Behind artificial intelligence, the “little hands” of the shadows

The task promises to be easy: copy the contents of the photographed purchase orders into various boxes. Text entry, paid two cents Americans by good copied. It takes two minutes to fill in all the boxes. In one hour, it is therefore possible to earn 60 cents – or around 58 euro cents. As long as nothing crashes. Welcome to the world of click workers.

Unlike bike couriers, who provide real service but are algorithmically driven, click workers only operate online. Their office? Internet sites where they carry out simple tasks in exchange for often derisory remuneration. “With the Covid-19 crisis, which confined people to their homes and put many people out of work, the number of workers in these platforms have explodedsays Antonio Casilli, professor at the Polytechnic Institute of Paris. The Australian company Appen, for example, saw a 31% increase in activity. »

Amazon Mechanical Turk, the best known

The most well-known platform is that of the Amazon giant, Amazon Mechanical Turk. A nod to the mechanical Turk of the 18th century, this real fake automaton ” revolutionary “ behind which was actually a human. The current principle is the same: behind the artificial intelligencewe find small hands.

“On the Internet, it becomes difficult to disentangle what is a well-designed algorithm from what is human work.notes Pauline Barraud de Lagerie, sociologist at Paris Dauphine University. Microwork blurs the boundaries because there are ultimately tasks that, although repetitive, remain better done or simply cost less to have humans do! » Examples ? The anonymization of CVs, the writing of product descriptions, the copying of texts, etc.

The robots that “destroy work” are therefore actually called Linda, Carlos and Anirban, and come from the Philippines, Madagascar, Venezuela, and even France. “Microworkers work just as well in the countries of the North as in those of the South, but with different profilesanalyzes Antonio Casilli. In the countries of the North, it is more a question of women, graduates, between 24 and 44 years old, who have another activity elsewhere and supplement their income. In the countries of the South, it is younger men who do not enter the labor market. »

Human supervised learning

Between human and artificial intelligence, vagueness reigns. Behind a so-called translation software can sometimes hide a microworker, paid at low cost. But most of the tasks carried out by these small hands aim to build and ensure the expertise of real software. “The majority of AI relies on supervised learning, i.e. they are provided with data to learn and derive models from it”describes Laurent Vanni, research engineer at the CNRS and specialist in these corpora.

Take self-driving cars. To ensure driving, engineers must have a gigantic database including each road sign and its meaning. Gigantic, because the car will have to know how to recognize a stop sign, whether it is in a city with buildings, in open country, or even half hidden by a post. And what better than a human eye for that?

“The annotated corpora, crucial for AI, are made by humans, through these platformsconfirms Laurent Vanni. The quality of the annotation and the data used is essential to avoid bias. Many research works focus on the evaluation of this quality of corpora. » As such, the large international panel of digital workers is an asset: if it is a question of identifying an emotion, collecting the opinion of people of different origins makes it possible to limit cultural biases.

Except that again, the machine has the last word. To ensure that Internet users do not fill in anything, the systems work by comparing the answers. ” If 80% of respondents say the dog looks aggressive and 20% say it looks happy, the correct answer will be considered the first and those who answer the second will not be paidexplains Pauline Barraud de Lagerie. If common sense is misleading, the chosen answer will be wrong. » The quality of the answers can also be ensured upstream, by asking the person to fulfill certain qualifications. Then charge the human to prove to a machine that he is capable of performing a task to help another machine.

Captchas, to feed AIs

These tasks in themselves are so fragmented and distributed among microworkers around the world that it is impossible for them to trace what type of AI will serve their work. Sometimes, this AI development assistance even becomes completely invisible to Internet users. This is the case of “captchas”, these short texts that must be completed to validate online forms. The best known had been set up by Google for its Google Books fund. The works had been digitized, then each Internet user confronted with a “recaptcha” transcribed word by word, free of charge, the handwritten or poorly printed passage.

Same unintended AI involvement when it comes to selecting red lights or crosswalks on an image, or matching two images. “The development of AI can also be based on the content posted by Internet usersrecalls Pauline Barraud de Lagerie. For example, for a meal ordering AI, you can leverage restaurant reviews, if only to acquire the lexical field. “The algorithms are decorrelated from the corpora, which can be used for very different and sometimes very distant applications”, confirms Laurent Vanni. Like that of an unwitting click worker, your contribution is then broken up and reworked by other little hands in the digital chain.

We would love to give thanks to the author of this article for this incredible web content

Behind artificial intelligence, the “little hands” of the shadows

Discover our social media profiles , as well as other pages that are related to them.