Gemini アプリに AI 画像検証機能を導入、SynthID で来歴確認 How we’re bringing AI image verification to the Gemini app

Google DeepMind has announced a new feature in the Gemini app that lets users verify whether an image was generated or edited by Google's AI. As synthetic media proliferates and visual misinformation becomes harder to spot, the company is positioning the tool as a consumer-facing entry point into content provenance.

At the heart of the feature is SynthID, DeepMind's watermarking technology. SynthID embeds an imperceptible signal directly into the pixels of an image, designed to survive common transformations such as cropping, compression, and color adjustments. When a user uploads an image to Gemini and asks about its origin, the model checks for the SynthID signal and indicates whether the image was created or modified by Google's generative tools, including Imagen and the recently popular Nano Banana image model. DeepMind has signaled that the verification capability will expand to audio and video over time.

The rollout arrives amid mounting concern over deepfakes and AI-manipulated media. Recent election cycles around the world have surfaced high-profile cases of synthetic imagery, prompting regulators and platforms to push for clearer provenance signals. The industry is converging on multiple, complementary approaches: the C2PA standard, backed by Adobe, Microsoft, and major camera makers, attaches cryptographically signed metadata describing how a file was produced, while watermarking schemes like SynthID embed the signal in the content itself. Metadata can be stripped when files are re-uploaded or screenshotted, so pixel-level watermarks are seen as a more durable, if imperfect, layer of defense.

A notable limitation is scope. Gemini's verification primarily detects content produced by Google's own models, since SynthID needs to have been applied at generation time. Images from OpenAI's DALL·E, Midjourney, Black Forest Labs' Flux, or open Stable Diffusion variants use different provenance approaches—or none at all—so a truly cross-vendor verification experience would require broader interoperability. DeepMind has open-sourced parts of SynthID's text watermarking work, and similar moves on image and video components could help, though no single watermark is considered fully tamper-proof against a determined adversary.

For everyday users, the practical takeaway is modest but meaningful: a quick way to ask Gemini whether a suspicious image came from Google's AI stack, without needing specialized forensic tools. For the broader ecosystem, the launch underscores that provenance is shifting from a back-end research topic to a user-visible product feature, and competitors are likely to follow with their own consumer-grade detection surfaces.