Why the Google SynthID Detector matters now
The significance is not that the internet has gained another website that labels media “real” or “fake.” SynthID Detector does neither. Its importance is narrower and, in some ways, more consequential: a major AI developer is letting the public interrogate a provenance signal embedded in media at the moment of generation or editing. What had largely been a platform-side capability is becoming evidence that a reporter, teacher, moderator, parent or ordinary social-media user can request directly.
That shift arrives as synthetic images, cloned voices and generated video move through election campaigns, fraud attempts and harassment at the speed of reposting. A viral clip can reach millions before a forensic review begins. An invisible AI watermark can travel with the file through many common edits, allowing the provenance check to move closer to the first moment of doubt: before publication, before a classroom accusation, or before a victim sends money to a voice on the phone.
But the asymmetry matters. A detected mark is affirmative evidence that a participating system touched the file. No detected mark is only an absence of that particular evidence. The Google SynthID Detector therefore changes the first step of verification, not the final standard. The right question is not “Can this site prove the file is real?” It is “Can this site find a known provenance signal, and what should I investigate next?”
From DeepMind research to a public checker
Google DeepMind introduced SynthID in 2023 as an invisible watermarking system designed to be embedded in generated media without making the output visibly worse. Unlike a corner label that can be cropped away, the signal is distributed through the file. Google says it is designed to remain detectable after ordinary transformations such as cropping, filtering and compression.
The system spread as Google's generation products expanded. Nano Banana and Gemini image tools, Veo video generation, Lyria music, Flow, ProducerAI and Vids all became part of a growing watermark footprint. At Google I/O 2025, the company announced a detector for journalists, media professionals and researchers, but access ran through an early-access waitlist. The October 7, 2026 opening at synthid.com removes that professional gate for the basic check and makes the service available worldwide in English.
Google had already built related checks into Gemini, Chrome and Search. According to the company, those integrated experiences now handle roughly one million verification requests each day. The public site matters because the user no longer needs to encounter a file inside a particular Google product. The evidence can be brought to the checker.
Google AI content detection now reaches beyond Google media
The detector is not limited to files made by Google. Google says media from OpenAI, NVIDIA and Kakao can now be detected through the service, with Apple support expected “soon.” That partner list turns SynthID from a single-company label into the beginnings of an ecosystem, though it remains far short of an industry-wide standard. Microsoft and Meta operate separate provenance and labeling approaches, while many smaller or open-source generators may add no durable watermark at all.
This is the central policy question behind the launch: whether provenance tools remain competing islands or become interoperable. The Coalition for Content Provenance and Authenticity, known as C2PA, takes a broader approach by attaching signed records about a file's origin and edits. SynthID focuses on a robust signal embedded in the media itself. The two approaches can complement each other — credentials explain a chain of custody, while a watermark can survive when metadata is stripped — but users should not have to consult a dozen incompatible checkers to understand one file.
Who benefits — and who can still slip through
Journalists are the clearest beneficiaries. A newsroom receiving a suspicious video can check for a SynthID match before devoting hours to frame analysis, geolocation and source interviews. A positive result may be enough to stop an unverified clip from reaching air. Teachers and schools gain a concrete signal when evaluating suspicious media, although the tool does not resolve authorship or intent. Platforms can use it to support labels, and everyday users get a faster way to question a viral post before forwarding it.
The gaps favor anyone who understands the limits. A bad actor can choose a generator that does not watermark, move among providers with incompatible standards, or attempt dedicated watermark-removal attacks. Apple's absence at launch leaves a major consumer ecosystem outside the current partner set. And research across AI detection has repeatedly shown that tools can fail even on material created by the models they are supposed to recognize. SynthID's design may survive routine edits, but robustness against common editing is not immunity from an adversary deliberately trying to erase or confuse the signal.
There is also an access tradeoff. The public website still requires a Google, OpenAI or Apple sign-in and operates with an approximate daily quota, according to Ars Technica's testing as cited in launch coverage. That is a much wider door than the old waitlist, but it is not anonymous, unlimited public infrastructure. For high-volume newsrooms, election monitors and trust-and-safety teams, those limits will determine whether the site is a convenient spot check or a dependable part of daily workflow.
