SynthID is now the layer doing the real work of identifying AI content. On 14 August 2026, Google announced that Gemini users can disable the visible watermark on AI-generated images, videos and songs — a change we unpacked in our guide to the Gemini watermark removal announcement. The corner badge is now optional; the invisible SynthID watermark and C2PA metadata are not.
That shift changes the practical question for everyone else. When the visible mark was mandatory, spotting Google’s AI output took nothing more than eyesight. Now the honest answer to “was this made by AI?” lives inside the file, and you need to know how SynthID works, where it is embedded, and which tools read it back.
This guide covers exactly that: what SynthID is, how it hides a signal in pixels, audio and text, the scale it has reached, how to check any file for it, what it survives, where it fails, and a verification workflow your team can adopt now that the visible badge no longer does the job for you.
Table of contents
- What SynthID Actually Is
- How the Invisible Watermark Is Embedded
- Watermarking at Scale: the Numbers
- How to Check a File for SynthID
- What the Invisible Mark Survives — and What Beats It
- SynthID After the Gemini Watermark Toggle
- A Practical Verification Workflow for Your Team
- Limits and Open Questions
- FAQ: SynthID and AI Content Verification
- References
What SynthID Actually Is
SynthID is an invisible watermarking system built by Google DeepMind. It embeds a machine-readable signal directly into AI-generated content at the moment of creation — not a badge drawn on top, but a pattern woven into the media itself, imperceptible to people and recoverable by detection software.
Where it came from
Google DeepMind launched SynthID in 2023, initially for images from its Imagen model family. It expanded to audio, video and text over the following two years. By Google I/O in May 2025, more than 10 billion pieces of content already carried the SynthID mark, and Google has kept it on by default across its generative products ever since.
What it covers today
Within Google’s ecosystem, SynthID spans all the model families behind Gemini’s media generation: the Nano Banana image models, the Omni video model, the Lyria music family and Gemini’s text output. There is no user-facing switch for any of them — the toggle Google shipped in August 2026 controls only the visible badge.
A signal, not a sticker
The distinction matters. A visible watermark is a design element any editor can crop or inpaint away in seconds. SynthID is a statistical property of the content: every region of an image, every second of audio, carries part of the signal. Removing it is not an edit; it is a reconstruction — which is why Google treats SynthID, not the badge, as the load-bearing transparency layer.
How the Invisible Watermark Is Embedded
Each media type gets its own embedding technique, tuned so the watermark changes nothing a human would notice while remaining readable to a detector.
Images and video
For images, the system makes tiny, coordinated adjustments to pixel values across the whole frame — shifts far below the threshold of human vision, arranged so a trained detector can recognise the pattern with high confidence. Video works frame by frame on the same principle. Because the signal is spread across the entire image rather than sitting in one corner, cropping out “the watermark” is not possible in the way it was with the visible badge.
Audio and music
For audio, including songs from the Lyria family, the pattern is embedded into the spectrogram — the frequency representation of the sound — before converting it back into a waveform. The result is inaudible to listeners but survives common transformations such as compression and re-encoding, which matters for music that will be uploaded, streamed and re-shared across platforms.
Text
Text is the hardest medium, because words cannot be nudged invisibly the way pixels can. Text is watermarked at generation time by steering the model’s word choices: where several tokens are similarly probable, the sampler favours a pattern that a detector can later test for statistically. Google open-sourced this text approach through the Hugging Face Transformers library in 2024, letting any developer apply it to their own models.
Watermarking at Scale: the Numbers
The scale is the strongest argument that SynthID, rather than any visible badge, is the backbone of AI content identification. At I/O in May 2025 Google reported more than 10 billion watermarked pieces of content. One year later, at I/O on 19 May 2026, the figure for images and videos alone had passed 100 billion — a tenfold rise — alongside roughly 60,000 years of watermarked audio.
Adoption beyond Google
The watermark stopped being a Google-only system in May 2026, when OpenAI, Kakao and ElevenLabs announced they would incorporate it into their products. Images generated by ChatGPT now carry the mark going forward, alongside OpenAI’s existing C2PA metadata. OpenAI’s stated reasoning captures the division of labour: watermarking is more durable through transformations like screenshots, while metadata can carry more information than a watermark alone.
| Organisation | Where SynthID is applied | Since |
|---|---|---|
| Nano Banana images, Omni video, Lyria music, Gemini text | 2023 onwards, on by default | |
| OpenAI | ChatGPT image generation, alongside C2PA metadata | May 2026 |
| ElevenLabs | AI audio and voice products | Announced May 2026 |
| Kakao | Generative AI products | Announced May 2026 |
How to Check a File for SynthID
Detection used to be the missing half of the story: an invisible mark is only useful if someone can read it. Through 2026 Google has shipped checking tools for ordinary users, not just researchers. Counting them up: Chrome’s right-click check, Circle to Search, and the Gemini app’s credential view are live for the public today, with the SynthID Detector portal in limited access — three of the four routes fully open.
