谷歌开源了其名为SynthID的文本水印技术，该技术旨在识别由大型语言模型生成的文本。SynthID通过在文本生成过程中微调每个单词的概率分数，从而在文本中嵌入不可见的数字水印，使软件能够识别出AI生成的文本，但不会影响人类对文本的理解。该技术已集成到谷歌的Gemini聊天机器人中，并可用于文本长度至少为三句话的文本，即使文本被裁剪、改写或翻译，也能识别。然而，SynthID并非完美无缺，在处理短文本、改写或翻译后的内容以及事实性问题的回答时，仍存在挑战。

👨‍💻 谷歌宣布开源其SynthID文本水印技术，该技术旨在识别由大型语言模型生成的文本。SynthID通过在文本生成过程中微调每个单词的概率分数，从而在文本中嵌入不可见的数字水印，使软件能够识别出AI生成的文本，但不会影响人类对文本的理解。

📊 SynthID的工作原理是，在生成文本时，根据上下文和概率分数选择最有可能的词语，然后调整每个词语的概率分数，从而在不影响文本质量和可读性的情况下，在文本中嵌入数字水印。这种调整后的概率分数模式构成一个独特的数字签名，可以用来识别由特定模型生成的文本。

🔍 尽管SynthID在识别由大型语言模型生成的文本方面取得了进展，但它并非万能的。它在处理短文本、改写或翻译后的内容以及事实性问题的回答时，仍存在挑战。因此，该技术还需要进一步改进，才能成为识别AI生成内容的可靠工具。

🚀 谷歌表示，SynthID已经集成到其Gemini聊天机器人中，并可用于文本长度至少为三句话的文本，即使文本被裁剪、改写或翻译，也能识别。这表明谷歌正在积极探索将SynthID应用于实际场景，并希望将其推广到更广泛的领域。

💡 尽管SynthID在识别AI生成内容方面存在局限性，但它仍然是一个重要的工具，可以帮助人们更好地理解和识别AI生成的内容，从而做出更明智的决策。随着AI技术的发展，识别AI生成内容的技术也将不断改进，以更好地应对日益复杂的挑战。

Google’s SynthID text watermarking technology, a tool the company created to make AI-generated text easier to identify, is now available open-source through the Google Responsible Generative AI Toolkit, the company announced on X.

“Now, other [generative] AI developers will be able to use this technology to help them detect whether text outputs have come from their own [large language models], making it easier for more developers to build AI responsibly,” Pushmeet Kohli, the vice president of research at Google DeepMind, told MIT Technology Review.

Watermarks have become increasingly important tools as large language models are used to spread political misinformation, generate nonconsensual sexual content, and for other malicious purposes. California’s already looking into making AI watermarking mandatory, while China’s government started requiring it last year. Yet the tools are still a work in progress.

SynthID, which was announced last August, helps make AI-generated output detectable by adding an invisible watermark into images, audio, video, and text as they’re generated. Google says the text version of SynthID works by making the text output slightly less probable in a way that is detectable by software but not humans:

An LLM generates text one token at a time. These tokens can represent a single character, word or part of a phrase. To create a sequence of coherent text, the model predicts the next most likely token to generate. These predictions are based on the preceding words and the probability scores assigned to each potential token.

For example, with the phrase “My favorite tropical fruits are __.” The LLM might start completing the sentence with the tokens “mango,” “lychee,” “papaya,” or “durian,” and each token is given a probability score. When there’s a range of different tokens to choose from, SynthID can adjust the probability score of each predicted token, in cases where it won’t compromise the quality, accuracy and creativity of the output.

This process is repeated throughout the generated text, so a single sentence might contain ten or more adjusted probability scores, and a page could contain hundreds. The final pattern of scores for both the model’s word choices combined with the adjusted probability scores are considered the watermark.

Google claims the system, which it’s already integrated into its Gemini chatbot, doesn’t compromise the quality, accuracy, creativity, or speed of generated text, which has long been an issue with watermarking systems. Google says it can work on text as short as three sentences, as well as text that’s been cropped, paraphrased, or modified. But it struggles with short text, content that’s been rewritten or translated, and even responses to factual questions.

“SynthID isn’t a silver bullet for identifying AI generated content,” Google wrote in a blog post in May. “[But it] is an important building block for developing more reliable AI identification tools and can help millions of people make informed decisions about how they interact with AI-generated content.”