AI tools like ChatGPT can draft letters, tell jokes and even give legal advice – but only in the form of computerized text.
Now, scientists have created an AI that can imitate human handwriting, which could herald fresh issues regarding fraud and fake documents.
Amazingly, the results are almost indistinguishable from the real thing drafted by human hands.
Below is one column of writing by the team’s AI model and another by humans, but can you tell which is which?
Scroll down to reveal the answer!
In this image, one column of writing was done by an AI model while the other was by actual human hands – but can you tell the difference?
The artificial intelligence programme can learn the handwriting style of a person and generate text scrawled in what looks like their hand (file photo)
If you guessed that the left column was written by AI and the right column by a human … you are correct!
The right column is a sample of handwriting by six different human writers that the team’s AI was trained with.
Meanwhile the left column is the imitations of each person’s handwriting by the new AI, called HWT.
HWT was developed by scientists at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi.
According to the experts, the ‘handwriting’ generated by HWT looks much more realistic than other existing AIs and would be preferable to use.
In their study, they showed the fake text from HWT and two other handwriting generation technologies to 100 people and asked which one they preferred.
The results revealed that participants in preferred HWT to the other text generators 81 per cent of the time.
What’s more, the participants could not distinguish the mimicked handwriting from the actual handwriting.
Two more columns of writing, both generated by other AIs – GANwriting on the left and Davis et al (2020) on the right. In the study, the majority of people preferred the writing by HWT over these two
Previous approaches to mimic a person’s handwriting had been developed using a machine learning model called a generative adversarial network (GAN).
This technique has become famous in recent years for creating fake faces and making new music by being trained on existing samples.
Handwriting generated by GANs capture the overall, general style of a writer – for example, the slant or width of the letters.
But GANs struggle to recreate how people create individual characters, as well as the little lines, known as ligatures, that tie characters together.
Instead of GANs, the researchers used ‘vision transformers’, a type of neural network designed for computer vision tasks.
Vision transformers are able to understand that parts of an image that are physically distant from each other are uniformly linked.
‘To mimic someone’s handwriting style, we want to look at the whole text, and only then will we start to understand how the writer ligated characters, how the writer connected letters, or spaced words,’ said study author Fahad Khan at MBZUAI.
AI-created handwritten text could be beneficial for people with disabilities or injuries that prevent them from holding a pen, the team think.
It could also be used to generate a large amount of data to improve the ability of machine learning models to process handwritten script.
Pictured from the team’s paper is the overall architecture of HWT to generate ‘handwritten’ text images. The tool could help people who have injuries that prevent them from taking up a pen
Study author Hisham Cholakkal, assistant professor of computer vision at MBZUAI and one of the inventors, acknowledged the potential fraud concerns.
‘It’s important to be aware that it’s possible to use AI to generate handwriting that matches the style of an individual,’ he said.
‘We wanted to know if you gave a model a few samples of someone’s handwriting if the model could learn about the style of that person and then write anything in the handwriting style of that person.’
While the study focused on generating handwriting in English, the researchers are now interested to apply their technology to other languages, like Arabic.
The study has been published as a pre-print paper, meaning it’s yet to be peer-reviewed, on the open-access repository arXiv.
