Amnesty International published AI-generated images of police violence, then deleted them

What happened

In spring 2023, marking the second anniversary of Colombia’s 2021 National Strike protests, during which police committed grave abuses, an Amnesty International account posted photorealistic AI-generated images of protest and police violence on social media. They carried a small AI disclaimer but were made to be read as documentary photographs.

Amnesty’s reason was protection: real photographs would have made identifiable people traceable by a state that had just been violent towards them. That reason is genuine and the problem behind it is unsolved.

Photographers spotted the fabrication almost immediately. The Colombian flag’s colours were in the wrong order, faces unnaturally smoothed, police uniforms out of date. Amnesty deleted the images within days and its Americas director publicly acknowledged the criticism.

The artist Molly Crabapple:

Either use the work of brave photojournalists, or use actual illustrations. AI-generated photos just undermine trust in your findings.

Showing a beating, a detention, a displaced family without making anyone identifiable to the security services is a daily problem in a five-person Baghdad office. What transfers is the checklist Amnesty skipped.

Five steps before you generate an image of a real event

1. Write the problem in one sentence. Amnesty’s was not “we need an image.” It was “we must not make identifiable people traceable.” Written down, that sentence usually rules out generating a photograph.
2. Exhaust the non-synthetic options against it. All free: crop or shoot from behind; blur or black-bar faces; silhouette; photograph the place with nobody in it; a hand drawing or a simple vector illustration; a properly licensed archive or agency photograph; or a caption saying you are not publishing an image because it would expose the people in it. That last one costs nothing and is itself a message.
3. Apply the documentary test. If the image will be read as a photograph of a real event that happened to real people, do not generate it. Generated illustration is defensible for the abstract: a metaphor, a header, a diagram. It is indefensible for the specific and the evidentiary.
4. Assume the label will be stripped. Amnesty attached a disclaimer and it did not save them. Screenshot an image and the caption is gone. The test is never “did we label it” but “is this image safe once it is separated from everything we wrote around it.” That is our reading, not a reported finding.
5. Show it to someone with local knowledge first. One Colombian would have caught every tell in thirty seconds. Your equivalents: uniform insignia, licence plates, Arabic and Kurdish signage, architecture, street furniture. Exactly what generators get wrong, and what a local audience notices first. A garbled Arabic sign in an image of Baghdad is a reversed Colombian flag.

What it cost

A synthetic picture of an atrocity hands every denialist a permanent line of attack on every genuine photograph your organisation has ever published, including work from years ago. It also takes paid work from the local photojournalists who took real risks to document the event.

Amnesty removed the images within days rather than defending them, which limited the damage. Its public framing leaned on not letting the row distract from the core message. Move fast, but write a removal statement that concedes the error. One that does not reads as evasive to the people you were trying to convince.