Checking an image or video for authenticity: What can you do when your footage is mistaken for AI?
You can check footage for authenticity if its origin is documented. The strongest way to do this is origin information that was set and signed at the moment of capture (Content Credentials), together with the original file. Without it, the original file, publication history and capture context remain as indications; camera metadata and detectors prove nothing on their own. The AI Act, the EU's AI regulation, requires no label for a real photo or video. If it comes under suspicion of being AI, all that counts is what you can present. We show which evidence holds up and why authenticity starts at the shutter release.
Three cases of suspicion: When a real photo is mistaken for AI:
Real photos come under suspicion of being AI today because the naked eye no longer settles anything; if you cannot present proof, the assumption wins.
In 2023, a photo competition in Sydney rejected a real smartphone photo because it looked too artificial to the jury. The organizer explained that it could not know for sure whether it was AI, but because it had the suspicion, it could not accept the image. The organizer apologized after the photographer came forward and confirmed that the photo was authentic. Three typical cases of suspicion since then:
- Competition. The jury or the rules demand proof, otherwise disqualification follows. At the 1839 Color Photography Awards in 2024, a real photo of a flamingo took third place and won the audience award in the category for AI images. The expert jury had not recognized it as real; after the photographer disclosed it, it was disqualified. Conversely, in 2026 a competition run by a camera manufacturer pointed to the RAW requirement for finalists when a finalist image looked suspicious, and excluded the image after review.
- News and event photo. Comments declare a real image a fake; it is proven through other footage and the publication history. A photo taken by a photographer of the news agency AP of the assassination attempt on Donald Trump on July 13, 2024 was taken for an AI image on social networks; a hand with six fingers served as evidence. PolitiFact, a US fact-checking newsroom, found that the circulating version had itself been manipulated; the original was confirmed by footage from other photographers taken in the same minutes and by the photographer's own account. An archive photo from 2007 showing Lionel Messi with the baby Lamine Yamal was considered an AI fake by many ahead of the World Cup final in July 2026; a fact check (Dubawa, August 2026) confirmed it was authentic through reverse image search and earlier publications.
- Platform label. A platform attaches an AI label to a real photo, triggered by metadata or a classifier. Metadata is invisible additional information stored inside the image file; people see none of it, checking software can read it out. From May 2024, Meta labeled images “Made with AI” as soon as their metadata contained an AI marking under an industry standard; that also hit real photos that had only been retouched with AI tools. In July 2024 this became “AI info”, and since September 2024 the notice appears only in the menu for content that was merely edited. Since April 2025, Pinterest has also been setting “AI modified” labels through its own classifiers, meaning detection programs that try to recognize AI content without a marking, from the image itself. In summer 2025, users reported labels on pins they had photographed or designed by hand without any AI; Pinterest acknowledged possible errors.
In none of these cases did the image itself settle the matter; what settled it was proof that the sender or third parties provided. In practice, the burden of proof lies with the sender. Why detectors and juries get it wrong: Deepfake detection and its limits.
What does it mean to check an image for authenticity?
Checking an image for authenticity means documenting its origin: which device or tool created it and when, and whether it has been changed since. The strongest evidence is signed origin information set at capture; after the fact, only indications remain. Whether the image shows reality truthfully is a different question that no proof answers.
What a suspicion of AI means legally
There is no labeling obligation for a real photo; the AI Act regulates AI content and is silent on real footage.
Article 50(2) AI Act obliges the providers of generative AI systems to mark their outputs in a machine-readable format; Article 50(4) AI Act obliges deployers, meaning anyone who uses an AI system under their own authority and puts the content to professional use, to disclose deepfakes visibly. Both duties presuppose content generated or manipulated by AI; an unedited photo or video is neither the output of an AI system nor a deepfake within the meaning of Article 3(60) AI Act. Who has which role is explained in detail in Who has to label AI images?.
The European Commission's guidelines on the transparency obligations from July 2026, which supervisory authorities go by, make clear that platforms that merely distribute other people's AI content are not deployers within the meaning of the AI Act. Whether a platform corrects a wrong label depends on its own procedures. Nor does the law recognize a detector's “probably AI” rating as a yardstick; detection within the meaning of the AI Act means establishing a provider's marking.
AI retouching is a different case. If AI did more than usual image editing, such as removing, replacing or inserting an object, the tool has to record this in the file. If a platform then shows “AI info”, that is no error: the notice points to a real intervention.
