SynthID Detector. A Watermark Hit Is Not a YouTube Label

SynthID Detector. A Watermark Hit Is Not a YouTube Label

On 7 October 2026 Google DeepMind opened SynthID Detector to everyone, in English, as a standalone check for whether an image, video, or audio file carries a SynthID watermark from Google or named partners. That is a provenance desk. It is not a YouTube Studio “AI use” toggle, a signed voiceover, or proof the file is the clip you meant to ship. ClipCanva is not affiliated with Google, YouTube, VEED, or Kapwing. Lock the spoken line on the AI script generator. Steal shot language from prompt ideas or the product-ad prompt pack. Lock the still, then animate it on image to video or generate picture on the AI video generator. Do not upload take one because Detector said “watermark found.”

Product facts (confirmed 9 October 2026)

Source	Public claim	Workflow implication
We're making it easier to identify AI-generated content globally (7 Oct 2026)	Standalone platform. SynthID launched 2023. Imperceptible watermarks on images, video, and audio. Over 180 billion images and videos watermarked, plus 240,000 years of audio. Early Detector was for media professionals; today it is global, English. Anyone can check if a file was made with AI from Google or partners including OpenAI, NVIDIA, Kakao, and soon, Apple. Joins built-in checks in Search, the Gemini app, and Chrome, which Google says now handle over 1 million requests daily.	A file check. Not a platform disclosure. Apple is soon, not now.
SynthID Detector	Live Detector hostname linked from the 7 Oct post.	Use the live form. The 2026 post does not republish click-by-click UI, quotas, or file-size limits. Do not brief a 2025 tester screenshot as today’s product.
SynthID Detector: Identify content made with Google’s AI tools (20 May 2025, I/O)	Early portal for testers. Upload image, audio, video, or text; scan for a SynthID watermark; highlight likely watermarked portions. Note on that page: product in development; video and text detection will be rolled out in the coming weeks.	Treat as history. The 7 Oct 2026 post names image, video, or audio files. It does not restated text detection. Do not copy 2025 “coming weeks” into a 2026 brief.
Disclosing use of GenAI content (YouTube Help)	Creators must disclose GenAI that: makes a real person appear to say or do something they didn’t; alters footage of a real event or place; generates a realistic scene that didn’t occur; or creates music that’s the main focus of the video. Studio path: Attributes → AI use → Yes/No. Realistic / meaningful changes require disclosure; non-realistic or minor edits do not. Production assistance that Help lists as not required: outline, script, thumbnail, title, infographic, captions, upscaling, idea generation, cloning one’s own voice. YouTube may auto-label content from YouTube’s GenAI tools, C2PA metadata, or internal detection. Labels from YouTube AI tools, C2PA, or manual review cannot be adjusted. Disclosure does not limit audience or monetization. Repeated non-disclosure can mean a manual label, removal, or Partner Program suspension.	Disclosure is a YouTube Studio desk. Detector is not that toggle.
VEED AI Video Script Generator	Own URL. Scripts from a prompt for social, explainers, marketing. Separate AI Video Generator.	Writer URL stays separate from renderer. A VEED.IO logo watermark on a free export is not SynthID.
Kapwing AI Video Script Generator	Own AI nav item. Separate AI Video Generator for text, images, and scripts.	Kapwing still splits script and video. Neither page is a YouTube “AI use” form.

Google did not publish, on the 7 October post, Detector file-size limits, a public API for the consumer portal, a price, or a statement that a miss means the file is camera-original. If a capability is not on the page you are briefing from, it is not a brief.

What Detector is — and what it is not

Operators will collapse four objects into “Google said the watermark is public, so the YouTube label is done.” The public pages refuse that.

