To check whether an image is AI-generated, look for its source and any record of how it was made before relying on visual clues. Upload it to the Content Credentials Verify tool to see whether it carries a signed record of how it was made, then ask Gemini whether it contains Google’s SynthID watermark. If neither turns anything up, a reverse image search and a careful look at the source usually tell you more than staring at fingers. No single test is conclusive, and a clean result never proves a picture is real.
Why your eyes are no longer enough
A few years ago, extra fingers and garbled text were easy clues. Image generators have improved, and a real photo can also be altered with an object removal or face swap. Visual clues still help, but they are only one part of the check: provenance, watermarks and the image’s original source can offer evidence that is not obvious to the eye.
Step 1: Check for Content Credentials (C2PA)
Content Credentials are an open standard from the Coalition for Content Provenance and Authenticity (C2PA). The C2PA describes them as a nutrition label for digital content: signed metadata that records where a file came from and how it changed. A camera that supports the standard can sign a photo at capture, an editing app can add which tools were used, and an AI image generator can record that the image is synthetic.
To check a file:
- Save the original image, ideally the highest-quality version you can find rather than a screenshot.
- Open contentcredentials.org/verify and drag the file in. It accepts common formats including JPEG, PNG, WebP, HEIC, AVIF, TIFF, MP4, MOV, MP3 and WAV.
- If credentials exist, you see who signed them, which app or device created the file, and whether AI tools were involved, along with its edit history.
The catch is coverage. The Verify page itself warns that Content Credentials are still rolling out, so most images you find online carry none. Metadata can also be lost when a platform strips it on upload or when someone screenshots the picture. So treat a valid credential as strong evidence, and a missing one as no evidence either way.
Step 2: Ask Gemini or Google Search about SynthID
SynthID is Google DeepMind’s invisible watermark. Google uses it across supported AI-generated images, video, audio and text. DeepMind says image and video watermarks are designed to withstand edits such as cropping, filters and compression. Unlike metadata, it sits in the pixels themselves, so it does not depend on a platform keeping the file’s metadata intact.
How to check:
- Gemini app: upload the image, video or audio clip and ask “Is this AI-generated?” or “Was this created or edited by Google AI?” Gemini checks for a SynthID watermark. Content Credentials checks are also available on the web and Android; Google says iOS support is coming.
- Google Search and Chrome: Google says you can ask Gemini, Search or Chrome to check for a SynthID watermark. The Gemini check currently identifies content made or edited with Google AI; a missing watermark does not rule out other AI tools.
Gemini may report that it found a watermark, found none, or could not tell. Here is how to read those results:
| Gemini says | What it means |
|---|---|
| SynthID watermark detected | All or part of the content was made or edited with Google’s AI models |
| No SynthID watermark | No Google watermark was detected. AI involvement is still possible, including content whose watermark cannot be recovered. |
| Unclear | Often “not enough details to watermark” or “likely too small an edit” |
A missing watermark only means Gemini did not find Google’s mark. It does not tell you whether another company’s model made the image.
Step 3: Run a reverse image search
A reverse image search answers a different question: where else has this picture appeared? Use Google Lens (in the Google app, Chrome or Google Images) or TinEye, and look for:
- The earliest copy. A picture that appeared years ago cannot document a newly unfolding event, even if the original photo is real.
- A higher-quality original. Viral images are often cropped. The full frame can reveal a watermark, a caption or a creator who labelled it as AI art.
- Fact-check articles. Widely shared fakes are often already debunked by news agencies and fact-checkers.
Step 4: Look closely, in the right places
If the provenance checks come up empty, zoom in. Focus on the parts generators and editing tools still find hard:
- Text: signs, labels, number plates and logos that are garbled, misspelled or change style mid-word.
- Hands, teeth and ears: fewer errors than before, but still worth checking, especially in group shots.
- Light and shadow: shadows that fall in different directions, reflections in glasses or windows that do not match the scene.
