in political contexts. Wrongly calling a genuine photograph fake hands a real grievance to whoever posted it, and a correction never reaches the audience the original did.

The practical rule most newsrooms have settled on is to use detection as an internal triage signal, and to publish only findings that a person can explain independently of the score: the earlier copy, the mismatched signage, the altered region visible when pointed out.

What newsrooms are doing differently now

The practical response in most verification desks has been to move detection later in the process rather than earlier. Provenance work comes first, because it produces findings that can be published, and analysis comes last, because it produces numbers that cannot.

That inversion matters. A checker who runs a detector first anchors on its answer, then looks for evidence supporting it. A checker who searches first arrives at the detector already knowing what the picture claims to be, which is the only context in which a score means anything.

Several desks have also stopped publishing scores entirely, while continuing to use them internally. The reasoning is that a number invites a reader to treat verification as a measurement, and the qualifications that make it honest do not survive being shared.

What to do when you cannot resolve it

A meaningful share of political images cannot be settled within a useful timeframe. The photograph is degraded, the earliest copy is unfindable, the score sits in the middle, and the deadline is now.

Saying so is a legitimate output. A note explaining what was checked, what was found and what remains unknown is more useful to a reader than an implied verdict, and it is far easier to update when better information arrives.

It also protects against the failure mode that does the most lasting damage in political coverage, which is a confident call in the wrong direction. A correction never reaches the audience the original claim did, and the error becomes evidence for the next argument.

Where a claim cannot be resolved and is spreading, the more productive move is usually to report on the claim rather than to adjudicate the image: who is circulating it, when it appeared, and what it is being used to argue.

One habit is worth adopting whatever your deadline. Write down what the image is claimed to show before running any check, then verify that claim rather than the picture. It is a small discipline and it prevents the most common analytical error in this field.

Questions people ask

Are AI images the main election misinformation problem?
No. Genuine photographs with false captions remain far more common, more shareable and harder to debunk, because there is nothing synthetic to find. A detector returns a clean result on them, correctly, which is why a reverse image search has to come first.
What does a detector actually catch here?
Two things well: wholly generated scenes, and localised alterations such as an enlarged crowd, a changed placard or an inserted figure. The second shows on the region map rather than in the score, which stays low because most of the frame is a real photograph.
Can I publish a detection score in a fact check?
It is unwise on its own. A number implies a precision the error rate does not support, and a wrong call on political imagery hands a grievance to whoever posted it. Publish findings a reader can verify independently, and use the score to direct where you look.
How do I check a photo quickly during a live event?
Reverse image search first, then read what is in the frame: signage, language, plates, season, architecture. Both are faster than any analysis and both reach the largest category. Run the pixel check third, for the altered and generated cases.
What is the liar dividend?
The effect where genuine evidence gets dismissed as synthetic simply because synthetic media exists. It is the second-order harm of this whole field, and it means a detector that confidently mislabels real photographs does damage well beyond the individual case.
Should platforms label political images automatically?
Labelling based on markers inside a file is reliable when it fires and silent the rest of the time, because most images carry no marker. Labelling based on a detector score at scale would apply a published error rate to millions of images, which is a different proposition entirely.