Counting fingers stopped working in 2024. So did looking for garbled text on shop signs, melted ears and the strange plastic sheen that gave away the first generation of image models. If you are still using the checklist that circulated three years ago, you are checking for problems that have been fixed. The reliable method in 2026 runs in the opposite order: interrogate the file first, then the picture, then the source.
What follows is the order to work through, what each step can and cannot tell you, and the one misreading that trips up almost everybody who tries this.
Step one: check the file before you look at the picture
Two invisible systems now travel with a lot of generated media, and both take seconds to query.
SynthID, through the Gemini app
Google embeds an invisible watermark called SynthID into media produced by its own models, and has applied it to more than 20 billion pieces of content since 2023. Since November 20, 2025 you can query it yourself: open Gemini, upload the file and ask “was this created or edited by Google AI?”. It now handles images, video and audio up to 100 MB, with a quota of roughly ten checks of each type in any rolling 24 hour window.
Google is blunt about the boundary. In its own words, Gemini “can currently only recognize content created by Google AI tools”. The watermark also survives rescaling, recoloring and compression in most cases, but Google warns that after enough alterations it may no longer be detectable. There is a separate SynthID Detector portal aimed at journalists and researchers, which localizes the watermark to specific regions of a file, but it is still behind a waitlist rather than open to everyone.
Content Credentials, through the C2PA verify tool
The second system is provenance metadata rather than a watermark. Content Credentials record what made a file and what was done to it afterwards, cryptographically signed. Paste or upload an image at verify.contentauthenticity.org and, if credentials are present, you get the generating tool, the edit history and the signing organization. Google began embedding C2PA metadata into output from Nano Banana Pro, Vertex AI and Google Ads in late 2025, and several camera makers now write credentials at capture.
You can also read the raw metadata yourself if you have the file locally. This is faster than any website when you are working through a batch.
# Look for a generator signature or C2PA claim in the file itself exiftool -a -G1 -s suspect.jpg | grep -i -E "software|creator|c2pa|claim|jumbf" # A real camera photo usually carries these; a generated image usually does not exiftool -Make -Model -LensModel -ExposureTime -FNumber -ISO suspect.jpg
Step two: the visual tells that still work
Modern models have fixed the famous failures. What they have not fixed is physical consistency across a whole scene, because the model is predicting plausible pixels rather than simulating a world. Look for places where two parts of the image have to agree and do not.
| Where to look | What tends to break | Still reliable? |
|---|---|---|
| Reflections in mirrors, windows, sunglasses, water | The reflected scene does not match the scene, or reflects something absent | Yes, one of the strongest |
| Shadow direction across multiple objects | Shadows point to two different light sources in one frame | Yes |
| Small text deep in the background | Foreground text is now clean, distant signage still dissolves into near letters | Often |
| Crowds and repeating patterns | The same face or the same shirt appears three times, brickwork loses its grid | Yes |
| Where hair meets a busy background | Strands blend into the background instead of overlapping it | Sometimes |
| Depth of field | Objects at the same distance are blurred differently, or blur ignores the lens | Yes |
| Hands, ears, teeth, jewelry | The classic 2023 tells | No, largely fixed |
Zoom to 200 percent and work the edges of the frame rather than the subject. Generators spend their quality where the eye goes. The corner of the image, the reflection in a car door and the text on a distant awning are where the model relaxes, and that is where the contradiction lives.
Step three: check the source, not the pixels
This is the step that actually settles most cases, and the one people skip because it feels less clever than pixel forensics.
Reverse image search the picture in Google Lens and TinEye, then sort by date to find the earliest appearance. Ask three questions. Who published it first? Does any organization with something to lose stand behind it? Does a second, independent image of the same event exist from a different angle?
A real event photographed by real people generates multiple images from multiple positions. A synthetic one almost never does, because whoever made it made one. That asymmetry catches more fakes than any detector. When a fabricated NASA announcement went viral this summer, the fastest way to break it was not analyzing the graphics but noticing that no primary source existed anywhere and the date had been chosen to make it feel plausible.
Google Search itself helps here. The “About this image” panel, reachable from the three dot menu on an image result, shows when Google first indexed the picture, where else it has appeared and how other sites have described it. That history is often the whole story.
