Google SynthID Detector: How to Check If an Image, Video or Audio Is AI-Generated (Step-by-Step Guide)

Every day, millions of people scroll past images, videos, and voice clips that look completely real — and a growing share of them were never captured by a camera or recorded by a microphone at all. If you have ever wondered how to check if an image is AI-generated before you believe it, share it, or spend money because of it, Google's SynthID Detector is the most direct answer available today: a public portal, opened worldwide in English on October 7, 2026 at synthid.com, that reads the invisible watermarks Google and its partners embed in AI-made media and tells you, file by file, whether one of those watermarks is present. This guide walks you through exactly what the tool is, how to use it step by step, and — just as important — what its results do and do not prove.

Why Everyone Suddenly Needs an AI-Media Checker

For most of the internet's history, seeing was believing. A photograph was evidence that something happened; a recorded voice was evidence that someone said something. AI generation has quietly dismantled that assumption. Somewhere around 2025 and 2026, generated images, video, and audio crossed the threshold where even careful, skeptical viewers can no longer tell them apart from the real thing by looking or listening. The fakes are not confined to research labs anymore — they are in social feeds, marketplaces, inboxes, and group chats, every single day.

The most expensive version of this problem is fraud. Scam operations now routinely use AI-cloned voices and faces of well-known public figures in fake advertisements for miracle products and investment schemes. The endorsement never happened, the person never said those words, but the video looks and sounds convincing enough that thousands of people hand over money. Voice cloning has also powered a wave of "emergency" phone scams, where a caller using an AI-cloned voice of a relative pleads for an urgent money transfer. None of these existed at this scale five years ago.

The second version is viral misinformation. AI-generated images of dramatic scenes — a flooded street, a burning building, a celebrity caught in a scandal — are shared as if they were news photographs, and they routinely travel further and faster than the corrections that follow them. By the time a fact-checker publishes a debunk, millions of people have already seen the fake, believed it, and shared it. Romance scammers use AI-generated profile photos that cannot be reverse-image-searched back to a real person. Online marketplaces host listings with AI-generated product photos of items that do not exist.

What ties all of these together is a brutal asymmetry. The core problem is speed: a convincing fake takes seconds to create and hours to debunk, so the only practical defense is a quick, reliable check before you believe, share, or pay. Until recently, ordinary people had no direct way to run such a check on a suspicious file — the detection tools belonged to researchers, platforms, and newsrooms. The public launch of the SynthID Detector in October 2026 is the first time a major AI company has handed that checking power directly to everyone.

What SynthID Actually Is: Google DeepMind's Invisible Watermark

SynthID is Google DeepMind's technology for marking AI-generated media, launched in 2023. At its core, it is a watermarking system — but not the kind you are picturing. There is no visible logo stamped in the corner, no audible tone hidden in the audio. Instead, SynthID embeds an imperceptible watermark directly into the content at the very moment of generation: into the pixels of an AI-generated image or video frame, and into the frequency components of AI-generated audio. The mark travels inside the file itself, not in the metadata around it, which means it survives things like format conversions and re-uploads that would strip an ordinary metadata tag.

"Imperceptible" means exactly what it says. A watermarked image looks pixel-for-pixel identical to the human eye; watermarked audio sounds identical to the human ear. The watermark is data woven into the media in a way only software can read. That design choice matters because it lets platforms and creators mark AI-generated content without degrading it — there is no ugly overlay, no quality trade-off for the viewer or listener.

The scale is already enormous. To date, Google has watermarked over 180 billion images and videos, along with 240,000 years of audio content. Music is part of the system too: output from Lyria, Google's AI music model, carries a SynthID watermark, so AI-composed tracks can be identified as such. When hundreds of billions of files carry the mark, a public tool that can read it becomes genuinely useful rather than a curiosity.

Crucially, the watermark was engineered for the messy reality of the internet. SynthID is not a visible stamp or a metadata tag — it is data hidden inside the pixels and sound itself at the moment of creation, designed to survive the normal edits files go through online. Per Google DeepMind, the watermark is built to remain detectable after cropping, filters, compression, MP3 compression, speed changes, and added noise — the everyday handling that files endure as they are shared, reposted, and re-encoded across platforms.

