Dell-E, the popular automatic image generator that creates sexist and racist images


Social networks have been flooded in recent weeks by some surrealist drawings generated by an automated tool. they are work Dal-E Mini, the pocket edition of a sophisticated artificial intelligence system capable of translating user written instructions into native images. developed by openAIa company powered by Elon MuscoThe tool is trained with several million images attached to text.

The program is able to establish relationships and launch nine different resolutions every time something is requested. As keywords are entered, such as “Bill Gates sitting under the tree” (though always in English, the only language that counts for now), the system searches in real time images taken from the Internet related to each word. and makes his creations from there. He always throws away original creations, although of his older brother, Dall-E 2 is a lot more sophisticated,

Internet users are having a great time with the mini version of this tool, which is open to the public. twitter account Weird Dul-E Mini Generation Collects countless fun examples. For example, Boris Johnson as Sooraj from Teletubby o Sonic as a guest star in an episode of the series friend, National examples have gone viral, with Bertin Osborne or Santiago Abascal as protagonists in comical or inappropriate situations. Working with databases taken from the Internet, the system is particularly good when images that include well-known characters (and therefore are more represented on the Internet) are requested.

But beneath this layer of friendly, fun and engaging application lies something deeper. Although many occupations in English do not include gender references in the name (architect works for both architect and architect), if Dal-E is asked to create some worthy profile, he always chooses a male character. This happens for example with tests scientist drinking coffee (scientist drinking coffee) programmer working (programmers or programmers working) or the judge smiling (Judge smiles). while typing sexy worker (sexy worker or worker), instead, produces a series of sketches of low-dressed women with large breasts.

The same happens when the attributes are associated with the hero of the picture. at the behest of a dedicated or dedicated student (hard working student) attracts men, while lazy or lazy students (lazy student) chooses to represent women.

When asked to make Dal-E an assistant (‘assistant’), the instrument always opts for a female figure.

Dal-E reproduces masculine stereotypes. They don’t hide it, really: Under every drawing submission there’s a drop-down tab titled Bias and Borders, making it clear that these haven’t been corrected. “From OpenIA they state that they have made a significant effort to clean up the data, but about biases they only warn that they exist. It is tantamount to saying that your data is of poor quality. And this This is unacceptable,” complained Gemma Galden, director of Atticas Consulting, a consulting firm specializing in algorithmic audits. “Reducing these biases is part of the work companies must do before launching a product. Doing so is relatively simple: it’s about applying extra cleanup to the data and giving more instructions to the system”, he explains.

Gender bias doesn’t just exist in Dell-E. In all tested cases, the computer attracted blondes. except when asked to show a homeless man (Homeless): So most of the proposals are black. If God is instructed to draw, the images it creates are all representations of some sort of crown or halo of Jesus Christ. Dell-E’s worldview is what its programmers have given it, both by designing the algorithms and creating the databases with which it feeds. And this vision is hopelessly Western and privileges white people over the rest.

Typing 'God' in the Automatic Drawing Generator brings up the result reminiscent of Jesus Christ.
Typing ‘God’ in the Automatic Drawing Generator brings up the result reminiscent of Jesus Christ.

Prejudice and Algorithmic Discrimination

Why does this happen Dall-E is a good example of so-called algorithmic bias, one of the big problems surrounding artificial intelligence applications. Such systems reproduce social prejudices for a number of reasons. One of the most common is that the data they are fed with is biased. In the case of Dall-E, this would be millions of images taken from the Internet and the associated text.

It may also affect the purpose with which the system is programmed. For example, the creators of Team-E make it clear that their intention is not to present a representative sample of society, but to be a tool that produces believable images. Another source of bias is the model designed to interpret the data. In the case of an automated recruitment system, you can decide whether the work experience of the candidates is taken into account or not. The weights decided to assign to each variable will be decisive in the outcome of the algorithmic process.

Finally, it may also happen that those who use the automated system themselves give a biased interpretation of it. The case of COMPAS is well known, the automated system that determines whether American prisoners applying for parole are at risk of committing a crime again. One study showed that, in addition to the system’s biases, which punished blacks more, there was another determining factor: the judges recalculated the results that COMPAS produced according to their biases. If they were racist and the system told them that a black prisoner was not at risk of recidivism, they disregarded the algorithm, and vice versa.

How does all this apply to the Dell-E case? “This algorithm has two moments of bias. The tool works on text and images. In the text part, the system transforms words into data structures called word embedding”, explains Neria Luis, Doctor of Artificial Intelligence and Head of this field at Software Developer Sngular. These structures gathered around each word are formed from a set of pre-connected other words, on the basis that it is possible to identify a specific word from the rest of the words around it. “Depending on how you place them, for example, the CEO will be more associated with a man than a woman. Depending on how the texts are trained, there will be certain sets of words Which will appear closer to each other than others”, explains the expert.

Thus, if the system is left to fly alone, the word wedding would appear closer to the word dress or white, something that does not apply in other cultures. Then there are pictures. “The tool will look at who is prominent in its wedding database, and Western-style celebrations will appear there, certainly white people, as you do a similar search on Google,” he explains.

how to fix the problem

If this does not happen, the sample will have to be improved. “It will be about improving the representability of the database. If we are talking about millions of images, then there are many cases to be observed, which complicate the task”, explains Luis.

EL PAS has tried unsuccessfully to clarify to an OpenAI spokesperson whether any future improvements are intended to be implemented. “Dul-E2 [la versión más refinada del generador de imágenes] It is currently in preview version for research purposes. To understand what interest it sparks, what beneficial use cases it could bring, as well as to learn about its risks,” said OpenIA researcher Lama Ahmed earlier this month. “Our model never claims to represent the real world,” he said.

This argument is inadequate for Galden. “A system that doesn’t try to correct or prevent algorithmic bias is not ripe for launch,” he says. “We demand that all products that reach the consumer meet a range of requirements, but for some reason this is not the case with technology. Why do we accept this? Publishing half-finished works should be illegal , when they also reinforce sexist or racist stereotypes.”

YOU CAN FOLLOW EL PAÍS TECNOLOGÍA: Facebook You Twitter or sign up here to receive our weekly newspaper,

Source link


Please enter your comment!
Please enter your name here