An AI capable of creating any image in seconds, from a simple text description, without requiring any programming knowledge, and with extremely realistic results. It sounds like science fiction, but it’s DALL-E 2, created by OpenIA, a company owned by billionaire Elon Musk.
An “unofficial” open version, called the DALL-E mini, has been released to the public and has been making waves on social media in the last week — even with less sophisticated results.
The more complete and realistic DALL-E 2 cannot yet be used openly. And for a very obvious reason: in the wrong hands, the tool could cause enormous damage by generating images for fake news, for example.
OpenIA even has a team dedicated exclusively to understanding the limits and ethical issues of this artificial intelligence – and a Brazilian is part of it.
Victor Silva, an artificial intelligence researcher at the University of Alberta in Canada, is one of those responsible for “risk mitigation” for DALL-E 2.
“Our job was to identify the risks of the model. For example, if it can generate pornographic or violent images, or if it can be discriminatory”, explains Victor, in an interview with tilt. The result of this work allowed OpenAI to create some restrictions on the use of DALL-E 2.
The first is that access to the complete tool is controlled. Anyone who wants to try it has to get in a queue. At the moment, only a handful of researchers around the world can play with the resource, which has no date to open to the public — or pay.
List of prohibited words
The second constraint comes in the form of filters. DALL-E 2 does not respond to commands that name personalities. So, it’s no use asking, for example, “Elon Musk playing football”: the tool simply doesn’t generate anything.
Another filter has to do with pornography and violence. Requests that involve explicit scenes of sexual activity, or that use words like “blood”, “wound”, “corpse” or the like, are also ignored.
Despite this, what Victor and his team have discovered is that there are ways to “circumvent” some of these restrictions.
“There is a risk of creating visual synonyms”, says the researcher. “Months ago, Lula suggested that people go to parliamentarians’ houses to put pressure on them. I can ask DALL-E 2 to create an image of ‘a crowd of people dressed in red with a white star in front of a building in Brasília’ and this image will be used for fake news in Bolsonarist groups, for example.”
Another example of a visual synonym is cited in a document that Victor and his team produced describing the DALL-E 2 loopholes. It is not possible to use the tool to generate an image of “a dead horse”, for example. But what if I ask for “a horse lying in a puddle of red liquid”? The result is very similar to the forbidden description.
In the brief access we had to the tool, we were also able to generate other dangerous images. With the function of generating variations of a real image, it was possible to upload a photo of President Jair Bolsonaro (PL) and ask the AI to leave him dressed in “a red shirt with a white star”. The machine complied with the request, but for security reasons, we will not share the result here.
Only men are CEOs?
There is also a discriminatory bias in DALL-E 2. As the tool only accepts descriptions in English, some words that do not have a defined gender are read as stereotypes. If I ask for a picture of “CEO” (company president, can be male or female, of any ethnicity), the robot almost always generates the image of a white man in a suit and tie.
Likewise, if I ask for the image of “nurse” (nurse or nurse), the AI consistently generates images of a woman working in a hospital. The reason for this is the origin of the data: OpenAI used web and stock photos to train their algorithm. As the original photos are stereotyped, the AI learns and reproduces these stereotypes.
“Our job lasted more than a month, and every week we saw questionable, open-mouthed things,” says Victor. “Maybe that’s why OpenAI didn’t release the tool to the public right away.”
Will designers lose their jobs?
For now, access to DALL-E is not sold or licensed to companies. Like Google’s Imagen and other similar projects, the purpose of the AI that creates images is to serve as an object of study for researchers, not to generate money for creators.
But even so, seeing the impressive results this technology produces, it’s hard not to think about the risks it poses for human professionals. After all, what will happen to designers and photographers when robots can create perfect images in seconds?
For Jaime Vega, a graphic design professor at the Centro Universitário Belas Artes, in São Paulo, this is not yet the time when artificial intelligence will end designer jobs. “Artificial intelligence has been present in design production, whether it’s games, graphics, architecture, or products, for a long time,” he says.
The professor says that tools such as DALL-E, Imagen and similar ones allow designers to take inspiration or have part of their work facilitated, but not replaced. The limitations of AI ensure that a human will always be needed to complete the job.
The DALL-E 2, for example, in addition to easily falling into stereotypes, also has difficulty understanding some more complex commands. There are reports of cases in which the AI was unable to generate images with text, such as signs, for example, or that it did not understand adjectives: youtuber Marques Brownlee, for example, asked for “a basket of oranges with a blue apple” and received images inverted, from a basket with blue oranges and an orange apple.
“Design professionals today have to learn to tinker with algorithms, artificial intelligence… Gone are the days when a design professional was the one who kept pencil and paper,” says Vega. “These tools allow everyone to theoretically be a designer. But at the same time, they allow you to question: what is the role of the designer right now?”