What the SynthID Detector Portal Is — and Why October 2026 Changed Everything

The watermarking technology has existed since 2023, but for most of that time ordinary people had no way to check a file themselves. The technology lived inside Google's products and partner systems, readable mainly by the companies that embedded it. That changed in two stages. At Google I/O in May 2025, Google announced an early version of a SynthID Detector portal and offered it to journalists, media professionals, and researchers, who tested it behind a waitlist. For over a year, it remained an insiders' tool — useful to newsrooms, invisible to everyone else.

The second stage is the news hook of this guide. On October 7, 2026, Google opened the portal to the entire world: the SynthID Detector launched globally in English at synthid.com, free for anyone to use. The launch was announced by Pushmeet Kohli, VP of Science and Strategic Initiatives at Google DeepMind, in an official blog post dated October 7, 2026 and updated on October 29, 2026. The announcement framed it plainly: as AI tools become better and easier to use, people need clear context about the content they encounter online to make informed decisions about what to trust.

Why does a public portal matter so much? Because it is the first self-serve way for a regular person — not a journalist, not a researcher, not a platform engineer — to take a suspicious image, video, or audio file and get a direct, authoritative answer about whether it carries a SynthID watermark. Before this, your options were guessing, asking an expert, or waiting for a fact-check. Now the check takes a few minutes and you can do it yourself, as many times as you need.

You can open the portal directly at synthid.com, and you can read Google DeepMind's launch announcement at the official Google blog post. October 7, 2026 is the date this stopped being an insiders-only tool: the SynthID Detector portal at synthid.com opened to everyone, worldwide, in English.

The Critical Distinction: A Watermark Verifier, Not a Universal AI Detector

Before you upload a single file, you need to understand the single most important fact about this tool, because misunderstanding it is how nearly every bad conclusion happens. The SynthID Detector is a watermark verifier — it checks whether a file contains a SynthID watermark. It is not, and does not claim to be, a universal AI detector. Google states this limitation explicitly: the tool is "not a general AI detector." Those six words are the entire key to using it correctly.

A positive result is meaningful and strong. If the portal finds a SynthID watermark in your file, that is solid evidence the file was made or edited with a participating AI system. Watermarks do not appear by accident — they are embedded deliberately at the moment of generation by systems that opted into the SynthID program. A detection is not a guess from visual patterns; it is the discovery of an actual embedded signal. When the portal says a watermark is present, you can treat that as a reliable answer.

A negative result is a completely different animal, and this is where most people go wrong. "No watermark found" means exactly one thing: no SynthID watermark was detected in that file. It does not mean the file is real. It does not mean it is human-made. It does not mean it is authentic. There are at least three common reasons a genuinely AI-generated file can come back clean: it was made with a generator that does not use SynthID at all (and many of the most popular ones do not, as the next section explains); the watermark was removed or damaged by aggressive editing; or the file is genuinely real and never had a watermark in the first place. The portal cannot tell you which of these is true — it can only report the absence of the signal it was built to find.

The practical rule is simple once you internalize it. A "watermark detected" result is strong evidence a file came from a participating AI system — but a "no watermark found" result is not proof the file is real. Treat a detection as an answer and a clean result as the absence of an answer, and you will never be misled by this tool. Treat a clean result as a certificate of authenticity, and you will eventually be fooled by exactly the kind of fake the tool was built to catch.

What the SynthID Detector Covers — and What It Does Not

On the coverage side, the portal checks for watermarks from Google's own AI products and generators, plus the industry partners that have adopted SynthID for their outputs: OpenAI, NVIDIA, and Kakao. In practical terms, this means images, video, or audio generated with participating Google tools — and content from those named partners' systems — can be flagged. Google also names Apple as a partner, but Apple support is forthcoming, and Google has not confirmed a date for it, so do not expect Apple-generated content to be detectable yet.

The "does not cover" list is just as important, and it is explicit. The portal does not detect content from Midjourney, from Meta's AI tools, from Microsoft's AI tools, or from open-weight models. None of these embed SynthID watermarks, so their outputs are invisible to this checker by design — a file from any of them will come back with no watermark found, no matter how obviously AI-generated it looks to your eyes.

