Authentic photos, false context: the fake no detector can find
Published September 1, 2026
Most of the visual disinformation reaching your newsroom or your brand’s mentions was never edited. The photograph is real, taken by a real camera in a real place on a real day. Only the caption is false, and it moves the picture to another month or another country.
That is awkward, because almost every tool built for this problem looks inside the file. Detectors of generated images, detectors of manipulation and forensic inspection of compression and sensor noise all read pixels, and an untouched photograph gives them nothing to report.
Nothing is wrong inside the file, and that is the point. The manipulation sits in the relationship between the content and the world, in the when and the where, and that relationship was never written into the file. That leaves one approach: fixing provenance at the moment of creation, so time and place travel with the image instead of being argued over later.
What an out of context image actually is
An out of context image is an authentic photograph, neither edited nor generated, republished with a caption that falsifies its time, its place or the event it is said to document. The file is original. Only the claim attached to it is false, which is why technical analysis of the image returns nothing.
The category is formalised, not an editorial invention. Among the ratings that Meta’s fact-checking partners apply to content, one covers authentic images, audio or video presented as evidence of an unrelated event. English-language fact-checkers describe the same thing as miscaptioned content, and researchers call it false context. The label matters because it separates two problems that routinely get treated as one: fabricating a piece of content, and misusing a piece of content that was never fabricated. Only the first of the two leaves traces inside the file.
Altered content, generated content, recontextualised content
The three get filed under one heading and behave nothing alike. Altered content was reworked after capture, so the file carries the inconsistency. Generated content never had a camera in front of it, so the model’s fingerprint sits in the pixels. Recontextualised content has neither problem, because nothing was done to the file at all.
| Type of content | What changes in the file | What analysis detects | What it takes to disprove it |
|---|---|---|---|
| Altered | Pixels reworked after capture | Inconsistent compression, sensor noise, shadows, edges | Forensic analysis of the file |
| Generated | No camera origin at all | Statistical traces of the generative model | Detection tools and disclosure rules |
| Recontextualised | Nothing, the file is original | Nothing to find in the file | External evidence of the real time and place |
Why recontextualised images are the most common form of visual disinformation
Real information used misleadingly is more common than invented stories, a point the European Parliament makes in its overview of the tactics used to mislead us, where context distortion counts as a technique in its own right. In the monthly fact-checking brief published by the European Digital Media Observatory in January 2026, 16% of 1,605 fact-checking articles published in Europe during December 2025 concerned content created or altered with artificial intelligence, a record high for that series; between April and May 2026 the share fell from 20% to 14%. Even in the peak month, then, more than four fifths of what European fact-checkers examined was not generated by a model.
Trust is paying for it. The Digital News Report 2026 from the Reuters Institute records 62% of respondents worldwide worried about telling real from fake news online, four points up year on year, with trust in news at its lowest since measurement began in 2015.
Why no file analysis can flag it
No file analysis flags an out of context photograph, because the file really is unaltered. A generated image detector searches for statistical signals left inside the picture by a generative model. A manipulation detector searches for discontinuities in compression, in sensor noise, in shadows and edges. Both work on the content of the file, and both are right when they report that an authentic photograph republished with the wrong date shows neither signal. The image is what came out of the camera. What has been falsified is the assertion wrapped around it, and that assertion has no pixels. Every detection tool can be right about the file and useless about the claim at the same time, which is how a photograph that has already been debunked twice keeps circulating with a new caption.
What downstream detection is actually looking for
Detection is a search for residue. A model that generates an image leaves regularities that never occur when light hits a sensor, and detectors are trained to find them. Editing leaves its own residue: a region compressed twice, a noise pattern that stops at a boundary, a shadow pointing the wrong way. Both assume that something interfered after capture, or instead of it, and left a mark.
Recontextualisation breaks that assumption without doing anything clever. Whoever republished the photograph needed no editing skills and no model, only a caption box, so nothing was left behind in the object being analysed.
The structural limit: the manipulation sits outside the file
A photograph does not record when and where it was taken in any way that survives contact with the internet. Time and place live in metadata that sharing platforms remove, and intact metadata is written by the device with nothing external vouching for it, so changing it is trivial. The two facts that decide whether a caption is true are the two the file is least able to defend.
It is also why transparency rules for synthetic media miss the problem. The EU AI Act obliges providers and deployers to mark artificially generated or manipulated content so people can recognise it, and here the obligation has nothing to attach to, because nothing was generated and nothing was manipulated. A perfectly labelled ecosystem of AI content would leave the most widespread form of visual disinformation untouched, which is why labelling AI content is not the same as proving what is true.