Why AI detector websites are a trap
Search for a detector and you will find dozens of sites that accept an upload and return a confident percentage. Treat those numbers as decoration.
They are statistical classifiers guessing from texture and frequency patterns, not readers of provenance. They produce false positives on heavily edited real photographs, on scanned film, on anything that has been through a phone’s night mode, and they degrade the moment a new model ships. A detector that says “97 percent AI” is telling you about its own training distribution, not about the image. The only tools that give you a checkable answer are the ones reading a watermark or a signed credential, because those either verify or they do not.
The wider point is that detection and generation are locked in an arms race, and detection loses more rounds than it wins. Researchers have been demonstrating for years how easily machine vision can be steered, right down to a printed shirt that makes a person invisible to an object detection model. Anything that classifies images from pixels alone can be gamed by someone who understands the classifier.
What changes now that the EU rules apply
Since August 2, 2026, Article 50 of the EU AI Act has required providers of generative systems to mark their output in a machine readable format that is detectable as AI generated, across text, images, audio and video. Systems already on the market before that date have until December 2, 2026 to comply. The obligation sits with providers rather than with you, and it does not apply where AI only performs assistive edits such as grammar correction.
In practice this pushes the industry toward exactly the two systems above, watermarking and signed credentials, and away from guesswork. It does not fix the internet. A model run locally on someone’s own hardware, or one operated outside the EU for an audience outside the EU, marks nothing. Expect labelled output to become the norm on mainstream commercial tools and to remain absent everywhere it matters most.
Troubleshooting
Gemini says it cannot tell
That usually means no SynthID watermark was found, which is the ambiguous result rather than a clean answer. Google also returns nothing useful when the edit was too small to watermark or the image is too low resolution to carry one. Move to the C2PA check and then the source check.
The verify tool shows credentials but I still do not trust it
Read the signer, not just the presence of a credential. Content Credentials tell you who signed the claim, and a claim signed by an unknown entity carries the trust of that entity, which may be none. A credential from Adobe, Google or a camera manufacturer means something. One from a name you have never seen means it was signed by someone.
The image came from a screenshot and has no metadata at all
Expected. Screenshotting destroys both watermark reliability and provenance metadata, which is why it is the standard laundering step for anyone spreading a fake deliberately. At that point only the visual checks and the source trail are available to you.
Two tools disagree
Prefer the one reading a cryptographic signature over the one making a statistical guess, and prefer the source trail over both. Provenance beats detection every time it is available.
Frequently asked questions
Does counting fingers still work?
No. Hands, ears and teeth were the tells of 2023 and current models handle them reliably. Building a judgement on them now produces both misses and false accusations. Physical consistency across the whole scene has replaced anatomy as the useful signal.
Can I check a video the same way?
Partly. Gemini now accepts video and audio uploads for SynthID checks alongside images, with the same limitation that it recognizes Google AI output only. For everything else, the source check matters even more, because a synthetic video almost never has a second independent recording of the same moment.
Do social platforms label AI images automatically?
Some do, when the file arrives with credentials or a watermark intact. That condition often fails, because uploading, recompressing and screenshotting strip exactly the data a platform would read. An unlabelled post tells you nothing either way.
Is there one tool that just gives me the answer?
No, and be suspicious of anything claiming otherwise. Watermark checks answer “was this made by a participating generator”. Credential checks answer “what does the file say about itself”. Neither answers “is this real”, and no classifier answers it honestly either.
How do scammers use generated images?
Mostly for fake profiles, fake product listings, fabricated proof of delivery and invented endorsements. The image is rarely the attack, it is the prop that makes the attack credible, which is why our guide to spotting online scams focuses on the request rather than the picture.
The bottom line
Run three passes in this order and you will be right far more often than the people arguing in the replies. Query the file for a watermark and for Content Credentials. Read the picture for physical contradictions, working the edges rather than the subject. Then find the earliest copy and ask who stands behind it.
The one habit that matters most is refusing to read an empty result as proof of authenticity. Most images on the internet carry no watermark and no credentials, and always did. Provenance tells you something when it is present and nothing when it is absent, and knowing that difference is most of the skill. If you want to get comfortable running these checks quickly, our beginner walkthrough of Gemini covers the upload flow you will use for the first step.