This gap matters enormously in practice because of where viral AI content actually comes from in 2026. A large share of the AI images and videos circulating on social platforms are made with exactly the tools on the not-covered list. That means the most common real-world scenario — "I saw a suspicious image, let me check it" — will very often end with a clean result that tells you nothing. The tool is authoritative within its coverage and silent outside it, and knowing which side your file falls on is part of using it well.

Coverage is also the dimension most likely to improve over time, as new partners adopt the watermarking standard. Before you trust a clean result, ask which tool made the file: if it came from Midjourney, Meta, Microsoft, or an open-weight model, the SynthID Detector was never designed to flag it.

How to Check a File in the SynthID Detector Portal, Step by Step

The check itself is straightforward and takes only a few minutes from start to finish. You can verify any file you have received, downloaded, or saved — and if you generate media with free AI tools yourself, you can upload your own creations and watch the watermark get detected firsthand, which is the fastest way to build intuition for how the tool behaves. Our guides to creating AI images for free and free AI video generators are good starting points if you want to run that experiment. Here is the full process:

1. Open the portal. Go to synthid.com in your web browser. The portal launched in English, so the interface you see will be in English regardless of your location.
2. Sign in. Per published reports, the portal asks you to sign in before you can upload files, accepting a Google, OpenAI, or Apple account. Google has not published full official documentation of the sign-in flow, so if the options you see differ slightly from reports, follow what the portal itself presents.
3. Choose the media type. The portal offers separate cards for Image, Audio, and Video. Pick the one that matches the file you want to check — selecting the right type ensures the file is analyzed with the correct watermark reader.
4. Upload your file. Files of up to 100MB are accepted. Make sure your file is in one of the supported formats listed in the table below; an unsupported format will not be analyzed.
5. Read the result. Per published reports, the portal reports whether a SynthID watermark was detected in all or part of the file — for example, flagging content as made or edited with Google AI. A partial detection (watermark in only part of a video or audio file) is a normal, meaningful result for files assembled from multiple sources.
6. Respect the limits. Per-modality upload quotas apply, meaning there are caps on how many files of each type you can check; Google has not published the exact numbers. The terms of use prohibit scripts and automated bulk checks — this is a manual, file-by-file tool, and attempting to automate it can get your access restricted.

Media type	Supported formats	Maximum file size
Images	JPEG, PNG, BMP, WebP	100 MB
Video	MP4, MOV, WebM	100 MB
Audio	MP3, WAV, OGG, AAC, M4A, FLAC	100 MB

If your file is larger than 100MB or in an unsupported format, convert or trim it before uploading — but remember the best-practices section below: every conversion is a small gamble with the watermark, so keep the highest-quality version you can. The whole check takes a few minutes: sign in, pick the media type, upload a file of up to 100MB, and read whether a SynthID watermark was found in all or part of it.

Built-In Checks You Already Have: Search, the Gemini App, and Chrome

Visiting a dedicated portal is not the only way to benefit from SynthID. Google has woven provenance checks directly into the products people already use every day: Search, the Gemini app, and Chrome now surface context about AI-generated content automatically, without you uploading anything or visiting a special site. When these products encounter watermarked media, they can tell you about it in place — a quiet, background layer of verification running while you browse normally.

The scale of these built-in checks is remarkable. Together, the verification features in Search, the Gemini app, and Chrome now regularly handle over 1 million provenance requests every single day. That figure tells you two things: first, that watermarked AI content is genuinely everywhere in the content people consume; and second, that automated, in-place checking has become a routine part of how the web works, not a specialist activity.

Think of the two systems as complements. The built-in features are the automatic background check — always on, zero effort, covering what you happen to encounter. The portal at synthid.com is the manual, deliberate check — for the specific file someone sent you, the clip you are about to publish, or the image you are about to believe. You don't always need to visit the portal: the built-in checks in Search, the Gemini app, and Chrome already run over a million provenance checks a day. Use the automatic layer for everyday browsing and the portal for the files that actually matter.

