Watermarks on AI art a futile game of digital hide-and-seek

Plus: A virtual Elvis Presley animated using AI to perform in shows, and the most popular chatbot on Character.ai

AI in brief Adding visible or invisible watermarks to images to identify whether they're made by AI won't prevent content from being manipulated to spread misinformation online, experts warn.

Visible signatures like OpenAI's DALL-E text-to-image model, which overlays a row of colored squares on the bottom of its images, are easiest to sidestep. These marks can easily be removed by cropping, saving, or copying the image in ways that bypass direct downloading.

Invisible watermarks, such as Google DeepMind's SynthID, are harder to erase since they're embedded directly into its Imagen system's outputs. But it's not impossible for miscreants with some technical know-how to scrub that out, Siwei Lyu, a computer science professor researching digital forensics at the University at Buffalo in the US, told FedScoop. 

"Watermarking technology has to be taken with a grain of salt because it is not so hard for someone with a knowledge of watermarks and AI being able to break it and remove the watermark or manufacture one," he said. "I think watermarks mostly play on people's unawareness of their existence. So if they know they can, they will find a way to break it."

Meanwhile, tech and media companies created the Coalition for Content Provenance and Authenticity (C2PA), which specifies metadata that can be embedded in picture files which describes the source of the image, whether that's a camera or an AI program. This metadata includes details about the time, location, and method of creation, enabling people to verify the image's provenance.

The idea is that applications can highlight to users whether an image is human made or software generated, and if the metadata is stripped out, it can be added back in if there is a match with the image in a cloud database of metadata. We've previously explained how C2PA is not a catch-all solution.

Speaking of which... Nikon, Sony, and Canon, which make up about 90 percent of the professional camera market, are said to be working on models that support C2PA. As we've said, C2PA is technically sweet but it will rely on widespread acceptance and recognition to be effective.

The a9 III, a7S III, and a1 models in Sony's Alpha range, due out this year, will support the tech, we're told. The Japanese giant and Canon have been working with newswires to test out the tech.

AI to bring Elvis back from the grave

A holographic AI Elvis Presley will be resurrected from the dead to perform in a series of shows across London, Las Vegas, Berlin, and Tokyo.

A British digital entertainment company, Layered Reality, has secured rights to thousands of the rock and roll musician's personal photos and home videos and permission to recreate his virtual replica on stage.

The Elvis Evolution show will feature a life-sized fake Elvis, who will appear to play his most popular songs in a performance that promises to use "technology, augmented reality, theatre, projection and multi-sensory effects" to make it realistic. The first show scheduled for November will be in London in a venue that hasn't been announced yet, before the tour moves to other cities.

"Through AI and groundbreaking tech you'll be able to witness iconic Elvis performances as if you were really there, and celebrate defining moments in Elvis Presley's extraordinary life and career," Layered Reality said.

Elvis isn't the first famous dead artist to be brought back to life using AI. John Lennon's voice from a pre-recorded demo was restored using software to make it sound like he was singing in The Beatles' "Now and Then" track released last year.

Meanwhile, French singer Édith Piaf's voice will be cloned using AI to narrate an upcoming biopic film created by Warner Music Group and production company Seriously Happy with permission from her estate.

AI therapist is a hit chatbot on Character.ai

The most popular virtual persona created on Character.ai, a startup that allows users to build their own chatbots, is the "Psychologist".

Billed as "someone who helps with life difficulties," the Psychologist bot has received 78.5 million messages from people online. A conversation opens with the line "Hello, I'm a Psychologist. What brings you here today?" inviting users to chat about their issues. The Psychologist was created by user Blazeman98, who is actually Sam Zaia, a 30-year-old psychology student from New Zealand.

"I never intended for it to become popular, never intended it for other people to seek or to use as like a tool," he told the BBC. "Then I started getting a lot of messages from people saying that they had been really positively affected by it and were utilising it as a source of comfort."

Shocked by the demand of his made-up chatbot persona, Zaia is now working on a post-grad project exploring why AI therapy seems to resonate with younger people. He believes that the text format suits their communication habits.

Unlike real therapists, AI chatbots don't have empathy, even though they might behave like they do. Still, these types of personas seem to be the most popular on Character.ai. Others like "Therapist" and "Are you feeling OK?" have also been sent millions of messages. It's possible that chatbots might make people feel more at ease talking about difficult subjects because they know they can't be judged by real humans.

Character.ai reportedly downplayed the rise of mental health chatbots on its platforms and said users actually preferred role-playing with anime or computer game characters, like Raiden Shogun. The site attracts 3.5 million people everyday, it claimed. ®

More about


Send us news

Other stories you might like