Google SynthID Detector: What Businesses Must Know
Google opened its SynthID Detector to everyone on 7 October 2026. Here is how the free AI watermark checker works, what it can and cannot tell you, and the critical gap businesses must understand.
The Short Answer
On 7 October 2026, Google handed the public a free, browser-based tool to check whether an image, video, or audio file was generated or edited by AI. The Google SynthID Detector, live at synthid.com, lets anyone upload a media file and learn whether it carries the invisible SynthID watermark. Until this week, the portal sat…
On 7 October 2026, Google handed the public a free, browser-based tool to check whether an image, video, or audio file was generated or edited by AI. The Google SynthID Detector, live at synthid.com, lets anyone upload a media file and learn whether it carries the invisible SynthID watermark. Until this week, the portal sat behind a waitlist reserved for journalists, media professionals, and researchers since its unveiling at Google I/O 2025. Now it is open worldwide, and for the first time it recognizes watermarks from Google and partners including OpenAI, NVIDIA, and Kakao, with Apple support expected soon.
The timing is no accident. The Pindrop 2026 Deepfake Readiness Index, published in September 2026, found that 74% of US security leaders had faced a suspected deepfake attack in the previous twelve months, and one in four of those hit reported losses above one million dollars from a single incident. A free verification tool will not end that threat, but it changes what every business should do about it. This guide explains what the detector is, how to use it, what it can and cannot tell you, and the critical gap most coverage misses entirely.
What the Google SynthID Detector Is
SynthID is Google DeepMind's invisible watermarking technology, introduced in 2023. When a participating generative AI tool creates or edits media, SynthID embeds a statistical signal directly into the content: into the pixels of an image, across the frames of a video, or into the waveform of audio. The signal is imperceptible to people and distinct from any visible badge or metadata label. It is designed to survive common edits such as compression and cropping.
The Google SynthID Detector is the public portal that reads those signals. Upload a file and it answers one narrow question: does this media carry a supported SynthID watermark? Google says it has watermarked more than 180 billion images and videos plus the equivalent of 240,000 years of audio, and that SynthID checks built into Google Search, the Gemini app, and Chrome now run more than one million verifications per day.
The October 2026 release is also the first cross-company watermark check in one place. OpenAI now embeds SynthID in images generated through ChatGPT, Codex, and its API; NVIDIA and Kakao have joined as partners; Apple is expected to follow, which would cover content from Apple Intelligence features such as Image Playground. Previously, a ChatGPT image carried a SynthID mark that Google's detector ignored. Now it gets flagged.
How to Use the Google SynthID Detector Step by Step
The portal is deliberately simple. Here is the full flow:
- Go to synthid.com. The service is available globally but currently operates in English only.
- Sign in with a Google, OpenAI, or Apple account. Sign-in is mandatory. Google requires it to enforce usage limits and to block automated probing of the detection logic. The terms of service ban automated queries, scraping, and attempts to study or defeat the watermark.
- Upload the image, video, or audio file. Supported formats include common image types (JPG, PNG, WebP, HEIC, GIF), video (MP4, MOV, WebM), and audio (WAV, MP3, OGG, FLAC, AAC, M4A). Google advises uploading the highest-quality copy you have, because heavy compression and repeated re-edits reduce detection accuracy.
- Wait for the result. Processing happens in real time. Google's privacy notice says uploaded files are deleted once results are returned, while a digital signature linked to your account is kept for 24 hours to enforce the daily quota.
- Read the result as a clue, not a verdict. The portal reports whether a supported watermark was detected. A positive result is strong evidence that a participating AI system generated or edited the file. A negative result means only that no SynthID mark was found. It does not certify that the file is authentic, and the site carries this disclaimer itself.
Each account gets roughly 10 checks per day. Google does not print the exact figure on the page; engineers told Ars Technica the cap exists to make it harder for anyone to use repeated checks to develop watermark-stripping techniques. For routine business spot-checks that quota is generous; for anything like batch moderation, it signals the portal is a manual tool, not an API.
What the Google SynthID Detector Detects (and What It Cannot)
The tool's name invites overconfidence, so the boundaries matter more than the feature list. Here is the precise split.
What it detects: media carrying the SynthID watermark. That covers content generated or edited by Google's AI tools, including Gemini, Imagen, Lyria, Veo, and Nano Banana, plus the products around them such as the Gemini app, Flow, ProducerAI, and Vids, and watermarked media from partners OpenAI, NVIDIA, and Kakao. One notable exclusion: plain text. The public portal accepts images, video, and audio only.
What it cannot detect: everything else. Meta and Microsoft use their own separate watermarking systems, so their synthetic media will not register. The AI video tools dominating viral feeds, Runway, Kling, Hailuo, Pika, and Seedance, do not use SynthID at all. Open-weight models that anyone can download and run on local hardware typically add no watermark of any kind. And even a genuine SynthID mark can be weakened or destroyed: researchers note that aggressive compression, heavy filtering, or regenerating the file can degrade the signal, and Google's own result page warns that a file may have been edited further after the watermark was applied.
The summary, in one line: the Google SynthID Detector answers whether a file carries one specific family's watermark. It is not a universal AI detector, and Google says so explicitly in its own FAQ.
What the Google SynthID Detector Means for Businesses
For most companies, the detector's value is defensive. In February 2026, Italy's largest bank, Intesa Sanpaolo, lost roughly 95 million euros in a deepfake voice scam: criminals used AI-cloned voices, fake WhatsApp messages, and forged emails impersonating senior executives to authorize international transfers. In 2024, a finance employee at the design firm Arup wired about 25 million dollars to fraudsters after a video meeting staffed entirely by deepfake colleagues. Voice cloning now needs only seconds of public audio, such as a podcast clip or a LinkedIn video.
