There is no doubt that the rise of large-scale language models (LLMs) has multiplied the amount of content written with AIs, thanks to their ability to generate fluid and coherent texts.
This has raised concerns about the authenticity of information and the risk of misinformation, as it is difficult to distinguish between human text and AI-generated text.
To address this problem, the Google DeepMind team has developed SynthID, a Gemini innovation ecosystem tool (watermark) that allows you to accurately identify whether content was created by an artificial intelligence model.
Additionally, they have recently released the technology behind SynthID for text, so that it can be used publicly and for free.
In this article, we explain how SynthID for text works, what its limitations are, and why identifying AI-generated content is crucial to ensure the safe and responsible use of these technologies.
What is SynthID-Text and how does it work?
SynthID-Text is a watermarking system that allows you to identify if a text has been generated by an AI model, without compromising the quality of the content or significantly increasing processing costs.
Unlike other methods, this approach is integrated during the text generation process, allowing imperceptible markup to be added without the need to modify the underlying language model.
SynthID-Text uses a process called Tournament Sampling, which subtly alters the selection of words that make up the generated text.
Instead of simply choosing the token with the highest probability, the algorithm adjusts the probability of the selected wordss, introducing a statistical pattern that works like a watermark.
This pattern is undetectable to human readers and does not affect the coherence or quality of the text.
The most relevant thing is that, during the detection phase, this brand can be identified without the need to perform computationally complex operations or have access to the original model, which is an advantage in production environments.
An example of how SynthID-Text works
Imagine they use a language model, like Google Gemini, that has SynthID built in. Let’s also imagine that you want to generate an article about environmental sustainability.
When generating the text, the language model assigns probabilities to each word based on the context and the text generated so far.
For example, if the sentence under construction is “Sustainability is important because,” the model could assign high probabilities to words like “reduces,” “promotes,” or “increases.” This is where the tournament sampling process comes into play.
Instead of selecting only the most likely word, the system introduces subtle modifications to the word selection probabilities to inject a watermark.
For example, the word “reduces” might have a probability of 60%, while “promotes” is set to 25% and “increases” is set to 15%.
With watermarking, SynthID model could adjust the probabilities so that “reduces” becomes 57%, “promotes” becomes 28%, and “increases” becomes 15%. This slight modification introduces a pattern that will be useful for later detection.
When someone reviews this text and wants to check if it has been generated by an AI model using SynthID-Text, a detection algorithm is applied that analyzes the statistical pattern of the text.
This algorithm can measure the “signature” of the text and determine if it meets the characteristics of the implemented watermarking.
If the detection is successful, the system will confirm that the text was generated by a model that uses SynthID-Text technology, thus allowing its origin to be identified without requiring access to the original model.
The importance of marking texts generated with AI
Marking texts generated by AI responds to the challenges that arise with the massive use of these technologies.
First of all we have trust in the information. The growing presence of AI-generated text poses the risk of content spreading without clear attribution. Identifying its origin is essential to maintain transparency.
Without an effective identification system, language models can be used to spread disinformation or generate malicious content, such as spam or phishing.
In certain contexts, such as education and scientific research, it is important to know whether a text was produced by a person or an AI model to avoid fraud and ensure originality.
Finally, there are potential regulatory challenges. In the near future, governments may require companies to identify when they use AI-generated content.
Implementing watermarking now will allow organizations to anticipate these demands and encourage responsible use.
Limitations and challenges of watermarking with SynthID
Despite its advantages, SynthID-Text is not a foolproof solution. There are challenges inherent to using any watermarking method.
Although the watermark is subtle, a malicious user could modify the text to try to erase the watermark without affecting consistency. Google claims that the mark is resistant to light editing and text cropping.
On the other hand, SynthID-Text focuses on identifying texts generated by specific Google models, which implies that its widespread use will require cooperation between different AI developers to establish a common standard.
Finally, although the system has proven to be safe and efficient, its effectiveness could decrease when applied to models other than those for which it was designed.
The text is just the beginning
Google’s release of SynthID marks an important precedent in the industry, with the hope that it will be possible to integrate watermarking techniques without compromising the user experience.
Ultimately, the ability to identify and attribute AI-generated text will be critical to maintaining transparency and trust in today’s digital ecosystem. And soon we will have similar tools for audio, images and video.
In a world where AIs have an increasing presence in our lives, tools like this are essential to ensure the safe and ethical use of technology.
This post is also available in: