how to tell if a news photo is AI-generated

fake news photos are already a documented problem

they already move money and shape how a disaster is read. on may 22, 2023 a fake, likely AI-generated image of an explosion near the Pentagon spread through ‘verified’ blue-check accounts, one of them posing as Bloomberg News, and briefly pulled the S&P 500 down about 0.3% to a session low before the hoax was exposed and stocks rebounded (CNN Business). after Hurricane Helene in october 2024, an AI image of a crying girl clutching a puppy on a rescue boat went viral and was shared by political figures to attack the federal storm response, until fact-checkers ruled it AI-generated (PolitiFact). audiences feel the ground shifting. in the Reuters Institute’s june 2025 survey across 48 markets, 58% of people said they still worry about telling real from fake in online news.

the visual tells, and why they keep fading

start with the picture, then keep going. generative images still slip in ways you can sometimes catch by eye:

• →details that shift between versions. the same scene reshared with a different dog, vest or background. models struggle to hold details steady, which is exactly what tripped the Hurricane Helene images.
• →physically impossible joins. fences that melt into crowd barriers, warped architecture, hands or signage that never quite resolve.
• →text and logos. invented letters, garbled captions and brand marks that dissolve up close, one of the more durable giveaways.
• →skin and texture. an over-smooth, waxy finish, or hair and pores that look painted on.
• →light that doesn’t add up. shadows pointing the wrong way, reflections that don’t match the scene around them.

each model release trains these tells away. a clean image no longer proves a human made it, and a real photo can carry a weird artifact too. that is why the eye alone settles nothing, and why detectors read statistical fingerprints in the pixels you can’t see.

check the source before you trust the image

the strongest tell usually isn’t in the pixels. it’s whether the moment exists anywhere else. a dramatic ‘breaking’ image with no other photos, videos or first-hand witnesses from the same scene is a warning sign, which is what investigators flagged on the Pentagon hoax. run a reverse-image search and a timeline check: does the picture have an earlier credible appearance, or does it surface only on partisan and anonymous accounts during a charged event?

watch for source laundering too. images that began life as AI ‘stock’ or as openly-labelled memes get stripped of that label when they’re reshared as real breaking news. and look across copies, the way fact-checkers did with the Helene image: the same ‘girl’ appeared with a different dog and a different life vest from one version to the next, a tell no single frame would show you.

look for provenance, the signed paper trail

real newsrooms now pair verification with cryptographic provenance. the Associated Press’s 2023 standards forbid using generative AI to add or remove elements from its photos, and say AP will refrain from transmitting AI images suspected or proven to be false depictions of reality (Poynter). AP, Reuters, AFP and the BBC also back C2PA Content Credentials, a tamper-evident ‘nutrition label’ that records who made a file, when, and with which tools (Content Authenticity Initiative). wire services have begun piloting signed capture in the field, straight from the camera.

provenance is strong evidence when it’s there, and silent when it isn’t. adoption is partial and opt-in, so most images circulating online still carry no credential at all. a missing manifest is not proof of fakery, only a gap where verification has to do the work instead.

watch the inverse trap: the ‘liar’s dividend’

a confident ‘it’s AI’ claim deserves the same scrutiny as the image. as fakes spread, real footage gets waved away as fabricated. in august 2024 a candidate falsely claimed a genuine photo of a 15,000-person rally was AI-generated; image-forensics expert Hany Farid found no evidence of AI or manipulation (PolitiFact). absence of provenance is not proof of fakery, and a ‘no AI detected’ result is not a clean bill of health. the risk runs in both directions, and it’s uneven, so steer clear of certainty on either side until corroboration and provenance line up.

what amige. does with a suspect news photo

it gives you a fast first signal that should send you to verify. it does not hand you a verdict to quote. amige. routes each scan to the detectors strongest for that kind of image, runs a panel of independent detectors built by different teams, and shows you where they agree and where they split. it names the maker as a best guess, like ‘looks like a diffusion-model image,’ always carrying the question mark. when the reads conflict it returns ‘uncertain’ rather than forcing a call, and it recalibrates as new generators ship.

treat the output the way the Reuters Institute frames detection: ‘a starting point as part of a verification process,’ a probability you weigh next to corroboration and provenance, never the last word. you can see how the routing and fusion actually work on the machine.

where photo detection falls short

the honest move is to name the limits, because they’re real:

• →screenshots and re-compression erode the signal. cropping, re-saving and screenshotting strip metadata and weaken classifier signals, one of the ways detection gets defeated (Reuters Institute). scan the original file when you have it.
• →a negative result proves nothing. the Reuters Institute puts it plainly: ‘even when an AI-detection tool does not identify any signs of AI, this does not necessarily mean the content is not synthetic.’
• →false positives happen. a real photo can read as AI after heavy editing or upscaling, so one number can’t carry a serious decision on its own.
• →your eyes are not a detector. the visual tells keep fading with each model release, and a picture that looks clean settles nothing by itself.

that’s why amige. shows a range and a panel instead of a single verdict, and why a news photo still deserves corroboration and provenance before you act on it. read the estimate as the start of the question, never the end of it.