Visual signs and why you cannot fully trust them
An image generator is most often built as a diffusion model: step by step, it turns random noise into an image that matches a description. In small details that need logic rather than beauty, it sometimes makes mistakes.
Hands and ears. An extra or fused finger, an earring in only one ear, a ring that sinks into the skin.
Text in the frame. Signs, car number plates and writing on clothes made of symbols that look like letters but mean nothing.
Light and shadows. Shadows fall in different directions, and a reflection in a mirror or glasses does not match the room.
Background. People in a crowd merge, building lines bend, and objects at the edges of the frame flow into each other. Skin is too smooth, like a doll’s.
Why this is not enough: new models fix such mistakes, and real photos can have glare, blur and editing. The International AI Safety Report 2026 notes that AI-made content has become harder to tell from real content, and detectors have limited success.
Bad: “The fingers are fine, so the photo is real.”
Good: “The fingers are fine, but the writing on the sign makes no sense. I will check the metadata and look for where the photo first appeared.”
C2PA metadata: how to check where a file came from
Metadata is the technical information inside a file: what took the picture, when, and in which program it was edited. C2PA is an open standard that records the origin of a file in this information and signs it. This signature is called Content Credentials.
According to C2PA, the signature is cryptographic: if the file or its information changes, the signature no longer matches, and the edit becomes visible. C2PA estimates that forging the signature itself is practically impossible with today’s technology.
How to check. The Content Authenticity Initiative points to the verification site contentcredentials.org/verify. Upload the file there: the site shows whether there is a signature and what it records, for example what created the file and what edits were made.
Do not upload personal photos, pictures of documents or other people’s private images to foreign websites. You do not need to in order to check a suspicious picture from the web.
The main limitation. C2PA states plainly that the signature is usually embedded in the file but can be separated, and metadata gets removed by accident or on purpose. The Content Authenticity Initiative adds that a screenshot, a photo of the screen or a re-recording in effect erase the signed information.
So “no signature” means only one thing: there is no signature. It proves neither that the photo is real nor that AI made it.
Bad: “No Content Credentials were found, so it was taken with a phone.”
Good: “There is no signature. I am looking for the original source and at the details.”
Invisible watermarks
An invisible watermark is a mark that a generator hides in the pixels, sound or text themselves. A person cannot see it, but a checking program from whoever added the mark can find it.
One example of such a mark is SynthID from Google DeepMind. According to the developer’s page, the mark is embedded in images, video, audio and text from Google’s generative services. In pictures it is designed to survive cropping, filters and lossy compression.
Google offers SynthID checks in its services and on the SynthID Detector portal launched in 2025. According to Google, the portal recognizes images, video and audio from Google and its partners, including OpenAI and NVIDIA. Google’s page does not say whether the check works from Russia, so here we only describe it.
The weak spot is the same as with the signature. Only generators that add the mark have it, and a picture from another generator gets no mark. The International AI Safety Report 2026 says that watermarks and metadata can be changed or removed.
Bad: “The check found no watermark, so it is not AI.”
Good: “There is no mark. So a generator without such a mark made the picture, or the mark was erased, or it is a real photo.”
Reverse image search
Reverse image search looks not by words but by the picture itself. You upload a photo, and the search engine shows similar images and the sites where this or a similar image was already published.
In Google, according to its help page, you do it this way: open google.com, click the “Search by image” icon and upload the file. You can drag the image into the search box or paste a link to the picture. In the Chrome browser, right-click the picture and choose “Search with Google Lens”.
What to look for in the results. Find where the picture appeared first: with a photographer, in the news, or with an author who labeled it “created with AI”.
If the same picture turned up in another news story, with a different date or caption, do not trust the caption, even if the photo itself is real.
The Google help page warns that the service may store the addresses of images you search with, to develop its products. Do not upload personal photos or pictures of documents to search.
Bad: “The search found nothing, so the photo is real.”
Good: “The search found nothing. I am looking at the details and the metadata, and I will not forward the photo until I know where it came from.”
If a photo is used against you
If a fake photo of you appears online, first save the evidence: links, dated screenshots of the page, the account name. Do not get into a conversation with whoever posted it, and do not pay to have it removed.
Under Article 152.1 of the Russian Civil Code, a person’s image may be published and used only with their consent, except in cases the law names. If such an image was spread online without consent, you have the right to demand its removal and a ban on further distribution.
Report the post through the platform’s own form and attach the evidence. If the photo is used for threats, extortion or to trick your relatives out of money, go to the police.
Warn your family and colleagues that such a photo may appear, and agree to call back on a known number before believing a message. Fake photos, videos and voices made by a neural network are called deepfakes. How to recognize them is covered on the page on how to spot a deepfake.
Bad: “I will delete my page and stay silent; maybe everyone will forget.”
Good: “I saved the links and screenshots, filed a report with the platform and warned my family.”