Best Practices for Getting a Reliable Result

A watermark check is only as good as the file you feed it. Follow these practices and your results will mean what you think they mean; ignore them and you risk false confidence in either direction.

• Always check the original file, never a screenshot or re-capture. Photographing a screen, screen-recording a video, or re-recording audio creates a brand-new file whose pixels and samples were never watermarked. The portal checks the file you upload, not the scene it depicts — a screenshot of a watermarked image will very likely come back clean.
• Upload the highest-quality copy you have. Every round of recompression, downscaling, or format conversion makes the watermark marginally harder to read. If you have the choice between the original download and a copy that has been through three messaging apps, always check the original.
• Check before you edit. Cropping, filters, and compression are within the watermark's design tolerances, but there is no reason to test those tolerances when you can check the untouched file first. Make verification the first step, editing the second.
• For compilations, check the parts, not just the whole. A video stitched from several clips, or an audio file assembled from multiple recordings, may carry the watermark in only part of its length. If a full-file check comes back clean but you are suspicious of one segment, check that segment on its own.
• Keep provenance notes. Write down where each file came from, when you received it, and who sent it. A detection result is far more useful when you can connect it to a source — "the watermark was found in the clip the supplier sent on Tuesday" is actionable in a way a bare result is not.
• Test the tool on your own generations first. Before you rely on the portal for a high-stakes check, upload a file you generated yourself with a participating tool and confirm it gets flagged. That single experiment teaches you more about the tool's behavior than any guide can.

Always check the original file: a screenshot, re-recording, or heavily recompressed copy gives the watermark its hardest possible test, and a clean result on such a copy tells you nothing.

Privacy: What Happens to the Files You Upload

Uploading a personal file to any online checker raises an obvious and legitimate question: what happens to it after the check? Per published reports describing the portal's on-screen notice, uploaded files are processed for the watermark check and then deleted — they are not kept, published, or added to any gallery. The only thing retained is a short-lived signature used to enforce the per-modality upload quotas, so the system can count how many checks you have run without keeping your actual files.

That is a reassuring design, but a few words of practical caution are still in order. First, "processed then deleted" is per published reports of the notice text — Google has not published a full, detailed data-handling document specifically for the portal, so treat the notice as the current statement rather than a legal guarantee you have personally verified. Second, as a general rule for any online tool, avoid uploading media you would be uncomfortable sharing with a stranger: deeply personal photos, unreleased creative work, or confidential business material. The portal is a checking tool for files you are already questioning, not a vault.

For most everyday uses — checking a viral image from a group chat, a suspicious product photo, or a voice clip of uncertain origin — the privacy posture described in the notice is proportionate to the task. Per published reports of the portal's notice, your file is processed and then deleted — only a short-lived signature is kept, to enforce the upload quotas.

Limitations and Edge Cases You Must Know About

Every honest guide to a detection tool must spend as much time on what the tool cannot do as on what it can. Here are the limitations and edge cases that matter, stated plainly.

First, the survival claims have boundaries. Per Google DeepMind, the watermark is designed to remain detectable after cropping, filters, compression, MP3 compression, speed changes, and added noise — the standard gauntlet of online sharing. But Google has not confirmed the exact thresholds where detection breaks down. A light crop is fine; whether an extreme transformation — a tiny cropped region blown up, a file re-encoded a dozen times at rock-bottom bitrates — still carries a readable mark is untested territory, and you should not assume it does.

Second, partial detection is normal and meaningful. A video assembled from five clips, only one of which was AI-generated, may report a watermark in only part of the file. An audio track with an AI-generated music bed under a real voiceover can behave the same way. "Detected in part of the file" is not a glitch or an inconclusive result — it is the tool accurately describing a mixed-origin file, and it deserves the same attention as a full detection.

Third, edits made after watermarking degrade the signal. The watermark is embedded at the moment of generation; everything done to the file afterward — however survivable in principle — chips away at it. A file that has been cropped, filtered, recompressed, screenshotted, and re-uploaded five times is a much harder read than the original, and at some point the signal drops below detectability. This is another reason the best-practices rule — check the original first — is not optional advice.

