Recent developments in the artificial intelligence (AI) field have made it almost impossible to discern between AI-generated content (writeups, images, music, etc.) and human-made content. This is where Google Deepmind’s SynthID comes into the picture.
Deepmind, an AI research laboratory, is a subsidiary of Google, that works on deep learning and reinforcement learning. Following the expansion of AI tools in the last year, it developed SynthID, a tool to watermark AI-generated content. Let’s learn more about this tool.
Demis Hassabis Net Worth – Co-founder and CEO of Google Deepmind
What is Google Deepmind SynthID?
SynthID is a toolkit for marking and identifying AI-created content. This new technology places hidden digital marks, also known as watermarks, into different kinds of media made by artificial intelligence – including text pieces, sound clips or tracks, pictures, and videos.
These watermarks are impossible to notice by humans but can be found for identification with SynthID.
How does it Work?
SynthID works in primarily two ways:
1. Watermarking AI-Generated Content
SynthID uses complex deep-learning models and algorithms to embed digital watermarks into AI-generated content. This method does not harm the quality of the original content.
The toolkit is capable of watermarking different types of media, guaranteeing that it is invisible to human eyes but still detectable by identification tools.
2. Identifying AI-Generated Content
SynthID checks for digital watermarks within the content. This scanning aids users in knowing whether the content or some parts of it are made by Google’s AI tools. The toolkit has the flexibility to work on various media formats like text, audio, images, and video.
What is the Best Generative AI: ChatGPT vs Copilot vs Gemini vs Pi vs Claude 2
SynthID for AI-Generated Text
SynthID can provide watermarking and identification for AI-generated text. This is very important considering large language models (LLMs) like Gemini and ChatGPT can generate human-like texts.
An LLM produces text token by token, where tokens can be single letters, full words, or even parts of phrases.
The model makes predictions about what the next probable token would be given previous words and their probability scores in order to create coherent text.
Deepmind also shares an example. Let’s assume when the phrase is “My favorite tropical fruits are __,” the model could guess tokens such as “mango,” “lychee,” and “papaya.”
SynthID makes small modifications to the probability scores of predicted tokens, keeping the quality, accuracy, and inventiveness of the output intact. This technique subtly embeds a digital watermark into the text that is being produced. The process of adding watermarks repeats many times within each generated sentence; in other words, one single sentence contains numerous adjusted probability scores. As the text length increases, the accuracy of SynthID’s watermarking also improves.
How to Add Gemini Extension to Google Chrome?
SynthID for AI-Generated Music
SynthID started watermarking AI-made music and audio late in 2023. Its initial use was with Lyria, a high-level AI model for making music from DeepMind. Every audio that Lyria generates has a SynthID watermark in its waveform.
The method that SynthID applies is transforming audio waves into spectrograms, which are visualizations of sound frequencies changing in time as two-dimensional images. After creating the spectrogram, the digital watermark gets added to it.
Following this, the spectrogram is transformed back into a waveform. This step guarantees that the watermark continues to be soundless for human listeners. The watermark is robust to common audio transformations, such as adding noise, MP3 compression, and speed changes.
SynthID can find a watermark in various locations within the audio track.
How to Use Gemini AI in Google Docs, Sheets, Slides, Gmail, and Drive?
SynthID for AI-Generated Images and Video
Additionally, SynthID adds digital watermarks to the AI-generated pictures and videos. In the case of images, the watermark is inserted within the pixels themselves, which guarantees that it does not alter its sharpness or visual quality.
As for videos, each frame includes a unique identification mark, making sure video quality remains intact and unharmed by any alterations.
The watermark stays visible even when doing alterations like cutting, filtering, color changing, brightness adjusting, frame rate modifying, and applying lossy compression methods such as JPEG. This guarantees that people can still recognize AI-created content even if it has undergone a number of modifications.
Availability
At present, SynthID can be used by Vertex AI customers with text-to-image models such as Imagen 3 and Imagen 2. These models are capable of producing images that are high in quality and have different artistic styles. Furthermore, SynthID watermarks image outputs on ImageFX too.
Also, SynthID has been included in Veo, the video generation model of DeepMind. Users can now check whether an image or video is made using AI by using the “About this image” feature in Google Search or Chrome. It enhances clarity and trust in the digital media.
The Bottom Line
It is critical to recognize AI-produced content in the present digital era. This is leading to an increase in misinformation and distrust. To combat it, SynthID provides a reliable way for marking and recognizing content generated by AI.
AI Virtual Rat by Harvard and DeepMind Sheds Light on Brain-Body Connection
