In a bid to enhance transparency and accountability in the use of images generated using artificial intelligence (AI), Google DeepMind and Google Cloud have unveiled SynthID. 

This tool embeds a digital watermark directly into the pixels of an image produced by Google’s AI model, Imagen. The tool is now available to a select number of users of Google’s Vertex AI platform.

The watermarked pixels are imperceptible to the human eye, but are detectable by specialised algorithms. According to Google, the tool works even after adding filters, altering colours, and compressing images. Deepmind acknowledges that extreme levels of image manipulation are likely to tarnish the watermark. 

At the moment, SynthID is exclusive to Google’s Imagen model. DeepMind has, however, expressed interest to make it available to third parties in the future.

“Being able to identify AI-generated content is critical to empowering people with knowledge of when they’re interacting with generated media, and for helping prevent the spread of misinformation,” said Google DeepMind in a blog post. 

The same conversation is going on in the world of language models like ChatGPT, which produce text based on user inputs. 

Back in January this year, OpenAI released a classifier trained to distinguish between text written by humans or AI. However, when OpenAI tested the tool it was only able to correctly identify AI text in about half of the samples it analysed, and marked human written text as AI-written almost one in ten times and the project was curbed in July. 

This newest development by Google therefore is a leap in the right direction, especially in a domain as influential as images. For example, in May of this year, an AI-generated image of an explosion at The Pentagon – the United States’ Department of Defence headquarters – went viral on X, causing a dip in the stock market and anxiety around the country’s national security. 

The controversial role of it-generated content is hotly contested in the realm of creative work. The vitriol stems from creatives worried the AI-models are being trained on their copyrighted material without licensing or attribution. 

In the UK, there is a limited exemption to copyright when the data mining is for non-commercial use. The move was made in an effort to attract AI companies and development into the country. Now, the government has proposed a new exception to the copyright law that would reduce arts and cultural content to “inputs” for AI systems. 

Recommended

• What can be done about AI-generated child abuse images?
• Scottish startups to count African elephants from space
• 29,000 UK web domains suspended for criminal activity in past year 

“We’re committed to connecting people with high-quality information, and upholding trust between creators and users across society. Part of this responsibility is giving users more advanced tools for identifying AI-generated images so their images — and even some edited versions — can be identified at a later date,” said DeepMind.

Watermarking generative art is not a novel idea, many startups have already built similar tools. However, the move by Google is significant due to the mounting pressure on tech companies to establish a technology for themselves. 

In May, Microsoft committed to watermarking its images using cryptographic methods, while Shutterstock and Midjourney are to embed markers indicating AI generated content.