Google launches SynthID Text to mark up AI-generated text

The Mountain View giant announced the launch of a free tool that promises to make it easier to identify texts generated by AI. This is the SynthID Texta solution that allows you to insert invisible “watermarks” into content produced by video systems. artificial intelligence.

This news comes at a crucial time, where the distinction between human and artificial content becomes increasingly challenging. With predictions indicating that by 2026, 90% of online content could be synthetically generatedand by 2030 this number could exceed 99%, the importance of tools like SynthID Text becomes evident.

Also read
Google releases advanced version of Gemini AI for free in Brazil
Gemini Live, Google’s voice assistant, now understands and responds in Portuguese

How does SynthID Text work?

SynthID Text operates by leveraging the fundamental process of AI text generation. This process is based on the creation of tokens — basic units that can be characters or complete words. The tool of Google intervenes in this process, subtly modulating the probabilities of generating these tokens.

In doing so, SynthID creates a unique, imperceptible pattern in the text that acts as a digital signature. This “watermark” does not alter the visible content of the text, but can be detected by specialized systems.

To identify the presence of this mark, the software compares the pattern of notes in the analyzed text with patterns expected from texts with and without a watermark. This allows you to determine whether the content was generated by AI or has another origin.

Limitations and challenges

Despite its potential, Google recognizes that SynthID Text faces some limitations. The tool performs poorly on short texts, content translated or rewritten from other languages, and on answers to simple factual questions.

These limitations arise because, in certain cases, there is less flexibility to adjust the distribution of tokens without compromising the accuracy or meaning of the text. For example, when answering “What is the capital of France?”, there is little room for variation without affecting the correctness of the answer.

Making SynthID Text free to developers and businesses was a wise move in the fight against misinformation and misuse of AI-generated content. The tool could be instrumental in identifying and possibly downgrading synthetically generated online content.

Additionally, SynthID Text can help solve the problem of inaccurate AI detectors, which often incorrectly label human articles and essays as machine-generated. With a reliable watermark, it will be easier to distinguish between authentic and artificial content.

Global context

The launch of SynthID Text comes at a time of growing global concern about the impact of AI on content production. China recently implemented regulations requiring the tagging of AI-generated content, and the US state of California is considering similar measures.

Other technology companies are also exploring similar solutions. OpenAI, the creator of ChatGPT, has been researching watermarking methods for years, although it has not yet released a public tool.

Therefore, Google’s SynthID Text represents a beautiful step forward in the responsible management of AI-generated content. By offering this tool for free and open source, Google demonstrates its commitment to transparency and encourages widespread adoption of ethical practices in the use of generative AI.

Source: Google AI for Developers