Google's new AI detector - SynthID from Google DeepMind

In a fast-paced AI environment, SOTA (state of the art) models improve on an almost weekly basis. Our team at Rephrasy knows how important it is to stay up-to-date to recent changes and improve our AI Humanizer whenever there is new information or any release on common AI Detectors.

In this post we introduce you to the new SynthID AI Detector and watermarking tool from Google itself.

SynthID AI Detector from Google

The newest release from Google DeepMind is a statement for people who use AI generated content and especially for people who work in the SEO space. Someone can assume that this piece of software is going to be a part of the Google Core Algorithm.

The Core Algorithm of Google Search is basically the mechanism to rank Websites in the SERPs (Search Engine Result Pages).

So based on this new tool, we can assume that it gets even more important that purely AI generated content, nevertheless if it's an image or a text needs to be humanized.

The mechanism of SynthID - How does it work?

Diving a bit more into the technical side of the SynthID AI Detector of Google, we can assume that the base technology is something which can be used to detect AI generated content not only from Google's LLM (Large Language Model) but also from OpenAI's ChatGPT and so on.

So now that this model is released, it will probably take a few weeks until modern AI Detectors adopt to these changes and it'll be even harder to bypass AI reliably.

One thing is sure: Google is able to detect AI generated text!

SynthID's watermarking technique is imperceptible to humans, but detectable for identification. This means that it's even harder for someone to predict whether a piece of text will be flagged as AI-generated.

Another important claim is, that the tool is currently launched in beta, which means the development is ongoing and it'll be even more advanced soon.

The tool is based on some advanced deep-learning algoriothms, meaning that even for someone with a fairly good understanding of Data Science, it is not trivial to fully understand.

In a nutshell

Our team provides a high level explaination to the mechanism of SynthID, so that everyone is able to understand.

So here is how it works:

When a large language model (LLM) generates text, it predicts one token (or text segment) at a time based on previous tokens and probabilities tied to each possible token. A token may be a part of a word, a full word, or a phrase, and the LLM will select the one that is most likely to be introduced in a natural, lay manner.

SynthID is relevant to this by manipulating the probability scores of the previous tokens. For instance, in the context mentioned, if the model is concluding a phrase “My favorite tropical fruits are __,” the list of predictions would look like “mango,” “lychee,” “papaya,” or “durian,” each of these with its probability score.

SynthID, in its turn, causes a slight shift of these scores, which leads to an unnoticeable watermark that can not be detected by the readers but rather by certain tools that have been specialized.

As a result, the AI-hash text could be marked individually while operating the natural, smooth, and coherent text.

The following screenshot from the official SynthID website shows an example of the likelihood of tokens being generated.

The whole process is repeated throughout the text, so that sentences can contain multiple adjusted probability scores, and pages or long documents can contain hundreds or more.

This means, that the longer the text, the more robost the results of the SynthID AI Detector gets.

Here is an example of a watermarked E-Mail:

How to bypass SynthID by Google?

After careful considerations, we are still working on understanding the AI Detector of Google to its best. Right now, our humanizer is capable of bypassing it, but as it is still in beta, we can assume that it will be more advanced in the coming weeks.

This is why our team is eager to dive deeper into the technology and make sure, that our AI generated content stays undetectable.

We are also looking into solutions which will ensure to be undetectable for images, videos and audio.