# How AI Watermarks Work and How to Detect Them Methods and tools for marking and detecting AI-generated content. - **Source:** [Lifehacker](https://brifing.si/source/f1400ced-a26b-4e58-a567-6187eca3708d) - **Published:** 2026-09-02 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/c7e06264-ee64-48cd-b2be-1bd016c9feeb ## The Basics of AI Watermarking AI watermarks embed digital signatures into generated content that are undetectable to the human eye but identifiable with specialized tools. These signatures verify that the content was produced by AI. For example, Google's SynthID adds a distributed signature to images, detectable even when cropped. Similarly, audio files can be marked with signals at frequencies beyond human hearing (below 20Hz or above 20,000Hz). ## Watermarking in Images Image watermarks are created by making slight modifications to pixel values. Google's SynthID adds a distributed digital signature to every generated image, detectable even if cropped. Additionally, Gemini images feature a visible gray symbol in the corner, but the invisible watermark remains even if this symbol is disabled. This method provides an effective tool for verifying image origins. ## Watermarks in Audio and Video Audio watermarks embed signals at inaudible frequencies (below 20Hz or above 20,000Hz). SynthID's audio component is detectable by watermark detectors but imperceptible to humans. Videos typically combine both image and audio watermarks. However, fake content creators can bypass this by adding AI-generated audio to real video, making detection more challenging. ## Text Watermarks and Detection Text watermarks manipulate the probability scores of words generated by large language models (LLMs). For instance, the likelihood of certain words (e.g., 'fluffy' after 'cat') is inflated to signal AI generation. This works better for longer texts, where improbable word combinations are more detectable. Currently, there are no public tools to detect text watermarks from models like Claude or Gemini. ## Verification with Metadata: C2PA C2PA is a metadata framework for verifying media origins. Some camera manufacturers use C2PA to let photographers trace an image's source, including whether it was AI-generated. However, C2PA metadata can be easily removed by taking a screenshot, making it less reliable than watermarks. ## Watermark Detection Tools and Limitations OpenAI and Google offer tools to detect SynthID and C2PA, but with significant limitations. OpenAI's tool only detects its own media, while Google's SynthID Detector is currently invite-only for journalists. Gemini provides some SynthID detection functions but requires specific prompts. Asking 'Is this real?' in Gemini triggers a SynthID check, but the same query on Google may yield a visual analysis instead. ## Removing Watermarks and Challenges SynthID watermarks resist typical modifications like cropping or filtering, remaining detectable. However, C2PA metadata can be stripped by screenshotting. Removing text watermarks is more complex and may require altering semantic meaning. The absence of a watermark doesn't prove content is human-made, as some companies don't use them or implement them inconsistently. ## The Future and Importance of AI Watermarks AI watermarks are becoming critical for verifying content origins and combating misinformation. However, inconsistent implementations and limited detection tools hinder their effectiveness. Wider adoption of standards like SynthID and C2PA, along with more accessible detection tools, could unlock their full potential, especially in journalism, academia, and law. ## Sources covering this story - https://lifehacker.com/tech/how-does-ai-watermarking-work-and-how-can-i-check-for-one?utm_medium=RSS — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._