The Google AI detection tool is now available to everyone.

Google has announced that SynthID is now available for any user interested in trying out this authentication system for content generated by artificial intelligence. This system allows for the embedding of imperceptible watermarks in images, videos, and texts created by machines, helping users verify whether content was produced by humans or algorithms.

The company emphasizes that SynthID has been made an open-source text watermarking tool, freely accessible to developers and businesses, with the aim of facilitating the identification of AI-generated content. SynthID made its debut in 2023 as a way to mark images, audio, and video generated by artificial intelligence. It was originally integrated into the Imagen platform, and its use was expanded in May 2024 when it was announced that it would be included in the Gemini chatbot during the I/O event.

This system works by encoding tokens, which are the basic units of data used by generative AI to interpret a request and predict the next word in the response. According to a DeepMind blog, the process involves inputting additional information into the token distribution at the time of generation, thereby modulating the probability of certain tokens being generated. By comparing the model's word choices with the "adjusted probability scores" in relation to the expected score patterns for marked and unmarked texts, SynthID can determine if a phrase was written by artificial intelligence.

A recent study published in Nature claims that this process does not affect the accuracy, quality, or speed of responses, and that the system cannot be easily circumvented. Unlike standard metadata, which can be easily deleted, it is asserted that the SynthID watermark remains even when the content has been cropped, edited, or modified.

Soheil Feizi, an associate professor at the University of Maryland, notes that achieving a reliable and imperceptible watermark in AI-generated text is a challenge, especially in situations where language model outputs are almost deterministic, such as in factual questions or code generation tasks. The open-source nature of the system allows the community to test the detectors and assess their effectiveness in various contexts, which is crucial for understanding the limitations of these techniques.

However, SynthID is not infallible. Although it is resistant to manipulation, its watermarks can be removed if the text is input into a translation application or if it is significantly rewritten. Additionally, its effectiveness decreases in short text passages and when trying to determine if a response based on a factual statement was generated by AI. For example, both humans and machines will answer "Paris" to the question about the capital of France, which only has one correct answer.

The SynthID tool is available for download on Hugging Face as part of Google’s updated Responsible GenAI Toolkit.