How to Detect AI Watermarks in Photos

AI watermarking has become a major topic of discussion after Anthropic announced that its Claude models will soon mark text output with invisible watermarks, while Google revealed that its visible watermark on Gemini images is now optional. These two examples highlight the varied methods companies use to identify AI-generated content. The technology generally involves embedding data that is undetectable to the human eye or ear but can be read by detecting tools, offering a way to verify content without relying on unreliable AI detectors.

Watermarking methods vary depending on the media type. For images, algorithms slightly modify pixel values to embed a digital signature. Google’s SynthID, for instance, spreads this invisible signature across an image so that even a cropped version retains detectable portions. This is different from the visible gray symbol in the corner of Gemini images; the visible mark is optional, but the invisible watermark remains active regardless of its setting.

Audio watermarking often places signature sounds outside the range of human hearing, typically below 20Hz or above 20,000Hz. Video usually combines these techniques. Text watermarking operates differently; every word an LLM generates comes with a probability score. The system inflates the chances that specific sets of random words will appear, making it possible to detect LLM-generated text without changing its meaning. This works better with longer pieces of text, which provide more opportunities to spot the less-likely word combinations.

C2PA is a framework for adding metadata that can help verify the origin of media. Some camera manufacturers use it to provide a traceable record of where an image came from, including whether it was generated with AI tools. SynthID originated with Google’s DeepMind but is now open-sourced and used by other companies, including OpenAI. However, some companies do not use watermarks at all, and implementations can be inconsistent between tools. The presence of a watermark can confirm something was made with AI, but the absence of one cannot prove it wasn’t.

Despite the prevalence of SynthID and C2PA, detecting these markers is still difficult. OpenAI offers a standalone tool to check for SynthID or C2PA in a file, but it only detects media generated by OpenAI. During testing, an image generated by Gemini that contained a detectable SynthID watermark was not flagged by OpenAI’s tool. Google created a SynthID Detector portal, but it is currently invite-only, specifically for journalists and verification professionals. Access to some detection functions is available through Gemini or Google, though prompts must be specific to invoke the correct check.

It is still possible that a piece of media contains a watermark that requires a different detector. Unless you have a strong indicator of which tool was used to create the content, it can be difficult to thoroughly check for every kind of watermark. Text-based watermarks, like those used by Claude or Gemini, are particularly hard to detect because neither offers a public way to check them yet.

Vulnerabilities and Limitations

While any watermark can technically be removed, SynthID is designed to be resistant to most common modifications. An AI-generated image with a SynthID watermark that has been cropped, filtered, or modified can still retain enough of the original watermark to be detectable. Removing metadata like C2PA is considerably easier; taking a screenshot creates a new image file based on the pixels visible on the screen. This wipes the C2PA data, though the watermark affecting those pixels may remain.

For text, the watermark is the easiest to get around. Since the method relies on altering the probability that certain words will appear, running the text through another AI tool or manually rewriting it can potentially remove the watermark. It is also important to note that a watermark does not automatically mean a piece of media is inauthentic. If someone uploads an authentic photo to a tool like Gemini for simple edits, the output will have a SynthID watermark. This does not make the whole image fake, but it will still be flagged by detectors.

Watermarks and metadata are simply tools to help determine where media likely came from and how it was modified. They can indicate AI usage, but they cannot prove the substance of an image is real or fake. The absence of a watermark cannot prove authenticity, and the presence of one cannot prove inauthenticity. Ultimately, verification requires human judgment and additional context rather than relying solely on these digital markers.