Surface	What the public page says it does	What it does not do	Fail condition
SynthID Detector (7 Oct 2026)	Check an image, video, or audio file for a SynthID watermark from Google or named partners	Sign the claim, set YouTube AI use, prove a miss is human-shot	Shipping because Detector returned a hit, or skipping disclosure because it returned a miss
Built-in checks in Search / Gemini / Chrome	Context while you browse; Google cites 1 million+ daily requests	Replace Studio Attributes, or scan a file that never touched those surfaces	Pasting a Search “about this image” screenshot into Legal as the YouTube disclosure
Partner list (OpenAI, NVIDIA, Kakao; Apple soon)	More generators may embed SynthID	Cover every model on Compare, or Apple Intelligence today	Briefing “all AI video is now detectable”
YouTube AI use (Help 14328491)	Creator-set label for realistic / meaningful GenAI; auto-label via YouTube GenAI tools, C2PA, or internal systems	Read a SynthID pixel watermark unless YouTube separately detects it	Treating Detector as the Studio toggle
Script / outline assistance	Help: not required to disclose	Become the photorealistic clip	Generating a realistic restock scene and calling it “just a script”
VEED / Kapwing script tools	Draft hooks and CTAs on their script URLs	Disclose on YouTube or watermark the master	Hitting generate on the unsigned draft, then arguing Detector will sort the rest

If you are uploading a file to synthid.com, you are on a check desk. If YouTube Help says the scene is realistic and generated, you are on a disclosure desk. If the bottle still needs a legal sentence, you are on a writer. If you cannot tell which, do not publish.

Pick the desk before you upload the master

The fail is treating “Detector is public” as “the channel is covered.” Map the job.

Job	Where it actually runs	What you must supply	What the system may invent	Fail condition
One claim the clip must carry	AI script generator — or VEED / Kapwing if that is already your writer	Audience, seconds, the signed sentence	A second CTA, a health claim, a competitor name	Skipping the card because “we’ll watermark it later”
Recap of a long cut	AI video summarizer	Source cut, claims to keep, claims to strike	Compression that changes meaning	Summarizing a webinar into a realistic “CEO said” scene you never recorded
Still that must stay on-model	Stills desk, then image to video	Approved pack shot as frame one	Camera path, warped type, a second SKU	Checking Detector on a still that Legal never signed
Motion from a locked still	AI video generator or a named model on Compare	Locked still and/or locked caption, crop, seconds	Extra props, a spoken line you did not approve	Generating first, checking Detector second, disclosing never
SynthID check	synthid.com, if you need a Google/partner watermark read	The export you will actually upload, not a proxy screenshot	A miss on a model that never embeds SynthID	Reading a miss as “this is documentary footage”
YouTube disclosure	YouTube Studio Attributes → AI use, per Help	Honest Yes/No against the realistic-scene / real-person / real-place / main-focus music tests	Auto-label from YouTube GenAI tools, C2PA, or internal systems that you cannot clear if C2PA or YouTube tools applied it	Assuming Detector already filed the form

If the brief is “see whether Google’s watermark is on this export,” that is R&D. If the brief is “the two-pack restocks Friday on YouTube,” that is script, stills, generate, then Studio. If the brief is “Detector will handle disclosure,” you do not have a brief.

VEED still keeps AI Video Script Generator as its own URL, then offers generate-video after the draft. Kapwing still keeps Script Generator and AI Video Generator as different URLs. VEED’s help article about a VEED.IO logo on free exports is a branding watermark, not an invisible provenance mark. None of those pages say a Detector hit is a YouTube label.

Creator / operator checklist

1. Read the 7 October line out loud. Global, English, image / video / audio, Google or named partners, Apple soon. If the next sentence in Slack is “AI video is now labeled automatically everywhere,” stop.
2. Name the surface. Detector ≠ Search/Gemini/Chrome overlay ≠ YouTube AI use ≠ C2PA in the file header. Brief the picker you can screenshot.
3. Cap the spoken line before you generate. Use the AI script generator. One sentence. Strike “#1,” ingredient claims, and competitor names you cannot defend. Native audio will say whatever the prompt implied, including the lie. Help does not require disclosing a script. Help does require disclosing a realistic scene that did not occur.
4. Decide realistic vs not before you open Detector. Unicorns, fully animated missiles, beauty filters, color, captions, upscaling, and idea generation are on Help’s don’t have to disclose list. A photorealistic restock in a real kitchen is not.
5. If music is the point, read the music row. Help now lists music that’s the main focus of the video as something creators must disclose, and AI generated music in the examples. A lyrics draft on the AI lyrics generator is still a writer object. The upload is not.
6. Check the export, not the proxy. Run Detector on the file you will upload. A compressed Slack clip is a different object.
7. A hit is not a ship. It means this file looks like it carries SynthID from Google or a named partner. It does not mean Legal signed the line, the bottle is this year’s SKU, or Studio is filled in.
8. A miss is not a documentary. Partners are a list, not the whole model shelf. Apple is soon. A generator that never embeds SynthID will miss. Absence of a watermark is not camera-original.
9. File Studio separately. Attributes → AI use → Yes or No. Auto-labels from YouTube’s own GenAI tools, C2PA, or manual review cannot be cleared by arguing with Detector.
10. Bill the right object. A script, a still, a generate, a Detector check, and a YouTube disclosure are five things. Do not hide all five under “we ran SynthID.”