Fourth, adversarial removal is an open contest. Google designed SynthID to be difficult to strip out without visibly or audibly degrading the file — removing the mark should, in theory, ruin the media. But no watermarking scheme is provably indestructible, attackers keep improving, and Google has not confirmed resistance against every conceivable removal technique. Treat the watermark as strong evidence, not as a mathematical proof.

Fifth, re-captures defeat the check by construction. Photographing a watermarked image off a screen, or holding a microphone up to a speaker playing watermarked audio, produces a new file whose pixels and samples were never watermarked in the first place. There is nothing for the portal to find, and a clean result in this case is expected — which is exactly why screenshots must never be treated as checkable originals.

Finally, the coverage limitation from the earlier section bears repeating because it is the limitation users will hit most often: generators that never embedded SynthID — Midjourney, Meta's tools, Microsoft's tools, open-weight models — are invisible to this portal by design. The watermark is built to survive normal online handling, but it is not indestructible — extreme edits, re-captures, adversarial attacks, and non-participating generators all sit outside what this portal can promise.

Alternatives: Other Provenance and Checking Tools

The SynthID Detector is the authoritative reader for SynthID watermarks, but a complete verification toolkit has more than one instrument. The most direct complement is OpenAI's own verification page at openai.com/research/verify, which checks uploaded images and audio for provenance signals associated with OpenAI's tools — including both C2PA metadata and SynthID watermarks. It is designed to detect content generated with ChatGPT, the OpenAI API, or Codex, and it accepts common formats including PNG, JPG, WEBP, GIF, MP3, WAV, AAC, FLAC, OGG, OPUS, and PCM. If the file you are questioning might have come from OpenAI's ecosystem rather than Google's, this is the natural second check to run.

Behind that tool sits the industry's other major provenance standard: C2PA Content Credentials. Where SynthID hides a signal inside the media itself, C2PA attaches tamper-evident metadata recording where a file came from and how it has been edited. The two approaches complement each other neatly — a watermark survives metadata stripping, while credentials carry a readable edit history — which is why OpenAI's checker looks for both. As more cameras, tools, and platforms adopt Content Credentials, checking them will become as routine as checking watermarks.

A word of caution about the crowded market of third-party "AI detectors" that claim to identify AI-generated images from visual patterns alone. These tools verify no watermark and check no provenance record; they make statistical guesses from pixels, and their accuracy varies wildly — including confident wrong answers in both directions. They can be a weak supplementary signal at best, and they should never overrule a watermark-based result.

And do not forget the automatic layer: the built-in provenance checks in Search, the Gemini app, and Chrome, running over a million requests a day, mean that much of what you encounter has already been quietly screened before you ever think to check it manually.

Tool	What it verifies	How it works	Access and notes
SynthID Detector	SynthID watermarks in images, video, and audio	Upload a file of up to 100MB	Public portal at synthid.com, in English
OpenAI verify	C2PA metadata and SynthID watermarks in images and audio	Upload a file	Public page at openai.com/research/verify
Built-in checks in Search, Gemini app, and Chrome	Provenance information on content you encounter	Automatic, over 1M requests a day	Inside Google products, no upload needed
Third-party pattern-based detectors	Statistical guesses from pixels alone	Upload or paste a file	Varies by service; no watermark authority, treat with caution

No single checker sees everything: pair the SynthID Detector's watermark authority with provenance tools like C2PA-based checkers for the widest net.

Who Should Use the SynthID Detector — and When

Newsrooms and fact-checkers are the most obvious users. Before publishing a viral image or clip as genuine, a two-minute watermark check can confirm AI origin outright or — equally valuable — tell the desk that this particular test came back empty so other verification methods are needed. In a breaking-news environment where the pressure to publish first is intense, a fast, authoritative positive result is worth its weight in gold.

Teachers have a compelling use case too. Media literacy is most persuasive when it is hands-on: students who upload files themselves, watch a watermark get detected, and then watch a Midjourney image come back clean learn the tool's power and its limits in a single lesson. The portal turns an abstract lecture about AI fakes into a lab exercise.