AI generated image detector or deepfake detector? The distinction matters
Calling SynthID a deepfake detector overstates its purpose. A forensic deepfake detector examines visual, acoustic or statistical artifacts and estimates whether content was manipulated. SynthID Detector instead looks for a known invisible AI watermark. It can answer whether that participating provenance signal is present. It cannot examine every possible route by which media might have been generated, edited or deceptively presented.
That difference explains the one-way certainty. “Detected” is meaningful: a compatible system says the media is generated or edited. “Not detected” may mean the file is authentic, but it may also mean it came from an unsupported provider, never received a watermark, lost the signal through an attack, or uses a different provenance standard. A negative result is not an AI image authenticity check certificate.
What you can upload to synthid.com
The public checker accepts a wide range of common media formats. It does not check text, and the public version has also simplified the output. An earlier iteration could highlight regions where a watermark appeared; the public site reports a plain detected-or-not-detected result.
| Media | Supported formats | What the result says |
|---|---|---|
| Images | JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, TIF, GIF | Whether a compatible SynthID watermark is detected |
| Video | MP4, MOV, WEBM | Whether a compatible SynthID watermark is detected |
| Audio | WAV, MP3, OGG, FLAC, AAC, M4A | Whether a compatible SynthID watermark is detected |
| Text | Not supported on the public site | No text verdict |
How to tell if an image is AI generated
Start with provenance, not appearance. Upload the original or highest-quality version available, because screenshots and repeated downloads can discard useful context even when a watermark survives. If SynthID is detected, preserve the file and the result, identify where it first appeared, and ask whether the post describes the media accurately. Generated does not automatically mean deceptive: illustrations, disclosed edits and creative work can all be legitimate.
If the result is negative, continue. Use reverse-image search, inspect timestamps and metadata, look for an earlier upload, compare landmarks and weather, listen for edits, and contact the claimed source. The practical answer to how to tell if an image is AI generated is rarely one detector. It is a chain of corroboration.
How to check if video is AI generated
Video deserves extra caution because a genuine clip can be paired with a false caption, spliced into a different sequence or stripped of its original audio. A watermark check can reveal compatible generated or edited media, but it does not validate the story told around the clip. Check key frames, source accounts, location details and continuity. For audio, seek an independent recording or direct confirmation from the speaker's organization before treating a voice as authentic.
The numbers show scale — and the size of the blind spot
Google says more than 180 billion images and videos have been watermarked and that the amount of protected audio is equivalent to 240,000 years of continuous playback. Those figures make SynthID one of the largest deployed provenance systems in the world. If a person listened without stopping, the audio total would reach back to a time long before modern humans had spread across the planet.
Scale, however, is not coverage. The figures measure media touched by participating products, not the universe of generated content. The internet contains outputs from competing closed models, open-source systems, older tools and modified pipelines. Even 180 billion marked items can coexist with an enormous unmarked population. The million daily checks inside Google's products show demand for verification; they do not establish the accuracy rate of a universal detector, because this is not one.
The launch is part of a broader expansion of Google's AI infrastructure. The company is simultaneously pushing more capable models and agent tools, including the workplace automation covered in our analysis of Google's Gemini agent for work and the expected model competition around Gemini 4 “Argon”. Provenance is becoming a necessary counterweight to the volume those systems can produce. Longer-term projects such as Project Suncatcher's proposed AI computing expansion underline the same tension: generation capacity is rising faster than most verification systems can coordinate.
What happens next
The first near-term test is Apple support. “Soon” is not a date, and the value of that integration will depend on which Apple-generated or Apple-edited media receives a compatible signal. The larger test is regulatory. The European Union's AI Act is increasing pressure for machine-readable labeling of synthetic content, but mandates matter only if systems can communicate across companies and survive normal distribution. Interoperability could turn a collection of corporate tools into public infrastructure; fragmentation could force users to guess which checker matches which generator.
The second test is adversarial. Every public detector creates a feedback loop for attackers: test an edited file, change it, test again. Google says SynthID is built to withstand cropping, filters and compression, but dedicated removal research will continue. A trustworthy system will need transparent evaluation, rapid updates and a way to distinguish “unsupported” from “checked and absent.” A binary answer is easy to understand, but it can hide very different kinds of uncertainty.
For readers, the safest workflow is simple. If the Google SynthID Detector says yes, treat the file as generated or edited by a compatible system and investigate the context. If it says no, do not call the file real. Preserve the original, check other provenance systems, trace the source and corroborate the event. The public launch gives everyone a useful new instrument. It does not remove the need for judgment.