Real photo or AI? How to use SynthID Detector, Google’s tool available to everyone

A photo that looks too perfect circulating on WhatsApp, a shocking video of a politician, a voice message from a loved one asking for money: when faced with this sort of content, the question ‘Is this real?’ arises every day. Since Wednesday 7 October 2026, Google has been providing part of the answer. Its fact-checking tool, SynthID Detector – previously reserved for journalists, researchers and media professionals – is now available to the general public worldwide.

The service is free and accessible via a web browser at synthid.com. It analyses images, videos and audio files. The only notable limitation for French-speaking users is that the interface is currently only available in English, although Google has promised to add other languages at a later date.

An invisible watermark, embedded right from the start

To use the tool effectively, you need to understand what it is designed to do. SynthID is not a detector that ‘guesses’ whether an image looks artificial. It is a watermarking system developed by Google DeepMind: when a compatible AI generates an image, video or audio clip, it inserts a digital signature that is imperceptible to the eye and ear. This mark is designed to withstand common forms of manipulation, such as compression, slight cropping or the addition of a filter. SynthID Detector then simply searches for this signature.

This is not a new development. Google has been watermarking its own creations since 2023, and the scale is considerable: according to the company’s figures reported by Unite.ai, more than 180 billion images and videos and the equivalent of 240,000 years of audio now bear this watermark. The verification feature is already integrated into Google Search, the Gemini app and Chrome, where it processes over a million requests a day.

The main change lies in the partners. SynthID no longer recognises content generated solely by Google’s models (Gemini, Imagen, Veo, Lyria): OpenAI, NVIDIA and South Korea’s Kakao have adopted the same watermark for their own generators. Apple is set to follow suit, and images created using its AI tools will bear the mark ‘soon’, according to Google. The company hopes to make SynthID a de facto standard for the entire industry.

This trend is no coincidence. Regulators are increasingly requiring that AI-generated content be flagged: in the European Union, the Artificial Intelligence Regulation has, since August 2026, required providers of generative systems to label their output in a machine-readable format. Rather than each developing their own signature, several major players have chosen to share Google’s, which makes verification much simpler for the public.

A four-step guide

• Visit synthid.com on a computer or mobile phone.
• Log in with a Google, Apple or OpenAI account: the analysis is not available without logging in.
• Upload the file you want to check: an image, a video or an audio recording. If you’ve received an image via messaging, save it to your device first.
• Read the result: ‘watermark detected’ or ‘watermark not detected’. For a video or audio file, the tool highlights the sections where the watermark appears.

The number of scans per user per day is limited, and the terms of use prohibit the creation of multiple accounts to circumvent this limit. As regards privacy, Google states that uploaded files are processed in real time and then deleted as soon as the result is returned; only a digital fingerprint of the file is retained for twenty-four hours, to ensure compliance with the daily limit. It is still advisable not to submit any private or sensitive documents.

There’s also a shortcut for those who already use Google’s services. In the Gemini app, you can import an image and simply ask whether it was created using Google’s AI: the assistant checks for the same watermark and responds in French. Google Search and the Chrome browser offer similar verification tools. These shortcuts are handy for everyday use, but the website synthid.com remains the most comprehensive tool, particularly for videos and audio files.

What the tool can’t tell you

Google states this itself on its website: it is ‘not a general-purpose AI detector’. The limitation is structural. An image produced by a generator that does not use SynthID – whether it is an open-source model installed on a computer, Midjourney or one of the many free online tools – contains no watermark. The tool will therefore respond with ‘no watermark found’, even though the content is entirely artificial.

A negative result therefore never proves that a photo is genuine. It simply means that the signature of Google’s partners was not found. Extensive editing can also weaken the watermark to the point where it becomes undetectable, and the company acknowledges that false positives do occur on rare occasions. Finally, SynthID does not distinguish between content created entirely by AI and a genuine photograph that has simply been edited using an AI tool, for example to remove an object or change the sky.

Conversely, a positive result is very telling: if the watermark is detected, the file has been run through one of our partners’ AI systems. This is where the tool is most useful, for example in exposing a purported news photograph created using a consumer-grade image generator.

Supplement with further checks

SynthID is just one tool amongst many. OpenAI offers its own verification system for files generated by its services, and Adobe provides Adobe Inspect, which reads ‘Content Credentials’: this provenance information, based on the C2PA standard, can trace a file’s origin, the software used and certain modifications. However, this metadata often disappears when an image is shared on social media or messaging platforms.

Good habits therefore remain essential. Reverse image searches, using Google Lens or TinEye, allow you to trace the first appearance of a photo and check whether it has already been debunked. It is also important to trace the source: a major event is rarely shown by just one anonymous account. Details can still sometimes give away artificial images, such as illegible text on a placard, distorted hands or inconsistent shadows, even though recent generators are making fewer and fewer of these mistakes. For a suspicious voice message, nothing beats calling the person back on their usual number.

The issue goes beyond mere curiosity. In Africa, as elsewhere, fake images first circulate on messaging apps, where they are shared without context amongst family and community groups. Extreme photo manipulation during election periods, fake photos of disasters, cloned voices used to demand money: the malicious uses of generative AI are on the rise. A free tool, available to everyone, helps journalists, fact-checkers and ordinary users. It does not replace critical thinking, but it provides an additional tool for it.