That notice is not your duty, though. You owe a visible label only if a deepfake was created. Retouching, light and color correction or a swapped background are usually not enough for that. A fashion retailer whose real product photo with a new background carries “AI info” is therefore seeing an accurate notice and still does not have to label anything itself.
If your real photo carries a platform's AI label, first check where it comes from. Pinterest sets labels based on metadata, its own detection program or self-disclosure, and according to its help page accepts appeals through support. Meta (Facebook, Instagram, Threads) reads AI markings under an industry standard from the file; Meta does not document an appeal route for wrong labels. There, check whether your editing tool wrote an AI marking. Tools for “removing AI labels” delete metadata and with it your own proof of origin.
Legally, then, a real photo needs no explanation; in practice, what counts is what you can present.
Checking photos and videos for authenticity after the fact: What you can still present
Without proof from capture, indications remain: the original file, publication history, capture context and a signature added later. Each indication proves something different, none proves authenticity on its own.
| Evidence | Shows | Limit |
|---|---|---|
| Original file (RAW or unedited camera JPEG) | The raw data of the capture exists; the usual test standard in competitions. World Press Photo, the foundation behind the competition of the same name, requires RAW files (one image is enough) or camera JPEGs as a series of at least seven images, checked by two independent analysts. | Strong, but not conclusive in itself: a RAW file carries no signature, and a fake is technically possible. Only someone who kept it can present it. |
| Camera metadata (EXIF) | Plausibility: camera model, exposure, time, location where available. | Unsigned and changeable: NIST, the US standards body, notes that metadata does not reveal manipulation and cannot be reliably attributed to anyone. Anyone who copies the file can rewrite or delete the metadata; it is lost on upload to many platforms and with a screenshot (AI watermarks: what remains when the metadata disappears). A note in an envelope that anyone can rewrite. |
| Publication history | The image was already published before the suspicion arose, with source and context; reverse image search is the standard method of fact-checking; this is how the 2007 archive photo was confirmed. | Only for images already published. Shows age and attribution, but not how the image came about. |
| Capture context | The event took place and you were there: burst shots before and after the subject, footage from others present, commission, invoice, location data; this is how the 2024 AP photo was confirmed. | Laborious and often impossible to reconstruct. Shows the event, but not an unchanged file. |
| Signature added later: Content Credentials (signed origin information under C2PA, an open standard that stores it directly in the file) via a signing service (Checking and understanding Content Credentials) | From the time of signing: who declared what about the file and that its signed parts have not changed since. | Proves from today onward, nothing retroactively. For anything created before that, no proof can be produced anymore. From the first entry on, however, the chain grows: every further edit attaches to the existing passport instead of overwriting it. |
| Detector report or marker check | “No marker found” or “no forensic anomalies” as one indication among others. | Absence proves nothing; detectors err in both directions (Deepfake detection and its limits). |
Whether a file already carries proof is shown by on:mint Verify, a free browser extension for Chrome and similar browsers: with a right-click, it checks any image online for Content Credentials, on:mint proof of origin and AI labeling.
All six types of evidence in the table have one thing in common: they come too late. Proof that holds up is created when the shutter is released.
Recording authenticity from the shutter release
Authenticity can be proven if it was recorded at capture: cameras from several manufacturers and the first smartphones now sign under C2PA when the shutter is released (C2PA explained). The signature proves the device and that nothing has changed, and with an embedded timestamp also the time; it does not prove that the content is true.
When the shutter is released, the camera creates proof of origin with checksums and device information, signs it digitally and embeds it in the image file. This is cryptography: the checksum (a cryptographic hash) is calculated from the signed file content; if a single byte changes there, it no longer matches and the signature becomes invalid. As the type of creation, the proof states “capture” (in the standard: digitalCapture), in the same field in which AI generators record “AI-generated”: the same list of ingredients, a different ingredient.
Authenticity cannot be certified after the fact. It is recorded at the shutter release or not at all. Whoever takes a photo today is already deciding whether that image can be proven tomorrow.
What the signature from the camera proves and what it does not:
- Proves: the device or tool (the signing identity), the time (with an embedded trusted timestamp), the type of creation (capture) and that the signed parts have not changed since signing.
- Does not prove: truthfulness, staging, framing. C2PA, the standards body behind Content Credentials, states itself that its standard makes no value judgments about provenance data; Germany's Federal Office for Information Security (BSI) adds that the source itself may have manipulated the material beforehand. And only what was included in the proof is signed; an independent security analysis by the University of Maryland, Baltimore County (UMBC, 2026) shows this with editable location data outside the protected area.