Right-click in Chrome or Circle to Search
At I/O on 19 May 2026, Google announced SynthID and C2PA verification built into Chrome and Google Search. As CEO Sundar Pichai put it, users can “simply circle to search an image, or right click in Chrome and ask, ‘Was this generated with AI?'” The browser checks the file for the SynthID signal and its content credentials and reports what it finds — no upload, no separate tool.
The SynthID Detector portal
For deeper checks, the SynthID Detector is a standalone portal announced in May 2025, where you upload an image, video or audio file and the tool highlights which portions carry the watermark — useful when only part of a file is AI-generated. Access began with journalists, media professionals and researchers, and a public waitlist is open via Google’s announcement page.
Content credentials in the Gemini app
The Gemini app itself can now read C2PA Content Credentials, letting users inspect the signed provenance record attached to a file from inside the assistant. For files from any source, the Content Authenticity Initiative’s public Verify tool reads the same records in a browser.
Building checks into your own software
Alongside the August 2026 watermark announcement, Google open-sourced Credentio, a C2PA library for validating content credentials locally inside your own applications. Paired with the open-source SynthID text library, a development team can now build provenance checking into a review pipeline without sending files to a third party.
| Route | Where it lives | Who can use it | Best for |
|---|---|---|---|
| Right-click check | Chrome browser | Everyone | Images found while browsing |
| Circle to Search | Google Search on mobile | Everyone | Quick checks on a phone |
| SynthID Detector portal | Standalone upload site | Journalists, researchers, waitlist | Detailed, region-level analysis |
| Gemini app credentials | Gemini app | Everyone | Reading C2PA provenance records |
| Credentio + open-source libraries | Your own applications | Developers | Automated pipeline checks |
What the Invisible Mark Survives — and What Beats It
No watermark is unbreakable, and Google has never claimed otherwise. The honest picture is a spectrum: the watermark shrugs off the edits ordinary users make, degrades under aggressive transformation, and cannot say anything at all about content from models that never embedded it.
Everyday edits: the strong ground
Because the signal is distributed across the whole file, the mark survives the transformations that defeated the visible badge: cropping, resizing, compression, filtering, colour changes and — for images — screenshots. This durability through screenshots is precisely why OpenAI paired the watermark with metadata rather than relying on metadata alone, since C2PA records are stripped by a simple screenshot while the in-pixel signal persists.
Aggressive transformation: the grey zone
Heavy editing erodes the signal. Substantially repainting an image, re-recording audio through a speaker, or extreme compression can degrade the embedded pattern until a detector’s confidence drops. For text the grey zone is wider still: light edits leave the statistical pattern detectable, but thorough paraphrasing or translation largely destroys it. A negative result is therefore evidence, never proof.
The structural blind spot
SynthID can only mark content from models that apply it. Output from open-weight models a user runs locally, or from labs that have not adopted the system, carries no signal — and never will. Detection tells you “this was made by Google’s or a partner’s AI”; it cannot tell you “this was not made by AI”. That asymmetry is the single most important thing to teach anyone doing verification work.
| Transformation | Does the mark survive? |
|---|---|
| Cropping, resizing, rotation | Yes — signal spans the whole image |
| Compression and re-encoding | Yes, within normal limits |
| Screenshots of images | Generally yes — unlike C2PA metadata |
| Filters and colour adjustment | Usually yes |
| Heavy repainting or re-generation | Degrades — confidence drops |
| Thorough paraphrase or translation of text | Largely no |
| Content from non-adopting models | Never marked in the first place |
SynthID After the Gemini Watermark Toggle
The August 2026 change did not touch the invisible layer, but it rearranged the responsibilities around it. With the visible badge now a setting — off by choice in most countries, and gated to Google AI Ultra subscribers in India, South Korea and Vietnam — the invisible watermark stops being a backstop and becomes the primary identifier.
What changed on 14 August 2026
Josh Woodward, Google’s VP for Gemini, announced that the visible watermark on AI images, videos and songs becomes a user-controlled setting, rolling out in the Gemini app and Flow first with Search to follow. The setting defaults to on. Crucially, Google confirmed the SynthID watermark and C2PA metadata stay embedded in every generation with no user-facing switch of any kind.
Why the invisible layer now carries the weight
Before the toggle, a missing badge meant “probably not Google AI”. After it, a missing badge means nothing — the file may simply come from a user who switched the badge off. Every workflow that quietly leaned on the visible mark, from newsroom triage to marketplace moderation, now needs a detection step that actually reads the file. The viewer’s free clue is gone; the durable signal remains.
