Google announced on Tuesday (October 7) the general availability of the SynthID Detector, an independent web portal at synthid.com that allows anyone to verify whether an image, video, or audio file was generated by artificial intelligence. The tool marks the end of a cycle that began in 2023, when Google DeepMind first introduced SynthID as an invisible digital watermarking technology for AI-generated media.
The SynthID Detector works in a straightforward manner: users upload a media file — whether an image in JPG, PNG, WEBP, HEIC, or GIF format, a video in MP4 or WEBM, or audio in WAV, MP3, or AAC — and the system scans the content for imperceptible watermarks. If a watermark is detected, the tool highlights the regions of the file most likely created by AI. For audio, the detector indicates specific segments where the watermark is present. For images, it marks individual suspicious areas.
The figures Google presented alongside the expansion announcement are substantial. According to data disclosed in the official announcement, the company has now applied SynthID watermarks to more than 180 billion images and videos, as well as 240,000 years of audio content. Google's Gemini AI assistant integrates SynthID natively — meaning virtually all visual or audio content created by Google's Gemini models already carries the digital signature. By May 2026, Google had reported 100 billion watermarked images and videos, showing that the internal adoption pace accelerated significantly over just a few months.
But SynthID is not only a Google technology. One of the most significant aspects of today's update is the opening of the watermarking ecosystem to partner brands. Google confirms that companies including OpenAI, NVIDIA, and Kakao are already using the SynthID technology to watermark content generated by their own models. In the text space, OpenAI has developed its own detection system, but also adopts SynthID for visual and audio media. NVIDIA, meanwhile, is integrating SynthID into the Cosmos video generation models from World Labs. Apple is expected to join the ecosystem soon, according to Google.
The availability of the detector via an independent website resolves a practical issue that persisted since the beta version: previously, the only way to verify SynthID watermarks outside of Google's trusted tester program was through the Gemini model itself. That approach was fragmented — users had to interact with the model to receive an analysis, which could sometimes be confusing when seeking a simple binary answer. Now, synthid.com offers direct verification, without intermediaries, with clear results.
Beyond the web portal, the SynthID Detector is already integrated into the Gemini app, Google Search, and Google Chrome. Google reports that users make one million verification requests per day through these integrations. The AI Content Detection API is also being launched as part of Google Cloud's Gemini Enterprise Agent Platform, enabling businesses to integrate verification directly into their own workflows — whether for internal operations such as feed curation and fraud prevention, or for end-user features like synthetic media identification on social networks.
SynthID operates differently from traditional authorship metadata systems. Instead of relying on information embedded in a file that can be stripped or lost during processing, the watermark is encoded directly into the image pixels or the audio waveform. This means the watermark remains detectable even after common edits like cropping, applying filters, lossy compression, or changing video frame rates. Google's internal testing showed the technology withstands more than 300 cycles of compression and manipulation.
The SynthID Text technology, meanwhile, operates on a different layer. As Google described in its technical documentation, it is a logits processor that modifies the probability scores assigned to each token during text generation. In a sentence like "My favorite tropical fruits are mango and...", the word "bananas" would receive a higher score than "airplanes". SynthID subtly adjusts these scores to encode the watermark information without perceptibly affecting text quality. The watermarked text remains indistinguishable from human-written text in most cases, but the detector can identify the watermark's presence based on probabilistic patterns. The source code for SynthID Text is already available as open source for any developer who wants to integrate the technology into their own models.
However, the SynthID Detector — and watermarking technology in general — has important limitations worth noting. Detection is probabilistic and can produce false positives or false negatives. Watermark application is less effective on highly factual text responses, where there is less creative freedom to adjust scores. And as Ars Technica observed in prior coverage, SynthID is not a universal detector: it only identifies content carrying a SynthID watermark, meaning AI-generated content without the technology goes undetected. Competing systems, such as those developed by OpenAI for text detection, suffer from the same limitation.
The broader question that remains is whether watermarking can be effective against deliberate disinformation. Malicious actors can simply use AI models that do not adopt SynthID, or attempt to strip or obfuscate existing watermarks. Google itself acknowledges that SynthID does not directly prevent motivated adversaries from causing harm — it merely makes it harder to pass AI-generated content as authentic. Real effectiveness depends on adoption at scale: if only a few major players adopt the standard, unmarked AI-generated content will continue to proliferate.
What the SynthID Detector demonstrates, however, is a paradigm shift in how major technology companies are approaching content authenticity in the age of generative AI. Rather than trying to detect synthetic content after creation — a technological arms race that has historically favored AI creators — the SynthID strategy starts from the premise of marking content at the point of generation, making authenticity verifiable from the source. If this approach consolidates with growing partner adoption, it could represent one of the few viable mechanisms for restoring trust in online digital content authenticity.
The SynthID Detector is now available for public use at synthid.com, and Google is already expanding verification for text and C2PA Content Credentials over the coming weeks.
Sources: Google Blog, TechCrunch, The Decoder
✓ Independent sources cross-checked and verified before publishing