How to know when and where a photograph was really taken
Establishing when and where a photograph was taken is a question of provenance, not of image analysis. Provenance is the documented record of where a piece of content came from and what happened to it since, and it is either produced when the content is created or reconstructed later from whatever survived, which is a different discipline with a much lower success rate. The distinction matters in practice, and it maps onto the difference between data provenance and data lineage: one tells you the origin, the other tells you the path. For a contested photograph you need the origin, in a form somebody else can check. What remains after the fact are partial routes: each tells you something real, and none of them closes the question on its own, which is the whole difficulty.
EXIF metadata, and why sharing platforms strip it
EXIF metadata is what a camera writes into a file at capture: date, time, device and, when location is on, GPS coordinates. It survives a direct transfer, not a platform. WhatsApp, Instagram, X, TikTok, LinkedIn and Reddit recompress uploads and strip EXIF from the file other users can download, capture time and coordinates included. The stated reasons, privacy and file size, are legitimate; the side effect is that the file stops carrying its own history. Sending the picture as a document preserves the metadata, but only if the sender cooperates and the file has not already crossed another platform. On what EXIF can and cannot establish once a date is contested, see EXIF metadata and the date of a photo as court evidence. Newsrooms use TrueScreen to capture images in the field with date, time and coordinates recorded as the shot is taken, rather than depending on metadata platforms remove.
Reverse image search and geolocation: what they prove and what they do not
Reverse image search shows where a picture has already appeared. Finding an older publication proves the image is old, and that move settles many cases. Failing to find one proves nothing: the test can falsify a claim, never confirm it, and it fails whenever the image is new, has been cropped or mirrored, or the earlier posts have been taken down.
Geolocation from visual clues is the more demanding route, and often the more convincing. Signage, road markings, architecture, vegetation, the angle of the sun, comparison with satellite imagery and street-level views: fact-checking desks work this way daily, and on the right frame it is decisive. It needs time, trained eyes and one identifiable element in the frame. A patch of sky, an anonymous crowd or a plain interior yields nothing, and those are the frames that get recaptioned.
Verification that got lucky, and verification you can repeat
Lucky verification depends on a favourable circumstance: a readable sign in the frame, an archived copy someone kept, a metadata field that survived by accident. It works when it works, and gives you nothing on the next file. Web archives show the gap: they capture whatever a crawler reached at a moment nobody chose, hence the evidentiary limits of the Wayback Machine.
Repeatable verification rests on information produced at the moment of capture and bound to the file so a third party can check it: when, where, from which device, and with what integrity since. ISO/IEC 27037, the international standard for handling digital evidence, rests on the same principle: a documented, reproducible process beats a fortunate outcome. The difference is not the result, it is whether the method survives being questioned. So the problem is addressed upstream, by recording provenance at the moment of creation instead of reconstructing it afterwards.
Certification at source: binding a photograph to its own time and place
TrueScreen certifies photographs and videos at the instant they are captured, binding the content to the moment and the place of acquisition with a timestamp and a digital seal. The capture runs inside a controlled process rather than the ordinary camera roll, so date, time and coordinates are recorded as the shot is taken and sealed with the file. Under eIDAS, the European framework for electronic identification and trust services, electronic timestamps and electronic seals carry legal effect and are recognised across member states. The result is not an opinion about whether an image looks authentic. It is a photograph whose evidentiary value rests on its own provenance, documented at the only moment when that documentation can still be produced honestly. Unlike systems tied to one specific camera model, it runs on the devices people already carry.
Take an ordinary newsroom exchange. A contributor on the ground sends a photograph of a damaged building. If the shot was certified at source, the desk has date, time and coordinates recorded at acquisition and bound to the file, and can publish without reconstructing anything. If it arrives forwarded through a chat, the desk has the contributor’s word and the time the message arrived, which will not survive a challenge.
Detection asks whether something is fake, and that answer weakens as generation improves. Digital provenance asks what is verifiably true about a file, and holds however convincing the fakes become. For images used out of context it is the only answerable question, because the falsification never entered the file.
Frequently asked questions
What is the difference between an edited photo and a photo used out of context?
Why do AI image detectors not flag an authentic photo used out of context?
How can you tell when a photo was actually taken?
Why does a photo received on WhatsApp no longer have its capture date?
Can you find out where a photo was taken?
Is reverse image search enough to verify a photo’s context?
What makes a verification repeatable rather than lucky?
Fix provenance when the content is created
Capture photos and videos with the date, time and place recorded at the moment of capture and bound to the file, so the context never has to be reconstructed later.
