Knowing how to tell if an image is AI generated has become a basic digital skill. Fake photos of politicians, videos of disasters that never happened and cloned voices spread every day on messaging apps and social networks. On 7 October 2026 Google opened SynthID Detector to everyone: a free tool that checks whether an image, video or audio file was made with Google AI or with AI from its partners. This guide shows how to use it step by step, which other methods are worth knowing, and where each of them falls short.
What SynthID Detector is and who can use it
SynthID is Google DeepMind's watermarking technology, introduced in 2023. It embeds a signal into generated media that people cannot see or hear but software can read. Google says the watermark is designed to survive common edits such as cropping, filters, frame-rate changes and lossy compression, and for audio, added noise, MP3 compression or speed changes.
The detector was shown at Google I/O 2025 but was limited to selected journalists, researchers and media professionals. Since 7 October it is available to anyone, globally, in English only for now, according to Google's announcement. Some figures Google shared:
- More than 180 billion images and videos and 240,000 years of audio have been watermarked with SynthID.
- The built-in verification features in Search, the Gemini app and Chrome handle more than one million requests a day.
- Besides Google content (models such as Nano Banana, Veo and Lyria, and tools such as Gemini, Flow and Vids), it detects content from partners that adopted SynthID: OpenAI, NVIDIA and Kakao, with Apple coming soon.
How to check if an image, video or audio is AI with SynthID Detector
- Go to synthid.com, the official detector site, on desktop or mobile. It is free.
- Upload the file. According to TechCrunch, it accepts images in JPG, PNG, WEBP, AVIF, HEIC, HEIF, TIFF, BMP and GIF; video in MP4, MOV and WEBM; and audio in WAV, MP3, OGG, FLAC, AAC and M4A.
- Read the result: the tool tells you whether it finds a SynthID watermark.
- Interpret it carefully: a positive result is strong evidence that the file came from Google or partner AI. A negative result does not prove the content is real.
If you already use Gemini there is a shortcut: upload the image to the Gemini app and ask "Was this created with Google AI?". Gemini looks for the SynthID watermark and adds context from its own reasoning. Google is building the same checks into Search and Chrome.
The big caveat: it is not a universal AI detector
This is the most important point in this guide. SynthID Detector only recognises its own watermark. An image made with Midjourney, Flux, Stable Diffusion or an open model running locally carries no SynthID signal, so the detector will not flag it even though it is AI. Nor will it catch content created before a partner adopted the technology, or watermarks from Meta or Microsoft. Google has not published accuracy or false-positive rates either, and TechCrunch notes that other watermarking systems have often failed to identify content from their own makers' models.
Other ways to tell if an image is AI generated
No single technique is foolproof, so combine several.
1. Content Credentials (C2PA)
C2PA is an open standard backed by Adobe, Microsoft, Google, OpenAI and others. It attaches provenance metadata to a file describing which tool created or edited it. You can read it by uploading the file to contentcredentials.org/verify. The weakness: it is metadata, so a screenshot or the recompression many social networks apply strips it out.
2. OpenAI's verification page
OpenAI runs its own Verify OpenAI-generated content page. It checks images and audio for both C2PA metadata and the SynthID watermark in content made with ChatGPT, the OpenAI API or Codex. Use it when you suspect an image came from ChatGPT.
3. Visual clues
- Hands, teeth, ears and jewellery with impossible or asymmetric shapes.
- Garbled text on signs, number plates or labels (though 2026 models are much better at this).
- Inconsistent lighting, shadows and reflections.
- Overly smooth skin, backgrounds that melt together or repeated patterns.
- In video: objects that appear or vanish, odd blinking, strange physics. In audio: no breathing or a flat intonation.
4. Context and reverse image search
Often the most effective test is basic journalism. Run the image through Google Lens or a reverse image search to find where it first appeared, check whether a reliable outlet has published it, and be wary of brand-new accounts pushing it.
5. Third-party AI detectors
Some services analyse statistical patterns without relying on watermarks. They can be a hint, but false positives and negatives are common, so never treat their percentage as a verdict.
Comparison: which method to use
| Method | What it detects | Main limitation |
|---|---|---|
| SynthID Detector | SynthID watermark (Google, OpenAI, NVIDIA, Kakao) | Misses other generators |
| Content Credentials (C2PA) | Provenance metadata | Lost in screenshots and recompression |
| OpenAI verifier | C2PA and SynthID in OpenAI content | Only OpenAI images and audio |
| Visual clues | Typical generation errors | Increasingly subtle |
| Reverse search | Origin and context | Useless for brand-new images |
Why spotting AI content matters
Cloned-voice scams and fake celebrity videos promoting investments are now an industry, as we covered in our piece on how AI is scaling crypto fraud. Meanwhile tools like Wan 2.6, which turns photos into video, put convincing synthetic media within anyone's reach. Opening SynthID Detector to the public does not solve the problem, but it gives everyone a free, reliable check for a growing share of AI content. For the bigger picture, see the real impact of AI on everyday life.
The rule of thumb: a positive SynthID or Content Credentials result is strong evidence; the absence of any signal proves nothing. When in doubt, look for context and trusted sources.
Frequently asked questions
Is SynthID Detector free?
Yes. Google opened it to everyone at no cost on 7 October 2026, although the interface is English-only for now.
Does SynthID detect Midjourney or Stable Diffusion images?
No. It only detects content carrying the SynthID watermark, meaning content generated by Google or partners such as OpenAI, NVIDIA and Kakao. For other generators, rely on C2PA, visual clues or reverse search.
Can the watermark be removed from an AI image?
Google says SynthID survives cropping, filters and compression, but it does not publish accuracy rates, so survival under every manipulation cannot be guaranteed. C2PA metadata, by contrast, is easy to strip.
How can I tell if an audio clip is AI?
Upload it to synthid.com, or to OpenAI's verifier if you suspect OpenAI tools. If no signal appears, listen for breathing, pauses and intonation, and confirm the message through another channel with the person who supposedly said it.