This is where the detector earns a place in standard procedure. When a suspicious invoice, voice note, or video arrives with an urgent request, running it through synthid.com takes seconds and can immediately confirm whether a participating AI system produced it. A positive hit escalates the investigation: pause the transfer and verify through a known channel, such as calling the executive on a saved number. A clean result tells you nothing conclusive, so the verification step happens regardless. Employees in finance and HR face the majority of targeted voice attacks, and urgency plus authority are the scammers' real weapons. A strict "never act on an unsolicited request without a second channel" policy, paired with a free technical check, is a low-cost upgrade to any security posture.
For higher-risk companies, the detector is only a first taste of provenance verification. Serious protection means layered defenses: deepfake detection tuned to audio and video biometrics, verification workflows for payments, and regular security reviews. That is the work a dedicated cybersecurity partner handles: threat modeling that treats the watermark check as one signal among several. As the Pindrop index noted, only 10% of surveyed enterprises have purpose-built tools against deepfakes while 90% lack them. The gap between threat and readiness is itself the opportunity.
Businesses should also think about provenance from the other direction. The industry is moving toward a world where watermarked content is traceable and unwatermarked content is treated with suspicion. Building your own AI products with provenance in mind, watermarks, audit trails, and disclosure labels from day one, is becoming part of trustworthy engineering. Agencies that develop custom AI and machine learning solutions increasingly build these safeguards into the products they ship, so verification is a feature rather than an afterthought. Our guide on data privacy and ownership covers the wider case for controlling your content pipeline.
The Gap No One Mentions: "Not Detected" Does Not Mean Human-Made
Here is the insight this article exists to deliver, and the sentence every business leader should internalize: a "watermark not detected" result from the Google SynthID Detector does not mean a file is human-made.
Consider the asymmetry. The portal can only recognize watermarks deliberately embedded by a participating generator. Every other path to a convincing fake bypasses it silently. Content from Meta's or Microsoft's systems returns "not detected." Content from open-weight models run on local hardware, with no watermark at all, returns "not detected." Content from the viral AI video apps dominating social feeds returns "not detected." Content from which the watermark was deliberately stripped returns "not detected."
This creates a dangerous failure mode: false confidence. An employee checks a suspicious voice note, sees "watermark not detected," and concludes it must be genuine. In reality, a scammer who uses an unwatermarked generator gets a clean result by design. The check becomes worse than useless; it becomes a credential for the fraud. The same trap applies to customer-facing verification. A company that stamps content "verified human" based on a negative SynthID result would be certifying something the tool never claimed.
The honest framing, and the one Google itself uses, is that SynthID is a positive-signal system. A positive result is informative: a participating AI system touched this file. A negative result is the absence of information, and absence of evidence is not evidence of absence. Businesses building moderation, fraud review, or provenance pipelines must therefore layer signals: watermark checks plus deepfake detectors trained on audio-visual artifacts, plus metadata analysis, plus source-of-origin verification, plus human review for high-stakes decisions. Text is another blind spot worth naming: AI-written phishing copy is the delivery vehicle for most AI voice and face fraud, and this portal does not address text at all.
The strategic takeaway: treat synthid.com as you would a single smoke detector in a large facility. It is genuinely useful, it is free, and installing it is a no-brainer. But you would not cancel the sprinkler system, the fire drills, and the insurance because one detector is on the wall. Layered provenance is the real system, and the businesses that build it now will not become the next headline.
For more on how AI is reshaping the security landscape at the device level, see our breakdown of AI agent security and Apple's new Mac controls.
Frequently Asked Questions
Is the Google SynthID Detector free?
Yes. Anyone can use it at synthid.com after signing in with a Google, OpenAI, or Apple account. Each account gets roughly 10 media checks per day, a limit Google says exists to prevent abuse of the detection system.
Can the Google SynthID Detector detect any AI-generated image?
No. It only detects the SynthID watermark from Google and partners OpenAI, NVIDIA, and Kakao, with Apple expected soon. Content from Meta, Microsoft, unwatermarked open-weight models, or tools like Runway and Kling will not be flagged. Google states plainly that it is not a general AI detector.
Does "watermark not detected" mean a file is real?
No. It means only that no supported SynthID watermark was found. Content from generators that never embed SynthID, or from which the watermark was removed or degraded, returns the same negative result. A negative result must never be treated as proof of authenticity.
Does the detector check text written by AI?
No. The public portal accepts images, video, and audio only. It does not analyze plain text.
What happens to files I upload to synthid.com?
Google's privacy notice says uploaded media is processed in real time and deleted after results are returned. A digital signature linked to your account is kept for 24 hours to enforce the daily usage quota.
Which AI tools' watermarks does the detector recognize?
Watermarks from Google's generative tools (Gemini, Imagen, Lyria, Veo, Nano Banana, Flow, ProducerAI, Vids), from partner OpenAI (images from ChatGPT, Codex, and the OpenAI API), and from NVIDIA and Kakao. Apple support is expected, which would cover Apple Intelligence Image Playground content.
The public launch of the Google SynthID Detector is a genuine milestone: cross-company watermarking checked in one free portal, backed by a corpus of more than 180 billion watermarked images and videos. Used correctly, as one positive signal inside a layered verification process, it gives every business a free first line of defense against synthetic media fraud. Used incorrectly, as a stamp of authenticity, it hands scammers a cleaner cover story. The difference is understanding, and now you have it. If your business needs help building layered media verification, deepfake-resistant workflows, or trustworthy AI products with provenance built in, contact our team and we will map out a plan that fits your risk profile.
Related Articles
Keep reading — more guides you may find useful.
Related Services
Services connected to this guide.