- Device limit: In early September 2025, a camera manufacturer suspended its signing service after a vulnerability in its camera software had allowed other people's images to be signed, including an AI image, and declared the certificates issued invalid; no resumption is known (as of September 2026). A signature is only as trustworthy as the device and the issuer behind it. Also, only new models sign, and the feature is not switched on by default everywhere.
- Transport limit: Platforms often remove metadata and with it the proof (Platform labels and the legal duty). YouTube shows “Captured with a camera” only if the metadata chain is intact and image and sound are unedited; if the notice is missing, YouTube says explicitly that this does not mean the content was edited. The way back is an invisible watermark in the pixels, that is, steganography: an identifier in the image itself that survives screenshots and re-uploading and leads to the stored proof (AI watermarks: what remains when the metadata disappears). Only C2PA plus an invisible watermark makes the proof hold up across uploads and screenshots.
Verification tools distinguish “valid” (signature intact), “trusted” (the signature chain ends at a listed trust anchor) and “No Content Credentials found” (Checking and understanding Content Credentials).
Conclusion: Checking an image for authenticity means proof instead of the naked eye
Real photos come under suspicion, and the naked eye no longer settles anything. The AI Act requires no label for a real photo; platform labels are measures taken by the platform, and with AI retouching, a label correctly reflects the tool's marking. There are two things to do: for existing images, secure the original file along with publication history and capture context, and have the images signed; in future, record authenticity when the shutter is released and protect the proof against upload and screenshot with a watermark.
That is what the on:mint signing service is for. It provides real footage and other originals with Content Credentials, meaning signed proof of origin under C2PA, and with an invisible watermark in the pixels. The proof is your image's digital passport: it names you as the creator and records when it was signed; existing Content Credentials, for example from the camera or from image editing, are carried over and extended. If a platform removes the metadata, the watermark leads back to this passport. on:mint signs with the certificate of a certification authority that is on the C2PA Trust List (the list of certificate issuers the standard trusts). Sign originals: Content Credentials by on:mint.
If, on the other hand, an image was generated or altered with AI, mintys, the AI labeling app by on:mint, labels it visibly and machine-readably under C2PA and secures both with an invisible watermark; when and by whom the labeling was done remains provable (AI labeling with mintys). on:mint has signed the Code of Practice on Transparency of AI-Generated Content and labels according to its procedures. In July 2026 the European Commission confirmed the Code as an adequate instrument for meeting the labeling obligations; according to its guidelines, it is so far the only EU-wide recognized practical framework for demonstrating compliance. If you label with mintys, you label in line with the Code and can expect fewer requests for information from supervisory authorities: the guidelines mainly expect additional evidence from companies that do not align their labeling with a recognized code.
On the cost side: the AI Act provides for fines of up to 15 million euros for breaches of the labeling obligation or, for companies, up to 3 percent of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises, the lower of the two applies. On top of that come warning letters: the Wettbewerbszentrale, a self-regulatory body of German industry for fair competition, treats breaches of the AI Act as unfair competition and set up a complaints office for AI cases in July 2026, explicitly including missing or incorrect labeling of AI-generated advertising. Warning letters, meaning cease-and-desist demands that come with a bill for costs, are therefore to be expected.
Three approaches compared. Toward the viewer, only the visible label counts; the proof counts when third parties spread your image without the label or someone disputes that it was labeled:
| Visible label only | AI tool's marking only | mintys by on:mint | |
|---|---|---|---|
| Visible label in the image, recognizable to the viewer without any tool (your duty as a deployer, Article 50(4) AI Act) | Met | Missing | Met |
| Label following the Code of Practice: EU icon with plain text, in the image, in a fixed position | Depends on implementation | Missing | Met |
| Labeling also machine-readable in the file, so that platforms and checking software recognize it | Missing | MetAs long as the metadata survives. | MetAs a Content Credential under C2PA. |
| Proof for each image that and when it was labeled, even if the label or metadata is lost along the way | Missing | Missing | MetWatermark in the image, passport in the file, record in mintys. |
With mintys you simply meet your labeling obligation under the AI Act in a single pass: label in the image, Content Credential in the file, watermark in the pixels, proof included. Label visibly under Article 50 AI Act: Label AI images now. For real footage, it works the other way round: sign originals with Content Credentials by on:mint before anyone asks.