Artificial intelligence tools already allow people to generate eerily convincing voice clones and deepfake videos. Soon, AI could also be used to mimic a person’s handwriting style.
Researchers at Abu Dhabi’s Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) say they have developed technology that can imitate someone’s handwriting based on just a few paragraphs of written material. To accomplish that, the researchers used a transformer model, a type of neural network designed to learn context and meaning in sequential data.
The team at MBZUAI, which calls itself the world’s first AI university, has been granted a patent by the US Patent and Trademark Office for the artificial intelligence system.
The researchers have not yet released the feature, but it represents a step forward in an area that has drawn interest from academics for years. There have been apps and even robots that can generate handwriting, but recent advances in AI have accelerated character recognition techniques dramatically. As with other AI tools, however, it’s unclear if the benefits will outweigh the harms.
The technology could help the injured to write without picking up a pen, but it also risks opening the door to mass forgeries and misuse. The tool will need to be deployed thoughtfully, two of the researchers said in an interview.
“We’ll have to create public awareness and develop tools to combat forgery,” said Hisham Cholakkal, an assistant professor for computer vision at MBZUAI. “It’s like developing an anti-virus for a virus.”
Despite the concerns, the inventors said they plan to apply their research to real-world applications within months – and they’re looking for commercial collaborators. “From decoding doctors’ handwriting to crafting personalised advertising, the potential of this development is huge,” said Rao Muhammad Anwer, also an assistant professor of computer vision at MBZUAI. The technology could also be used to generate a large amount of synthetic data to improve how other AI models process handwriting.
There’s still more work to do, however. The researchers’ transformer model, trained from publicly available handwritten texts, can learn and write in English and, with some success, in French. But the team said they are still trying to crack handwritten text in Arabic.
