News
Google SynthID Detector: What's New, Limits, and Who Should Use It
On this page6 sections
Google opened SynthID Detector to everyone on October 7, 2026. At synthid.com, anyone with an account can upload an image, video, or audio file and see whether it carries SynthID, an invisible watermark that AI tools from Google, OpenAI, NVIDIA, and Kakao add to what they make. Google says Apple will join "soon." No price is listed, and checks are limited per day.
For a small business owner, the new situation looks like this: a customer uploads a real photo of your work and gets "detected." That can happen, because retouching a photo with AI can leave the same watermark as generating one from scratch. A found watermark is evidence that a supported AI tool touched the file. It does not say how much was changed, by whom, or whether the scene is real, and a missing watermark proves nothing either. No file was tested for this article. Every fact comes from Google's, OpenAI's, and Meta's own pages and the portal's FAQ and terms.
SynthID Detector at a glance
- Opened to everyone: October 7, 2026; an early version had been limited to journalists and media professionals since 2025
- Checks: images, video, and audio; text is not supported
- Recognizes: watermarks from Google AI, OpenAI (images since May 19, 2026, audio since July 31, 2026), NVIDIA, and Kakao; Apple "soon," with no date
- Access: an account is needed, no price is listed, and checks are limited by a daily quota whose size the portal does not publish
- Result: detected or not detected; for video and audio, the portal reports which segments carry the watermark
What is new in SynthID Detector?
1. One public check for several companies' watermarks
Google's checks built into the Gemini app recognize only content made with Google AI, according to Gemini Apps Help. The portal also reads the partners' watermarks, so one upload covers all of them. OpenAI began adding SynthID to images from ChatGPT, Codex, and its API on May 19, 2026, and runs its own checker, openai.com/verify, which looks only for OpenAI signals. Google has not said which NVIDIA or Kakao products carry the watermark. Google says it has watermarked more than 180 billion images and videos since 2023.
2. A durable watermark, with limits
The portal's FAQ says the watermark is designed to survive filters, color changes, and lossy compression, and that in very rare cases the detector can trigger on a file that was never watermarked. OpenAI's Help Center gives the same picture for its images: the watermark is designed to withstand light edits, cropping, and screenshots, while heavy cropping or file compression can make it undetectable. Neither the portal nor OpenAI's Help Center gives detection rates or a false-trigger rate. Files from tools that do not use SynthID, or made before a tool adopted it, will not be flagged at all.
3. An account and a daily quota
The portal's terms forbid automated bulk checks, extra accounts to get around the quota, and using results to weaken or bypass the watermark. Its privacy notice ties the daily quota to the user's account and mentions signing in with a third-party account such as Apple, but it does not say how many checks a day are allowed. The number Google does publish is for the Gemini app: about 10 image checks, 10 video checks (up to 5 minutes in total), and 10 audio checks (up to 3 hours in total) in any 24-hour window.
What SynthID Detector changes for a small business
The same reading of a result applies to every case below.
| Result | What it tells you | What to do next |
|---|---|---|
| Detected, on an image | A supported AI tool likely generated or edited at least part of the file. It does not show whether the whole image was generated or only the background changed, who made it, or whether the scene is real. | Treat it as one signal. Ask for the original file or more photos before deciding. |
| Not detected | No supported watermark was found. The file may be real, or it may come from a tool without SynthID, an older file, or a heavily edited one. | Judge by source and context. A reverse image search can show whether a photo was taken from somewhere else. |
| Detected, on audio or video | A supported AI tool likely produced the marked segments. It says nothing about the sender's intent. | Confirm through a channel you already trust before acting. |
A real photo of your work comes back "detected"
This is the case owners are most likely to meet. A real photo of a haircut, a cake, or a clean kitchen can return "detected" if an AI tool edited it. Google Photos has added SynthID to images edited with Reimagine in Magic Editor since February 2025, although very small edits may not be marked. Gemini Apps Help describes a detection as all or part of the content being created or edited by Google AI, and OpenAI's Help Center says its watermark does not show whether the model edited a photo the user uploaded, or how much it changed. A genuine photo with an AI-replaced background therefore gets the same result as a fully generated image. Keep the unedited original of every work photo you retouch with AI, so it can be shown if a customer asks what was changed.