Creators and freelancers should treat it as professional self-defense. If a client hands you "original" photos, footage, or a voiceover for a paid project, a quick check protects your reputation — discovering after delivery that your video was built on undisclosed AI-generated assets is a client relationship problem you never want. Buyers of stock media, marketing agencies vetting influencer content, and editors reviewing contributor submissions all sit in this same boat.

And then there is everyone else: the everyday user deciding whether to share a shocking image, act on a surprising video, or trust a voice note. Anyone who publishes, teaches, creates, or shares media has a use for this — and the check pays off most exactly where the cost of being wrong is highest. Make it a habit at those moments, and the few minutes it takes will save you from the far larger cost of spreading or acting on a fake.

Common Mistakes to Avoid

Most misuses of the SynthID Detector are not technical failures — they are thinking errors. Here are the ones to watch for, in rough order of how often they happen.

• Mistake 1: Reading "no watermark" as "proven real." This is the big one, worth repeating a third time because it is that common. A clean result means no SynthID watermark was found — nothing more. It is not a certificate of authenticity, and treating it as one is how people get fooled by fakes from non-participating generators.
• Mistake 2: Uploading a screenshot instead of the original. A screenshot is a new file; the watermark lives in the original file's pixels. Checking a screenshot and concluding "it must be real" is testing the wrong object entirely.
• Mistake 3: Expecting coverage that does not exist. Uploading a Midjourney image or a Meta AI video and being surprised by a clean result means the coverage section of this guide was skipped. Know your generator before you interpret your result.
• Mistake 4: Trying to automate bulk checks. The terms prohibit scripts and automated checking, per-modality quotas cap your volume, and attempting to work around either can cost you access. This is a deliberate, file-by-file tool — use it that way.
• Mistake 5: Editing first, checking later. Every edit is a small gamble with the watermark. Check the untouched file the moment you receive it; you can always edit afterward.
• Mistake 6: Relying on this tool alone. The portal is one instrument. Pair it with provenance checks, source verification, and plain critical thinking — especially for high-stakes decisions.

Almost every bad conclusion from this tool traces back to one root error: treating it as a universal lie detector instead of a watermark verifier. Keep the distinction front and center and the mistakes largely disappear.

What's Next: More Partners and More Languages

The October 2026 launch is best understood as a starting line, not a finish line. Google names Apple as a partner with support forthcoming, but has not confirmed a date — that integration, when it arrives, will meaningfully widen the portal's coverage, given how much AI-generated content flows through Apple's ecosystem. It is the single most concrete expansion to watch for.

Beyond Apple, the obvious next steps are more partners adopting the SynthID standard and more interface languages beyond the launch English. A watermarking system becomes more valuable with every generator that joins it, because each new participant closes a little more of the coverage gap described earlier. But candor requires the caveat: Google has not confirmed timelines for additional partners or languages, so treat these as reasonable expectations rather than promises.

The place to watch for all of it is the official announcement post, which Google already updated once on October 29, 2026 — a sign the company intends to keep it current as the portal evolves. The portal launched in English with a first set of partners — Google has signaled more of both, but has not confirmed dates. Check back on the official sources periodically; this is a fast-moving area and today's limitations may not be tomorrow's.

Conclusion: Check Before You Trust

The story of this guide fits in one paragraph. Since 2023, Google DeepMind's SynthID has been quietly embedding invisible watermarks into AI-generated images, video, and audio at a staggering scale — over 180 billion images and videos, 240,000 years of audio. On October 7, 2026, Google handed everyone the reader: the SynthID Detector portal at synthid.com, open worldwide in English. A positive result is strong, trustworthy evidence of AI origin from a participating system. A clean result is the absence of that evidence — informative, but never proof of authenticity.

The habit this tool enables is simple and worth building: check before you trust. Check the startling image before you share it. Check the too-good product photo before you buy. Check the client-supplied footage before you build a project on it. Check the voice note before you act on it. Each check takes minutes, and the built-in features in Search, the Gemini app, and Chrome are already screening over a million files a day in the background for you.

Calibrated expectations are the whole game. This portal will not catch every AI fake — nothing can, while major generators sit outside the watermarking standard. But within its coverage it offers something the internet has never had before: a direct, authoritative, public answer to the question "was this made with AI?" Use it for what it is, respect what it is not, and you will navigate the age of synthetic media with far more confidence than guessing ever gave you.