The regulatory backdrop
Machine-readable marking is also what the law increasingly asks for. Article 50 of the EU AI Act requires providers to mark synthetic content in machine-readable form — a requirement satisfied by SynthID and C2PA rather than by any particular badge, as we covered when the EU moved to mandate labels on authentic-looking AI content. Platforms are converging on the same posture: TikTok already reads C2PA metadata to auto-label uploads, alongside its AI likeness detection tools.
A Practical Verification Workflow for Your Team
For most organisations the question is no longer whether to verify AI content but how to make checking routine. Provenance checks are becoming standard practice in cybersecurity operations, communications teams and editorial workflows alike, and the tools above make a lightweight process realistic.
The three-step check
First, read the credentials: right-click in Chrome, use Circle to Search, or run the file through the Content Credentials Verify tool to see any signed C2PA record. Second, test for the watermark through Google’s tools, or the Detector portal for a region-level view. Third, interpret with the asymmetry in mind — a positive result is strong evidence of AI origin; a negative result only rules out marked sources.
When to escalate
Escalate when the stakes are high and the checks disagree with the story around the content: a supposedly live photo with an AI provenance record, an executive’s voice with no credentials from an unknown number, a viral clip whose earliest copy fails every check. For those cases, pair provenance tools with traditional verification — source contact, reverse image search, corroboration — because the blind spot means technology alone cannot close the question.
Write it into your guidelines
Two policies are worth adding now. Outbound: decide whether your organisation keeps the visible badge on its own AI-generated media, and record that the invisible SynthID mark ships regardless — your published assets are permanently identifiable as AI output, badge or no badge. Inbound: name the tools staff use to check third-party content and log what was checked. Teams that already produce AI media at scale, from personalised Gemini image generation to full video, should treat this the way they treat brand safety: written down, owned, and reviewed.
Limits and Open Questions
A clear-eyed adopter should know where the system is still weak, because the gaps define how much trust the mark can carry.
Coverage will never be universal
Open-weight models running on local hardware answer to nobody’s watermarking policy. However many labs adopt SynthID, a determined bad actor can always generate unmarked content. The system raises the cost of casual deception and marks the mainstream tools honestly — it does not, and cannot, catch everything. Verification culture has to absorb that limit rather than pretend detection is a solved problem.
The removal arms race
Academic work has repeatedly shown that watermarks of every kind can be attacked, and a cottage industry of removal services already targets visible and invisible marks alike. Google’s counter is scale and secrecy: the detector models are not public, and the signal design keeps improving. Expect this to remain a moving contest rather than a settled outcome.
Who guards the detector
Concentrating detection in the hands of the same company that generates the content raises a fair governance question: outsiders must trust Google’s detector verdicts without being able to audit them fully. Open-sourcing the text watermarking and the Credentio C2PA library moves some checking into public hands, and the C2PA standard itself is multi-vendor — but for images, audio and video, SynthID detection remains Google’s word.
FAQ: SynthID and AI Content Verification
Does turning off the Gemini visible watermark remove SynthID?
No. The toggle Google shipped on 14 August 2026 controls only the visible badge. The invisible watermark and C2PA metadata are embedded in every generation from the Nano Banana, Omni and Lyria model families, and there is no setting, paid or free, that removes them.
Can I check content for SynthID myself?
Yes. Right-click an image in Chrome and ask whether it was generated with AI, use Circle to Search on mobile, read content credentials in the Gemini app, or check a file with the public Content Credentials Verify tool. The dedicated SynthID Detector portal offers deeper analysis, with access via waitlist.
Is SynthID only used by Google?
Not any more. OpenAI, Kakao and ElevenLabs announced in May 2026 that they are adopting it — ChatGPT-generated images now carry the mark going forward. Google also open-sourced the text watermarking method, so any developer can apply it to their own models.
Can SynthID be removed from a file?
Not by normal editing. It survives cropping, resizing, compression, filters and screenshots. Aggressive transformation can degrade it, and heavily rewritten text loses the pattern — which is why a negative detection result should be treated as inconclusive rather than as proof a file is human-made.
What does a “no watermark found” result actually mean?
It means no SynthID signal was detected — nothing more. The file could be human-made, could come from a model that never embeds the mark, or could have been transformed enough to destroy the signal. Treat detection as one input to a verification decision, not the decision itself.
References
SynthID Detector — a new portal to help identify AI-generated content
Google will now allow users to remove visible watermark from its AI generations — TechCrunch
Google SynthID comes to Chrome, Search, and ChatGPT — Yahoo Tech
Coalition for Content Provenance and Authenticity (C2PA)
Content Credentials Verify — Content Authenticity Initiative
SynthID Text watermarking in Transformers — Hugging Face
EU Artificial Intelligence Act, Article 50 — Transparency Obligations
