Behind, they are human beings located in Africa or Asia, a priori always made of flesh and bone and therefore not at all artificial, and “slightly” less well paid than the employees of Meta or Google.
The term artificial intelligence is a source of a lot of fantasies. The most alarmist and imaginative see in its development the premises for the advent of a conscious AI capable of controlling the world. Well, in reality, the overwhelming majority of AI systems are discriminating systems capable of distinguishing a dog from a cat, Peter from Paul, or answering a few basic questions. And far from being self-sufficient, artificial intelligence requires a multitude of very human brains to function and perfect. Humans very often exploited by large groups which, if they claim to change the world, nevertheless use an economic model that is not revolutionary. A survey of Vice takes us to the heart of the matter.
An invisible workforce
You may have already seen offers, and even participated in some, offering to repeat a few sentences or sort images for a fee. For Krystal Kauffman, turkish for seven years for Mechanical Turk (a micro-work platform owned by Amazon), and for tens of thousands of other workers, this kind of activity has been a daily life and a means of subsistence.
“ In the past, we have worked on several large AI projects. So there are tasks where people just have to repeat the same sentence six times, so it’s about training the AI to recognize different voices and stuff says Kauffman. ” So I have the opportunity to do a bit of everything, but there’s a lot of machine learning and AI data labeling going on right now. We have seen an increase in demand for this work “.
As you can imagine, these routine and unstimulating tasks do not interest large companies like Meta or Google; they outsource them to freelancers employed by companies like Sama, Alegion, or iMerit. These precarious workers, mainly from sub-Saharan Africa and Southeast Asia, strive to train AIs, for example to improve their image analysis or voice recognition skills. Indeed, when an image is sorted by an artificial intelligence algorithm, this process is in fact nothing artificial: on the contrary, the AI has previously “learned” to do it from the decisions taken by humans.
In practice, Krystal Kauffman and her working brothers are not informed of the outcome or purpose of their work. If they do not expect any recognition from the companies that solicit them, they are amused to see the ignorance in which bathe most end users at the other end of the chain.
” We’re used to working on things that we don’t know exactly what they’re for […]. And while I don’t need to be called an employee or anything like that, very rarely do you hear a big tech company acknowledging the invisible workforce that’s behind so much of that technology. They make people believe that AI is smarter and more advanced than it actually is, which is [la raison pour laquelle] we continue to train her every day Kauffman reports.
Remarks confirmed by Laura Forlano, associate professor at the Institute of Technology of Illinois. She explains : ” I think one of the myths around AI computing is that it actually works as intended. I think right now what human labor essentially makes up for is a lot of shortcomings in the way the systems work. On the one hand, the industry can claim that these things magically happen behind the scenes, or that a lot of what happens is IT. However, we know that in many, many cases, whether it is online content, how to operate an autonomous vehicle or medical devices, human labor is used to fill in the gaps in a system that is not really able to function independently “.
A supply chain, like others
Technology companies would thus give the illusion of a self-functional AI, masking a very different reality. In addition, this sector gives the feeling of being above ground, yet, as you now know, it is supported by tens of thousands of self-employed workers.
Kelle Howson, researcher at the Oxford Internet Institute, says: I think the public doesn’t have a good awareness that this is a supply chain. It is a global supply chain, it contains unequal geographical flows and relationships. And it relies on a huge amount of human labor “.
And like most global supply chains, like the textile chain, the AI chain is unbalanced. On the one hand, workers from developing countries who are the real fuel for its proper functioning, on the other, those who hold the controls as well as the users, located in developed countries.
” There are many more workers on micro-work platforms in southern countries than in northern countries. Thus, the majority of the labor supply on these platforms is concentrated in the South, while the majority of the demand is located in the North. sums up Howson.
Unenviable working conditions
Of course, between working standing in the fields, carrying heavy loads while being subject to the vagaries of the weather, and sitting in front of a screen, the second option seems less trying and more enviable. Physically, the two tasks are indeed incomparable. Nevertheless, content moderation can have psychological repercussions. Moreover, the status of these workers keeps them in precarious conditions.
” It is only a piece job where a person earns a part-time wage by spending an hour a day here and there, it is not an impact job because it does not does not really lead to career development and ultimately poverty reduction says Sara Enright, Project Director at the Global Impact Sourcing Coalition (GISC).
T. Okolo, a doctoral student in the Department of Computer Science at Cornell University and one of the authors of a paper on the role of the global workforce in the development of AI published in 2021, adds: ” Basically people [du Sud] are often paid on the basis of a so-called fair wage, based on the GDP or local income of their respective context. But the work is very routine, very manual and a bit tiring too, although obviously it cannot be compared to physical labor and plantation work in the colonial era. However, this work is entrusted to the same regions and similar companies “. He completes: Moderation is also a more disturbing activity for the workers themselves, who have to view different types of content all day long, which is mentally taxing for someone who is constantly exposed to it. “.
So, last May, Daniel Motaung, a former content moderator, filed a lawsuit in Nairobi, Kenya. It accuses Meta, Facebook’s parent company, and its main outsourcing partner, Sama, of forced labor, human trafficking and union busting. Accusations supported by an investigation by the Time Magazine. The remuneration of Daniel Motaung and his colleagues? $1.5 an hour. Not expensive paid for the moderation of content which resulted in two of them post-traumatic stress syndrome after they watched images and videos of rape, murder or even dismemberment on a loop.
You also work for AI… and for free
The more cynical among us may think that this is the economic reality of the world as it has always existed: exploiters and exploited, instead of masters and their slaves. But these people may not appreciate working for free, sometimes, for the tech giants.
Do you solve a CAPTCHA by selecting trains or planes from a panel of images, proving that you are not a robot? You have also contributed to improving an image discrimination algorithm.
Do you publish photos while striving to caption them correctly? Well, you certainly participate in perfecting deep learning models by populating databases. The LAION-5B collection (Large-scale Artificial Intelligence Open Network) for example brings together 5.8 billion captioned images, retrieved from all over the Web.
How to improve the situation?
The conclusion is there: is it possible to do better? Naturally. How ? In ” thinking seriously about how the human workforce gets involved in the development of AI. This workforce deserves to be trained, supported and compensated to be ready and willing to do important work that many might find tedious or too demanding. argue Mary L. Gray and Siddharth Suri, authors of the book Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass.
Overall, all of the speakers cited recommend greater transparency on the part of technology companies, the creation of policies that improve working conditions and wages, or even the establishment of training to allow individuals like Krystal Kauffman to contribute to AI models in ways other than labeling.
Furthermore, they suggest that everyone who contributes to forming an AI should be informed and recognized as such, and given the option to participate or not. If applicable, to be remunerated accordingly or to benefit from greater incentives.
Source : Vice
2