An Era When Real Photos and AI Images Are Hard to Tell Apart
Sometimes you see a photo on social media and wonder whether it's real or something an AI made. Say someone you find really attractive posts a stunning photo, and you start thinking, 'Wait, is this an AI-generated image?'
And these days, AI-generated images have gotten so good that it's not easy to tell them from real photos with your eyes alone.
The technology you need here is SynthID Detector, which Google opened to the public on October 7. Upload an image file and it finds the invisible watermark planted during AI generation. It can check not only video but audio too.
A Digital Signature Hidden in the Photo
SynthID is an invisible watermarking technology developed by Google DeepMind. It's different from slapping a logo on an image or writing 'AI' into the file info.
It tucks a mark that people can barely notice by eye or ear into the pixels of an image and into video and audio signals. The Detector's job is to read whether that mark is still in the file.
✦ Two Ways to Spot AI Images
Guessing from the look vs. reading a planted signal
A typical AI image detector analyzes features like an image's texture and shape to estimate how likely it is to be generated. SynthID Detector, on the other hand, checks for a specific signal inserted at the generation stage.
So if an image came from a supported AI, you can find the source mark no matter how natural and photo-like it looks. Of course, this approach has a clear limit. An image that never had SynthID in it isn't something the tool can analyze in the first place.
Anyway, the tool's early version was released to journalists, media professionals and researchers first, in May 2025. It was a limited trial service back then. And since October 7, 2026, users around the world can check files directly on the website.
It Catches ChatGPT Images Too
What's different with this release is the scope of the checks. Google expanded the Detector so it can check content from partners that have adopted SynthID, such as OpenAI, NVIDIA, Kakao and Apple, not just its own AI, all in one place.
Google's announcement, updated on October 29, includes Apple as a partner too, and Apple has started applying SynthID to Image Playground images and some photo-editing results.
✦ Gemini App vs. Web Detector
Same company, different reach
The built-in check inside the Gemini app still only recognizes Google AI content, but this web Detector can check partner watermarks too.
PCWorld reporter Ben Patterson asked Gemini and ChatGPT to each make an image on the same subject, a robot visiting Disneyland.
Once the two files were uploaded to the Detector, SynthID was detected in both, and they were sorted into Google AI and OpenAI (ChatGPT) output, one each. It's a test that shows images from different companies get sorted by source in the same checking window.
✦ C2PA vs. SynthID
Two different ways to mark a file
By the way, OpenAI apparently adds not only SynthID but also 'C2PA Content Credentials' to images generated in ChatGPT, Codex and the API.
So How Do You Actually Use SynthID Detector?
SynthID Detector is a web tool released for free, with no separate subscription fee. Just go to the official site (https://synthid.com) and you can use it.
✦ How to Use It
From the website to your first check
To upload a file, you need to log in with a Google, OpenAI or Apple account. It's free, but there's a usage limit. Google says the daily allowance is about 10 checks, with images, video and audio combined.
Supported formats include images in JPG, PNG, WEBP, AVIF, HEIC, TIFF and GIF, video in MP4, MOV and WEBM, and audio in WAV, MP3, OGG, FLAC, AAC, M4A and more.
The SynthID technology itself also has text watermarks, but what you can check in the publicly released Detector is limited to images, video and audio. And although the site opened worldwide, the interface is still only in English.
✦ What the Detector Reads
Not "is it AI?" but "is the SynthID signal there?"
Instead of a complicated score table, results center on whether a watermark was detected and what source label it carries. That's because the Detector doesn't analyze an image to confirm whether it's AI-generated.
It only reads the SynthID signal a model put in during generation or editing.
Content from models that use a different watermarking method, or that never applied SynthID, falls outside what it checks. A baby-and-puppy video that PCWorld tested is a good example.
The reporter uploaded a file that looked like an AI video, but SynthID found no trace of itself. The other AI videos the reporter dug up on Reddit and elsewhere came out the same, since no SynthID watermark turned up.
The SynthID watermark goes in during AI editing as well as AI generation. SynthID can stay in even an original photo shot with a camera, if an AI editing feature was used on it.
So as more companies join this program, it seems like picking out AI-made content will get steadily easier.
The Watermark Isn't Guaranteed to Survive
SynthID plants its signal in the pixels of an image, not in the file metadata. So even if you strip the author info or metadata from a file, the watermark may survive.
In particular, Google said it designed the signal to withstand cropping, filters, frame rate changes and lossy compression. But they say heavy cropping or strong compression can make the watermark harder to detect.
Ars Technica tested this property using two images made with Nano Banana Pro. One was generated by AI from scratch, and the other was a real photo edited with AI.
The reporter simulated 300 rounds of repeated compression and resizing, and the watermark was detected in the full frame of both images all the way to the end. The signal survived in a file saved as a screenshot too.
✦ Ars Technica's 300-Round Test
Full frame vs. cropped
But when the edges of the 300-round images were cropped by about 20%, the watermark wasn't detected in either one. A 50% crop cut off detection after about 250 rounds.
Sure, this is an experiment on two images degraded to an extreme, but it shows that even a watermark embedded in the pixels can lose its signal when part of the image is gone.
There Are Still Technical Limits, But...
SynthID Detector also drew controversy over false positives (detection errors) right after its release.
A prominent case: Full Fact, an independent fact-checking organization in the UK, found five photos posted by politicians that got a verdict of 'made or edited with Google AI.'
The posters said they hadn't used AI, and Full Fact also judged a false positive likely. On the specialist checking screen, watermarks even showed up in places like trees and grass that didn't seem related to any AI editing.
✦ After the Full Fact Report
What happened next
After the problem was reported to it, Google updated its judgment method on October 15. When Full Fact re-checked the five photos, all of them changed to 'Unsure' instead of a firm verdict.
That label is a stage used when the watermark signal is faint or the call is ambiguous. But even after the update, a case was reported in which a photo from a politician who said they hadn't used AI was flagged as AI-edited again.
There are still technical limits, and updates to cut down on false positives will need to keep coming.
SynthID Detector's strength is that it checks different companies' watermarks in one place. Detection cases have actually come out of Gemini and ChatGPT images, and uploading a file is simple.
It supports video and audio too, so you can check not just a single image but many kinds of generated content shared online. If you ever need to detect AI-generated content, for work or personal use, give it a try.
That's all for this post. See you next time.
Related posts
- Nano Banana 2.1: 4K Images at $0.113 Each, Worth It?
- YouTube Music's AI Labels: How to Disclose Your AI Music
- MAI Image 2.7 Catches Up to GPT Image 2.5 in Image Editing!
🌐 Korean version: "이거 AI가 생성한 이미지인지 확인해줘!" 《구글 SynthID Detector》가 무료로 공개됐다
🌐 Japanese version: 「AI製か確認して！」GoogleのSynthID Detectorが無料公開