Images you make for posts
A banner made in Gemini, or an image made in ChatGPT after May 19, 2026, carries the watermark, and a customer who saves it from a post can check it. The useful test is whether a person could take the image for your real product or your actual work. A generated pattern behind a sale announcement passes that test; a generated plate of food in a menu, or a generated hairstyle in a portfolio, does not, because the customer expects to receive what the picture shows. The Grow Your Business With AI course makes the same split in its lesson on creating campaign content: a real photo when customers need to see the actual product, team, place, or way of working.
Photos clients send you
A damage photo in a refund request, a reference photo for a booking, or a picture attached to a review can be checked in a minute. Upload the original file where possible: OpenAI advises against cropping or converting an image before checking, and Gemini Apps Help asks for screenshots to be cropped tightly around the image. "Not detected" is not the same as trustworthy. A real, unedited photo can still belong to someone else or show a different job, which is why Gemini Apps Help also suggests a reverse image search to find where a picture came from. The decision about a refund stays with you.
Voice messages about payments
The portal also checks audio, and voice output from ChatGPT and the OpenAI API has carried SynthID since July 31, 2026. If a voice message asks you to change the payment details on an invoice, a "detected" result says the voice was likely made with a supported AI tool, which a legitimate supplier might also use. A voice from a tool without SynthID passes unflagged. Either way, a call to a number you already have on file is the check that settles it.
Empacer is a voice-first AI business assistant for small service business owners. You say what needs doing, review the finished work, and decide whether it goes out.
Who should try it now and who can wait
- You post AI-made or AI-retouched images: try it now on a few recent posts to see what a customer would see, keeping the daily quota in mind.
- You receive refund or damage claims with photos: try it on the next disputed photo, using the original file, as one signal among several.
- You post only unedited camera photos: nothing to change in how you post; the portal is still useful for files that clients send.
- You make images with Apple's AI tools: do not count on the portal to flag those images until Apple support arrives. Checking incoming files works the same for you as for anyone.
What is published and what is not
| Question | Published by Google, OpenAI, or the portal | Not stated in those sources |
|---|---|---|
| Daily limit on synthid.com | A daily quota tied to the account (portal privacy notice) | The number of checks per day |
| Accounts | Third-party sign-in such as Apple (portal privacy notice) | The full list of accepted accounts |
| Apple support | Coming "soon" (Google) | A date |
| NVIDIA and Kakao | Named as partners (Google) | Which of their products add the watermark |
| Accuracy | Rare false triggers are possible; heavy edits can remove the watermark (portal FAQ) | Detection rates and false-trigger rates |
FAQ
Can SynthID Detector check text, such as a caption or a review?
No. The portal accepts images, video, and audio only. OpenAI has watermarked ChatGPT text in the EU since October 5, 2026, to comply with the EU AI Act, but its text detector is open only to approved organizations, so there is no public way to check a caption this way.
Are files uploaded to synthid.com stored?
According to the portal's privacy notice, the original file is deleted as soon as the result is returned. To enforce daily quotas, Google keeps a digital signature of the file, linked to the account, for 24 hours and then deletes it.
Do Instagram and Facebook use SynthID to label posts?
Meta's labeling policy does not mention SynthID. The "AI info" label relies on industry-shared signals or the uploader's own disclosure, and since September 2024 content that AI only edited shows the label in the post's menu, according to Meta's labeling policy.
How can I show that a photo of my work is real?
Keep the original file from the camera. Some phones go further: on the Pixel 10, Pixel Camera attaches Content Credentials to every JPEG photo it takes, and Google Photos adds them to images edited with AI tools, according to Google. OpenAI's Help Center notes that this kind of metadata can be lost through a screenshot or a file conversion, so the original file matters.