Worked example: a Friday restock that does not start in Detector

Say the job is a YouTube restock cutdown. The two-pack ships Friday. Legal line approved: Same bottle. Same seal. Restock Friday. The hero still exists. Someone read the 7 October DeepMind post, saw “anyone can easily check,” saw OpenAI / NVIDIA / Kakao, saw “soon, Apple,” and told the channel “Detector is public, so the AI label is covered.”

That sentence mixes five desks.

The spoken line is a writer object. Put the approved sentence on the AI script generator. YouTube Help says a script is production assistance you do not have to disclose. The photorealistic pour in a real kitchen is not a script. If you already live in VEED or Kapwing, their script tools are still script tools. Copy the signed line. Do not hit generate from the unsigned draft.

The beat list is a card object. Steal shot language from prompt ideas or the ads and Shorts prompt pack, then write four lines, not a chat novel:

• 0s — kitchen, bottle already on the board (do not invent a second SKU)
• 2s — hand in, label readable
• 4s — pour or use, same bottle
• 6s — end card with the signed sentence only

The still is a frame object. If the bottle was shot, you are in image to video from that frame. If you still need picture, generate one scene on the AI video generator. Pick a model on Compare because Detector’s partner list is not your model picker.

The Detector pass is a check object. Upload the export to synthid.com. Log hit / miss / inconclusive. Do not rewrite the Legal line from the result. Do not skip Studio because of a hit. Do not call a miss “filmed.”

The YouTube pass is a disclosure object. Studio → Attributes → AI use. This restock is a realistic scene that did not occur on that counter this morning. Select Yes. Help says that label does not, by itself, limit audience or money. Help also says repeated non-disclosure can cost the Partner Program. Detector did not file that form for you.

If someone then asks “did we check C2PA,” that is a third provenance channel. YouTube may auto-label files that contain it, and Help says you cannot clear that label. Do not strip metadata as a workaround. Do not invent a C2PA status you did not read.

FAQ

Does a SynthID hit mean I can skip YouTube’s AI use toggle? No. Detector is Google’s watermark reader for Google and named partners. YouTube Help is a separate creator disclosure in Studio. Run both when the clip is realistic generated media.

Does a SynthID miss mean the clip is real? No. The 7 October post lists Google plus OpenAI, NVIDIA, Kakao, and Apple soon. A miss can mean “this generator never wrote SynthID,” “the file was transformed past detection,” or “you checked the wrong export.” It is not a camera report.

Do I have to disclose an AI script? YouTube Help lists generative tools used to create or improve an outline, script, thumbnail, title, or infographic as production assistance that creators don’t need to disclose. The photorealistic video you generate from that script is a different object.

What about AI music? Help requires disclosure when you create music that’s the main focus of the video, and lists AI generated music among examples that need disclosure. A lyrics draft is still a writer object. Confirm the live Help page on the day you upload.

Will Search, Gemini, or Chrome label my YouTube video for me? Google says those surfaces now handle over a million verification requests a day. That is browse-time context. It is not the Studio Attributes survey, and it is not Detector’s file upload.

Do not mix these objects

Object	Tool	Done when
Signed line	AI script generator	One sentence Legal will defend
Shot card	Prompt ideas	Beats with times, no extra SKU
Locked still	Photography or stills desk	Label readable, type unwarped
Motion	Image to video or AI video generator	One scene, approved audio
Watermark check	SynthID Detector	Hit / miss logged on the export
Platform disclosure	YouTube Help: GenAI content	AI use set; C2PA/YouTube-tool labels left alone

Google made the watermark check public. It did not make the brief public. Write the line, lock the frame, generate the scene, then decide Detector and Studio as two different desks.