Sources

• synthid.com — the public SynthID Detector portal, opened worldwide in English on October 7, 2026.
• "Google expands SynthID Detector for AI content" — the official Google DeepMind announcement by Pushmeet Kohli, VP of Science and Strategic Initiatives at Google DeepMind, published October 7, 2026 and updated October 29, 2026. Source of the launch date, the scale figures (over 180 billion images and videos watermarked; 240,000 years of audio), the partner list, the I/O May 2025 early-portal history, and the 1M+ daily built-in provenance requests figure.
• openai.com/research/verify — OpenAI's public verification tool, which checks uploaded images and audio for C2PA metadata and SynthID watermarks (page verified live).

Note on sourcing: the launch facts, scale figures, partner list, and built-in feature statistics above come from Google's official announcement. Details of the portal's interface — sign-in options, result wording, upload quotas, and the file-handling notice — are per published reports, as Google has not published full official documentation of the portal interface; they are fenced as such in the text.

Frequently Asked Questions

Is there a cost to use the SynthID Detector?

Google has not announced any pricing for the portal. It opened worldwide on October 7, 2026 as a public tool that anyone can use, and the launch announcement describes expanded access rather than a paid product. If that ever changes, expect the news on Google's official blog — but as of now, there is no confirmed price, tier, or subscription involved.

Do I need an account to use the SynthID Detector?

Per published reports, yes — the portal asks you to sign in before uploading files, and it accepts a Google, OpenAI, or Apple account. Google has not published full official documentation of the sign-in flow, so follow whatever options the portal itself presents when you visit.

Does a "no watermark found" result prove a file is real?

No — and this is the most misunderstood point about the tool. A clean result means only that no SynthID watermark was detected. The file could have been made with a generator that does not use SynthID (such as Midjourney, Meta's tools, or Microsoft's tools), the watermark could have been removed or damaged by editing, or the file could genuinely be real. The portal cannot distinguish between these cases, so never treat a clean result as proof of authenticity.

Can it detect images made by Midjourney, Meta, or Microsoft tools?

No. Those systems do not embed SynthID watermarks, so their outputs are outside this portal's coverage by design and will come back with no watermark found. The portal covers Google's AI products and the partners that adopted SynthID — OpenAI, NVIDIA, and Kakao — with Apple support forthcoming on an unconfirmed date.

What is the maximum file size I can upload?

Files of up to 100MB are accepted. Supported formats are JPEG, PNG, BMP, and WebP for images; MP4, MOV, and WebM for video; and MP3, WAV, OGG, AAC, M4A, and FLAC for audio. If your file is larger or in another format, convert or trim it first — but use the highest-quality version you can, since heavy recompression makes the watermark harder to read.

Can the watermark survive cropping, filters, and compression?

That is exactly what it was designed for. Per Google DeepMind, the SynthID watermark is built to remain detectable after cropping, filters, compression, MP3 compression, speed changes, and added noise — the everyday handling files go through online. However, Google has not confirmed the exact thresholds where detection breaks down, so extreme transformations should not be assumed safe.

Is the portal available in languages other than English?

Not yet. The portal launched worldwide in English on October 7, 2026. More languages are a reasonable expectation for a global tool, but Google has not confirmed any timeline, so for now English is the only interface language.

What happens to the files I upload?

Per published reports of the portal's on-screen notice, uploaded files are processed for the watermark check and then deleted — they are not stored, published, or reused. Only a short-lived signature is retained, used to enforce the per-modality upload quotas. As a general precaution, still avoid uploading deeply personal or confidential media to any online checker.

Can I check files automatically or in bulk with a script?

No. The terms of use prohibit scripts and automated bulk checks, and per-modality upload quotas cap how many files of each type you can verify. The SynthID Detector is designed as a manual, file-by-file tool, and attempting to automate it risks having your access restricted.

When will Apple support be added?

Google has not confirmed a date. Apple is named as a partner with support forthcoming, which will widen the portal's coverage meaningfully when it arrives — but until Google announces it, do not expect content from Apple's AI features to be detectable. Watch the official announcement post for updates.