Google SynthID: What the Invisible AI Watermark Really Detects (2026)
Since 2024, Google DeepMind has been rolling out SynthID, an invisible digital watermark added to content generated by its AI models. Imperceptible to the eye and ear but detectable by a dedicated tool, it aims to answer a question that has become central: was this text, image or video made by artificial intelligence? As Google embeds SynthID by default across its models, the term is climbing in search. This guide explains what SynthID actually detects, its limits, and why detecting AI is not the same as certifying that a real photo is authentic.
What is SynthID?
SynthID is a watermarking technology built by Google DeepMind. The moment a Google model generates content, SynthID embeds an imperceptible signal directly into the data — a tiny change to an image's pixels, a sound's frequencies or a text's word choices. The result looks identical to a human, but a trained detector can later recognise the signature and report that the content carries a SynthID watermark.
Unlike metadata (easily stripped) or a visible logo (easily cropped out), the watermark is woven into the content itself. Google says it survives many common transformations: cropping, adding filters, compression, or changing a video's frame rate. According to Google DeepMind, more than 10 billion pieces of content have already been watermarked with SynthID.
What SynthID detects (and across which content)
Originally limited to images, SynthID now covers the four main content types generated by Google's models:
- Images — via the Imagen generator.
- Video — via Veo, Google's video generation model.
- Audio — via Lyria (generated music and sound).
- Text — via Gemini, by subtly steering word choice in a detectable way.
To check content, Google launched the SynthID Detector, a portal that analyses an uploaded file and flags the likely presence of a SynthID watermark, indicating which parts of the content are affected. This consumer-facing tool is what brought "SynthID" into everyday language.
The limits: what SynthID does not detect
Only Google's ecosystem (and a few partners)
This is the decisive limitation. SynthID only spots content produced by models that embed the watermark: Google's own (Gemini, Imagen, Veo, Lyria) and a small number of partners. An image from Midjourney, Stable Diffusion or Adobe Firefly carries no SynthID watermark — the detector will return "no detection" even though the image is genuinely AI-generated. SynthID therefore sees only a fraction of the synthetic content in circulation.
The watermark can erode
Robustness has limits. Heavy transformations — intensive editing, strong re-compression, screenshots, and for text, translation or a full rewrite — can weaken or erase the watermark. Text watermarking is the most fragile: a few rewordings are often enough to collapse the detector's confidence. The output is expressed as a probability, not a binary certainty.
Absence of a watermark proves nothing
The corollary of the two points above: a "no SynthID watermark" result does not mean "authentic content." It could be a real photo… or an image from a non-participating model, or content whose watermark was destroyed. SynthID can confirm an AI origin; it cannot rule it out.
Detecting AI ≠ certifying a real photo
This is the most important nuance, and the most misunderstood. SynthID answers "was this content made by an AI?". It does not answer the opposite question — often the only one that matters to an insurer, a court or a client: "is this real photo authentic, taken where and when claimed?".
These are opposite problems. The first tries to unmask the synthetic. The second tries to prove the real — a capture's provenance, its timestamp, its integrity. A SynthID watermark on a generated image says nothing about the reliability of a damage photo taken by a policyholder; and a real photo, by definition, carries no SynthID watermark.
SynthID, C2PA, capture-time certification: complementary logics
Three approaches coexist, and they do not replace one another:
- SynthID (watermark) — marks AI-generated content at the source. Useful for flagging the synthetic, limited to the ecosystem that emits it.
- C2PA / Content Credentials — an open standard (Adobe, Microsoft, the BBC…) that attaches a signed "where this file came from" history. Useful when the device or software actually writes it.
- Capture-time certification — seals the authenticity of a real photo or video at the moment it is taken, with a verifiable cryptographic fingerprint.
That third path is the one Truth-Check takes: the photo is captured straight from the device sensor inside the app, then the image, its metadata, its location and a SHA-256 hash are sealed. Any later modification breaks the hash, and anyone can verify the certificate publicly. Where SynthID tries to spot the fake after the fact, certification proves the real at capture — a positive, binary proof that does not depend on knowing which generator did (or did not) mark the content.
Should you rely on SynthID to verify evidence?
SynthID is a real and welcome advance against AI-generated disinformation. But to establish that a photo or video is reliable evidence, it has three blind spots: it is proprietary (controlled by Google), it does not cover content outside its ecosystem, and above all it does not certify the authenticity of a real capture — so it does not replace a certificate. A watermark can, at best, reveal that content is artificial. Proving that content is authentic remains a separate task. To go further, see our guides on how to detect an AI-generated photo and the 7 tools to verify a photo's authenticity.
FAQ
Is SynthID 100% reliable?
No. It detects content marked by participating models with good confidence, but the result is probabilistic and the watermark can be weakened by heavy editing, strong re-compression, or — for text — a rewrite or translation.
Can SynthID detect a Midjourney or DALL·E image?
SynthID only detects content that carries its watermark, i.e. content from Google's models and its partners. An image from a model that does not implement SynthID will not be recognised as such.
Does a SynthID watermark prove a photo is authentic?
No, it's the opposite: a SynthID watermark indicates that content was generated by an AI. Proving that a real photo is authentic is a matter of provenance and capture-time certification, not AI detection.
Does SynthID replace an authenticity certificate?
No. SynthID marks synthetic content at the source; a certificate proves the integrity and origin of a real capture. The two approaches are complementary and answer different needs.
Sources
Try Truth-Check for free
Certify your photos and videos in seconds. 3 free credits, no commitment.
Download
