You can now hand a short clip to Google’s Gemini app, type something as plain as “Was this generated using Google AI?”, and get a specific, machine-backed answer — not a guess, not a shrug. It’s a single-tap provenance check tucked into a consumer chat app, but it lands squarely in the middle of a much larger fight over deepfakes, provenance, and who gets to decide what counts as “real” on the internet.

Under the hood, the feature does one job and does it quietly: it looks for SynthID, Google DeepMind’s invisible watermark that the company embeds into media produced or edited with its own models. That watermark is designed to survive ordinary recompression, cropping, and most common edits while remaining imperceptible to human viewers — it’s meant to be read by software, not seen by people. That’s why Gemini can tell you not only whether a clip shows traces of Google AI, but also which parts contain the fingerprint.

Using it is intentionally simple. Upload a short video to Gemini, ask “Was this generated using Google AI?” in plain language, and the assistant runs a scan across both the visual track and the audio track. If SynthID shows up, Gemini will report things like “SynthID detected within the audio between 10–20 seconds. No SynthID detected in the visuals.” That per-segment detail is not trivia: many modern clips are stitched together from multiple sources — real footage, AI-generated B-roll, synthetic voiceover — and knowing which layer carries an AI stamp is essential for a useful provenance check.

There are practical limits and a few important caveats. The tool is built for snackable social clips: it accepts files up to around 100MB and about 90 seconds in length, which fits the world of TikToks, Reels, and Shorts far better than feature films. And crucially, this is not a universal deepfake detector — it can only detect Google’s own SynthID watermark. If a video was generated or heavily manipulated by a third-party model that doesn’t embed SynthID (or that uses a different watermarking scheme), Gemini won’t flag it as AI-generated, even if every frame is synthetic.

The feature is rolling out inside the Gemini app wherever that app is available, and it supports the same languages Gemini already understands; people only need a Google account and to accept Gemini’s terms to use it. Google has been progressively folding image and video provenance tools into consumer products rather than keeping them behind research dashboards, reflecting a push to make provenance more discoverable to everyday users.

That design — model-level watermarking plus a one-tap verifier — is a pragmatic, platform-driven answer to a gnarly problem. On the one hand, embedding a machine-readable signal at generation time and giving users a simple way to check for it can speed up newsroom triage and help platforms and creators label content more responsibly. Google says it’s already watermarked and tracked billions of pieces of AI-generated media since SynthID’s launch, and folding detection into Gemini is another step toward making that provenance visible at the consumer edge.

On the other hand, the protections are only as broad as the ecosystem that adopts them. SynthID helps when Google models are involved; it does nothing to expose content created by other companies’ tools unless those vendors also agree to embed compatible provenance. That fragmentation matters: a world where each major model family writes its own invisible signature is better than nothing, but it’s still far from the cross-platform, standardized labeling advocates have been arguing for. Industry efforts such as content credentials and C2PA-style metadata have been floated as ways to bridge different toolchains, but the reality today is patchy, with a mix of invisible watermarks, visible stamps, and platform-specific signals.

For journalists, moderators, and curious consumers, Gemini’s check is a fast and practical triage tool: it answers one narrow question quickly — was Google AI in the loop? — and that answer can meaningfully narrow an investigation. For example, if a viral clip contains SynthID only in its voice track, that suggests an edited real video whose audio was swapped; if SynthID appears throughout the frames, it points to wholly synthetic footage. But remember: a negative result (no SynthID) is not a green light for authenticity; it simply means Gemini didn’t find Google’s watermark. Detecting maliciously created media in the wild still requires complementary techniques: cross-referencing timestamps, checking camera metadata when available, reverse-image and reverse-video searches, and human reporting.

At scale, the real question is whether provenance becomes a default expectation on the internet or stays an optional tool you must seek out. Google’s move makes it easier for regular people to check content without juggling specialized tools, which is a meaningful nudge. But long-term trust will depend on wider interoperability, consistent platform policies, and — perhaps most difficult of all — incentives for creators and platforms to adopt provenance practices even when it’s not in their short-term interest. Until then, Gemini’s video check is a useful, narrowly scoped instrument in a much larger orchestra of verification work.

If you want to try it yourself, the instructions and limits are spelled out in Google’s help pages and the DeepMind synthID documentation, which also links to technical papers and developer tools for people who want to understand how the watermarking actually survives edits and compression. For now, the takeaway is simple: when a dubious clip lands in your feed, asking Gemini “Was this generated using Google AI?” is an easy and revealing first question — provided you read the answer in the right context.