- Edges and backgrounds: hair that blends into the background, objects merging into each other, repeated faces in a crowd, architecture that does not quite connect.
- Skin and texture: an even, airbrushed look across every person in the photo.
- Physics and context: weather, clothing, uniforms or vehicles that do not fit the place and date claimed.
Treat these as reasons to dig further rather than proof. Real photos can have odd shadows, and heavy phone processing or beauty filters create the same smooth skin.
Step 5: Check the source and the story
A real photo with a false caption can mislead just as much as an AI-generated one. Ask who posted it first, whether they have a history of real reporting, and whether reliable outlets show the same event from other angles. A dramatic image that only exists on one account, with no other photos or video from the scene, deserves suspicion however convincing it looks. Tools such as the Instagram Edits AI assistant make AI edits accessible to ordinary users. An edit alone does not establish deception; the claim attached to it matters.
How to spot deepfake video and audio
The same order applies to video and audio: provenance first, then context, then your eyes and ears.
- Check provenance. The Content Credentials Verify tool accepts MP4, MOV, MP3 and WAV files, and the Gemini app checks video and audio clips for SynthID.
- Watch the face and mouth. Lip movements that drift out of sync, teeth that blur, blinking that looks unnatural, and a face that wobbles when the head turns or a hand passes in front of it.
- Listen for flat delivery. Cloned voices can sound right but carry odd pacing, missing breaths, or the same tone whatever the words.
- Verify through another channel. If a “relative” or “boss” calls or sends a voice note asking for money or codes, hang up and call them back on a number you already have. An agreed family code word helps too.
Tools at a glance
| Tool | What it checks | Where |
|---|---|---|
| Content Credentials Verify | C2PA signed metadata and edit history | contentcredentials.org/verify |
| Gemini app | SynthID watermark from Google AI, plus Content Credentials | Gemini app on web and mobile |
| Google Search and Chrome | Checks for Google’s SynthID watermark | Google Search, Chrome |
| TinEye | Earlier copies of the same image | tineye.com |
Pixel-based AI detectors estimate whether an image resembles generated content. They can miss edited images or flag real ones, so treat a score as a reason to investigate further.
What to do when you find a fake
- Do not share it, even to mock it. Reposts help it spread.
- Reply with what you found, such as the original source or a fact-check, rather than just calling it fake.
- Use the platform’s report option for manipulated or misleading media.
- If it targets you, such as an explicit deepfake or an impersonation scam, save the evidence, report it to the platform, and contact the police.
The same habit applies to claims about AI itself. The details of the Google Gemini hacking test matter more than the viral description of the result.
FAQ
Can I tell if an image is AI-generated with 100% certainty?
No single check can classify every image. Content Credentials can describe a real camera capture, an ordinary edit or AI generation; read what the signed record actually says. A detected AI watermark supports AI involvement, but an absent watermark or credential does not prove a photo is real.
Does a screenshot remove the AI label?
A screenshot usually drops Content Credentials metadata, because it creates a new file. SynthID is built into the pixels rather than the metadata, and Google says the watermark usually survives resizing, recolouring and other edits, though many alterations can make it undetectable.
Can Gemini detect images from ChatGPT or Midjourney?
Gemini’s help page describes its SynthID check as identifying Google AI content, with Content Credentials providing another way to read a file’s history. Separately, Google’s SynthID Detector supports content from participating partners, including OpenAI. This is not a universal detector for every image made by ChatGPT or Midjourney.
Are AI detector websites reliable?
Treat them as a hint. They estimate a probability from the pixels and can be wrong in both directions, especially on compressed or edited images. Use them alongside provenance checks and source checks, not instead of them.
What should I do if someone uses a deepfake to target me?
Save the original file, screenshots and links, report the content to the platform, and contact local police if it involves threats, impersonation or intimate imagery. In India, you can also submit a cybercrime complaint through the National Cyber Crime Reporting Portal. The legal position depends on the content and circumstances.
- SourceGoogle DeepMind: SynthID
