How to Tell If Content Is AI-Generated: Google's SynthID Detector Is Now Public

On October 7, 2026, Google opened synthid.com to anyone with an account. You can upload an image, video, or audio file, and the tool will tell you if it contains SynthID—the invisible watermark that Google and a few partners embed in the content their models generate. Until now, the detector was reserved for journalists and researchers who requested access.

If you create short-form videos, this is a big deal. TikTok, YouTube, and Instagram require you to label realistic AI-made content and also actively scan for it using technical signals in the file. If a video is AI-generated and isn't labeled, the platform might add the label itself, limit its reach, take it down, or issue a warning or strike. Repeat offenses can lead to demonetization. The SynthID detector shows you what these systems see in your file. Let's break down what it finds, what it misses, and why "no watermark found" doesn't mean "filmed by a human."

What Is SynthID and What's Changed?

SynthID is an invisible watermark. When a Google model generates an image, video clip, or synthetic voice, it alters the file in a way that's imperceptible to you but readable by software. The technology has been active since 2023. According to Google, it has already watermarked over 180 billion images and videos, plus 240,000 years of audio.

The verification feature was already integrated into Search, the Gemini app, and Chrome, handling over a million queries daily according to Google's own data. What launched on October 7 is the standalone public hub: synthid.com, available worldwide and currently only in English.

What the Detector Checks For

The key thing to understand is that it doesn't look for general "signs of AI"; it looks for a specific watermark. That's why the list of participating companies matters more than the list of file formats.

Content from	Status
Google models (Gemini, image, video, and audio generation)	Detected
OpenAI	Detected
NVIDIA	Detected
Kakao	Detected
Apple Intelligence	Promised "soon"
Any other generator	Not detected

Formats accepted by the hub, according to Unite.AI:

Type	Formats
Images	JPG, PNG, WEBP, HEIC, and others
Video	MP4, MOV, WEBM
Audio	WAV, MP3, OGG, FLAC, AAC, M4A

The result is binary: watermark found or not found. According to FourWeekMBA, the public version doesn't show where in the file the watermark is located; that fragment-level view was available in the version for researchers. The same outlet reports that you need a Google, OpenAI, or Apple account to log in, and the free limit is around ten checks per user per day.

Why the Watermark Survives Editing

This is the part creators often underestimate. SynthID doesn't live in the file's metadata, which is easily wiped out during recompression. The watermark is embedded within the pixels or the sound wave itself, so it withstands:

Cropping, including reframing to a vertical aspect ratio;
Filters and color correction;
Frame rate changes;
Lossy compression, including the compression applied by platforms on upload;
For audio: added noise, MP3 recompression, and speed changes.

In practice, if you take a background generated with Gemini, add your own voiceover, cut it down to 30 seconds, and post it to Shorts, that background will still likely test positive.

What It Doesn't Tell You

Google explicitly states that this is not a general-purpose AI detector. The limitations:

A lack of a watermark proves nothing. If the generator isn't part of the SynthID initiative—and most aren't—the file will come up clean. A negative result doesn't mean a camera was involved.
It doesn't distinguish between "fully generated" and "lightly touched up." The watermark indicates that the file was processed by a participating model. It doesn't clarify whether it was an entirely synthetic scene or your real clip with an AI-expanded sky.
It doesn't say which model created it. It tells you a watermark is present, not who put it there.
Screen recordings and re-photographing can break the signal. The watermark withstands file processing, but it's not a forensic-grade tool.

Alongside SynthID is C2PA, the Content Credentials standard: an open-source record attached to the file that describes how it was made and edited. They are complementary, not rivals. That record provides detail but is easily removed; the invisible watermark offers no detail but survives a full editing workflow.

What This Means for Your Work

Platform rules already require you to disclose realistic AI-made video. TikTok, YouTube, and Meta ask you to declare it, and their systems also read source signals in the file to apply the label themselves. Having an AI label isn't a punishment. The penalty comes when it's missing: if you don't add it and the platform detects the generation, it can limit the video's reach or take it down, your account gets a strike, and repeat offenses can lead to demonetization.

1. Sort Your Tools into Two Buckets

Know which of your generators embed SynthID—Google, OpenAI, NVIDIA, Kakao—and which don't. This isn't a reason to ditch the first group. It's a reason not to pass that material off as camera-shot footage.

2. Check the AI Box Yourself

TikTok, YouTube, and Instagram all have an "AI-generated" toggle. Using it won't hurt your reach. A label added later by the platform looks like an attempt to conceal it, and on YouTube, the label becomes permanently attached to the video.

3. Vet Your Third-Party Assets

This is the everyday use case. A "real" clip of an event you found for a compilation, a stock footage pack you bought, raw material a client sent you: run it through synthid.com before you start editing. If it's watermarked, you'll know while you can still change your plan.

4. If You Buy UGC, Check It on Delivery

For a brand, it's a cheap intake filter: if the brief asked for a real person on camera and the file says a generator was involved, you have your answer. Keep the asymmetry in mind: a clean result doesn't confirm anything. It's better to specify in the brief whether AI material is allowed than to rely on detection.

5. Don't Argue with Screenshots

"No watermark found" doesn't prove you filmed it, and a "Watermark found" result doesn't prove the entire scene is fake. Understanding that distinction is more valuable than any screenshot of the results page.

What We Still Don't Know

When other interface languages will be added, as the tool is currently English-only.
When Apple Intelligence will be fully integrated, beyond the promise of "soon."
Whether other major video generators will adopt SynthID, which is crucial for broader coverage.
Whether platforms will start reading SynthID directly on upload to apply labels automatically.
The official free usage quota, as Google has not published a specific limit.

Frequently Asked Questions

How can I tell if an image is made with AI?

Upload it to synthid.com to check for the invisible SynthID watermark. It works with images, video, and audio created by models from Google, OpenAI, NVIDIA, and Kakao. This method won't work with other generators, as their content doesn't carry a watermark to be read.

Is it free, and do I need an account?

The detector is free, but you log in with a Google, OpenAI, or Apple account, and daily checks are limited to around ten per user, according to FourWeekMBA.

Can you remove the SynthID watermark from a video?

Not with normal editing: it withstands crops, filters, frame rate changes, and compression. There's not much point, either. Having the watermark doesn't violate any rules, but failing to disclose AI use to a platform does.

If the detector finds nothing, is the video authentic?

No, and this is the most common misinterpretation. A clean result only means the file doesn't carry the SynthID watermark. It could still have come from a generator that doesn't add one.

Does SynthID detect AI-written text?

No. The public hub works with images, video, and audio. Checking generated text is not part of its function.