Google is open-sourcing its AI text watermarking toolkit, SynthID to developers and businesses, the company announced on X. The company had integrated the tool to its Gemini AI model for watermarking AI-generated content in a way that it said was “imperceptible to humans” but could be detected by an algorithm easily. 

Announced by the company in August last year, SynthID can add the watermark to text, images, audio and videos.

Large Language Models generate the next word in a sentence simply by predicting what the most likely word might be next. SynthID assigns a probability score for each word to determine if it was predicted by the AI model, and then judges the entire text after considering all the scores.

Google has previously said that the SynthID toolkit is accurate without disturbing the quality, accuracy, creativity and speed of the AI-generated text. 

This tool can help in identifying Gen AI usage in texts that are longer than two sentences. And it is capable of spotting AI usage even if the text is cropped, paraphrased, or modified. However, it does struggle to function if the text has been translated. 

In a blog post released in May, Google admitted that SynthID may not be the perfect solution to detect AI generated content but could be “an important building block for developing more reliable AI identification tools.”