On this page
Triage: decide how sure you need to be
In the first minutes of a breaking story, pictures arrive faster than anyone can check them: from agencies, from social media, from readers, from people at the scene. Not every picture needs the same work, so start with two questions.
- What happens if it's wrong? A photo that will lead the homepage, identify a person or show violence needs full verification. A generic street scene used as illustration needs less, and a caption that says what it is.
- Who sent it? A staff photographer or a trusted agency is a very different starting point from an anonymous account that posted the picture a few minutes ago.
Then sort what you have into three piles: confirmed, being checked, and not usable yet. Nothing moves from the second pile to the first without passing the checks below, however fast the story moves.
Find the person who took it
The strongest evidence that a photo is real is the person who took it, reachable and able to describe the moment. Work backwards from the copy in front of you:
Trace the earliest copy
Run a reverse image search on the whole picture and on crops of distinctive details, and look for the oldest upload. Our reverse image search guide covers the engines and how to sort by date.
Check the account
How old is it? Does it usually post from that place? Has it posted other pictures of the same scene, before and after? A real witness usually has more than one frame.
Contact them
Ask what happened, where they stood and what they saw just before and after. Ask for the original file, for permission to use the photo and for how they want to be credited.
Ask for the neighbouring frames
People rarely take a single photo of a dramatic moment. The frames before and after, straight from their camera roll, are hard to invent and quick to send.
If nobody can be found, or the account appeared the same day and vanished after posting, treat the picture as unverified however convincing it looks.
Rule out recycled photos
Many of the misleading pictures in a breaking story are not AI at all. They are real photos from an earlier event, another country or a film, relabelled as new. A reverse image search catches most of them in seconds. Look for:
- an earlier copy of the same picture, from before the event happened;
- the same scene in a different story, such as last year's flood or a protest in another city;
- stills from films, games or adverts, which can pass for real footage once cropped;
- a real photo cropped to remove the context that would change its meaning.
A recycled photo will usually come back from Perspicax as likely real, because it is real. That is exactly why provenance comes first: a detector answers whether an image was generated, not whether it shows what the caption claims.
Check the place and the time
Geolocation and chronolocation, working out where and when a picture was taken, are the core skills of visual verification. They need patience more than special tools:
- Landmarks and signs: buildings, hills, bridges, shop fronts, street names, and the language and script on signs. Match them against satellite imagery and Street View.
- Street furniture: road markings, kerbs, lamp posts, number plates and police vehicles differ from country to country, and often from city to city.
- Shadows: their direction and length tell you roughly when the photo was taken. Sun-position tools such as SunCalc let you compare them with the claimed time and place.
- Weather: rain, snow, haze or clear skies can be checked against weather records for that day.
- Season: leaves on the trees, the quality of the light, what people are wearing.
AI images often get the general look of a place right and the specific details wrong: a skyline that is almost the city, signs in nearly the right script, a bridge with the wrong number of arches. A place you can match point by point is strong evidence. A place that is only "sort of" right is a warning.
Get the original file
Ask for the file as it came off the camera or phone, not a screenshot and not a copy from a messaging app. The original gives you more to work with:
- Camera metadata: the device, the time and sometimes the location. It supports a story but is easy to edit or strip, so it never proves one on its own.
- Full resolution, for zooming in on signs, faces and small details.
- Content Credentials, when the device or app adds them. Some newer phones sign their photos at capture, and a valid record from the camera is good evidence of where the file came from.
- Neighbouring frames and video from the same moment, which are hard to invent consistently.
Check it against everything else
A real event leaves many traces. Compare the picture with other photos and videos of the same moment, taken by other people from other angles, and be clear about which signals carry weight.
Still worth checking
- The same details from different angles: the same vehicles, banners, damage and clothing.
- Timing that fits: the light, the shadows and the weather match other footage from the same hour.
- Video from the scene whose sound and movement match what the still shows.
- Reports from people on the ground that describe what the picture shows.
No longer reliable
- Many outlets ran it: they may all have copied the same unverified post.
- An agency or channel logo on the image: logos can be pasted onto anything.
- A large, sharp, high-resolution file: AI tools produce those too.
- No visible mistakes: the newest generators rarely get hands, faces or text wrong.
- Camera details in the metadata: they can be edited in seconds.
Use a detector as one input
A detector reads what the eye can't, and it is fast enough to run on every picture in the queue. Use it like any other source: as evidence to weigh, not as the verdict. Perspicax combines proven computer vision techniques with TrueEye 2.3, our state-of-the-art detector, examines every region of the image and answers in three bands:
Three possible answers
In testing, it caught 98.6% of images from the newest OpenAI, Google and ByteDance models on average, and 94.6% of AI images from 13 generators it never saw in training. On the other side of the ledger, the side that matters most to a newsroom, it wrongly flagged fewer than 1 in 2,000 original camera photos.
Measured the way Perspicax answers: an inconclusive answer counts as a miss. See every number
For newsroom tools and publishing systems, the same check is available through the Perspicax API, so every incoming image can be screened before an editor sees it.
Document and disclose
Write down what you checked and what you found, picture by picture: the source, the earliest copy, the location match, the detector's answer and any remaining doubts. When you publish, tell readers how you verified the picture, especially when it is central to the story.
When you can't verify a picture that is already everywhere, say that too. "We have not been able to verify this image" is useful information, and it stops your audience treating your silence as confirmation.
If you published a picture that turns out to be AI-generated, correct it prominently and explain how it got through. Readers forgive mistakes faster than they forgive silence, and the explanation helps other newsrooms avoid the same trap. For the same checks in everyday life, see check a photo before you share it.
Check an image now
Drop, paste or link a photo. Perspicax checks every region of it with computer vision and TrueEye 2.3, and answers likely AI-generated, likely real or inconclusive.
Questions people ask
How do journalists check if a photo is real?
They find the person who took it, trace the earliest copy with a reverse image search, check the place and time against maps, shadows and weather, get the original file, compare it with other footage, and use a detector as one more input.
Can a detector prove a news photo is real?
No. A detector tells you whether the pixels look generated. It can't confirm where or when a photo was taken, or whether the caption is true. Perspicax wrongly flags fewer than 1 in 2,000 original camera photos, but provenance still comes first.
What if a viral photo can't be verified?
Don't present it as real. If it is already widely shared and newsworthy, you can report that it is circulating and that you could not verify it, and explain what you checked.
Is photo metadata enough to verify a picture?
No. Metadata can support a story, but it is easy to edit or strip. Use it alongside the source, the location and other footage of the same moment.
Can Perspicax spot an edited news photo?
Not reliably. TrueEye 2.3 is built for fully generated images, so a real photo with a local edit may come back likely real or inconclusive. Compare it with other footage and the original file.
Can newsrooms check images automatically?
Yes. The Perspicax API lets a newsroom tool send each incoming image and get the same three-band answer, so editors see the result before they decide.