Finnish start-up Metroc has entered into an agreement with the maximum security women's prison in Hämeenlinna (Finland) to employ female inmates to train an artificial intelligence-based search engine, as reported by Wired. One of the inmates, interviewed by Wired's correspondent, described the work as “boring and repetitive”, but still better than “sitting in a cell all day”. This is not the first time that subjects with few other employment opportunities, such as refugees, citizens of poor countries and now prisoners, have been used to train AI, often with alienating and mechanical tasks. and paltry remuneration.
How to train an AI that speaks Finnish
Programs based on artificial intelligence can recognize images, faces and perform hundreds of functions automatically. To train AI, English-speaking companies call on workers from the South, who live in low-income countries. For example, OpenAI, the company that develops ChartGPT, uses an external company that hires workers in Kenya, Uganda and India. These workers are responsible for explaining to a program how to tell a dog from a car, or how to recognize if a post on Instagram incites hatred or racism. In short, it involves spending many hours in front of the computer cataloging images (labellyng) or answering simple questions in order to provide the AI with a database of information.
Maximum security women's prison in Hämeenlinna, Finland. Source: RISE for Wired
When AI works in English, there is no shortage of cheap workers. The Global South is full of English-speaking citizens willing to work for a few dollars an hour. But how to train an AI in Finnish? In the countries of the South, people speaking Finnish are few or even non-existent. Additionally, Finland is one of the richest countries in the world, with very high average salaries, so finding cheap workers to train AI is difficult.
Metroc leaders found the solution in prisons. Thanks to an agreement, Hämeenlinna inmates will be able to develop software in the Finnish language.
Prisoners train on artificial intelligence
Metroc is a Finnish start-up that has developed software that connects supply and demand in Finnish public procurement. The program analyzes a project and determines if it already has all the resources necessary to carry it out or if it needs to call on professionals. The algorithm is based on artificial intelligence and can determine whether a hospital already has an architect to design rooms or a supplier for doors and windows.
To train the AI, Metroc entered into an agreement with the high security prison in Hämeenlinna. Unlike other Finnish prisons, where inmates enjoy some freedom during the day, here inmates are never allowed to leave the prison. A Wired reporter interviewed one of the inmates participating in the program. She explained that the work consists of simple tasks, such as answering “yes” or “no” to questions on texts displayed on a computer. For example, a classic question is: “does the ad displayed refer to the housing market rather than a job offer? The sessions last a maximum of 3 hours per day, with remuneration of 1.54 euros per hour.
Prison management is promoting the deal with Metroc, saying it provides income for inmates and prepares them for the world of work that awaits them on the outside. However, the tasks are described by the interviewee as “rather boring and repetitive”. Furthermore, she also said that she did not know exactly what the work she was doing was for.
Public reaction
The initiative proposed by Metroc and Hämeenlinna Prison has met with a certain consensus among Finnish public opinion. Pia Puolakka, director of the Smart Prison project in Filnandia, was one of the first to join the initiative proposed by Metroc: “The goal is to increasingly introduce the world of the Internet into prisons , so as not to isolate the detainees from the world,” said Ms. Puokkala; in addition, “inmates are absolutely free to participate in the initiative, they can work in shifts for up to three hours and they also have other activities at their disposal,” she added.
Gig Economy workers demonstrate in California to demand more rights and protections. Source: Megan Rose Dickey/Protocol.
However, other voices were raised to criticize the project. “The narrative that we are moving toward a more automated and efficient society tends to forget that there are real humans behind many digital systems,” said Amos Toh, a Human Rights Watch researcher specializing in intelligence. artificial. Toh recalled that the sharp evolution of technology has led to a spasmodic search for cheap workers to train AI. Companies also seek labor from groups of people who have no other choice, such as refugees, asylum seekers, citizens of low-income countries and, in this case, prisoners.
Does training an AI allow you to acquire skills? Not really
Toh and the University of Helsinki researchers also pointed out that it is still unclear how repetitive, mechanical tasks performed to train an AI can prepare a person for the world of digital work. Human Rights Watch reiterated that more effective and proven initiatives could be prioritized to provide skills to prisoners, such as online programming and computer courses.
The Finnish public wholeheartedly accepted Metroc's initiative because they have confidence in the quality of the Finnish prison system and are convinced that the prisoners concerned are completely free to choose the activities they wish to do. But it is worth questioning the effect that a similar project applied in countries where prison systems are much harsher and less effective could have. According to a report from the ACLU (American Civil Liberties Union), in the United States, 75% of prisoners say they are forced to work. Under these conditions, even the training of an AI could become a form of exploitation for all purposes.
The Mechanical Turk
The social impact of the functioning of an AI can be explained by the story of the Mechanical Turk. In the late 1700s and early 1800s, a Hungarian inventor created a wooden automaton in the shape of a human dressed like a Middle Eastern man (the Turk) and powered by a box filled with gears automated. The Mechanical Turk was able to play chess thanks to the box's secret mechanisms and was such a powerful automaton that he beat many chess players of the time, much like today's AIs. Legend has it that he even beat Napoleon at chess.
Illustration of the Mechanical Turk designed in 1770 by Hungarian inventor Wolfgang von Kempelen. The original was destroyed by fire in 1854. Source: Joseph Racknitz – Humboldt University Library.
Everyone wondered about the secret of the mechanism. Eventually it turned out that there were no automatic gears in the box that drove the Turk, but simply a little man, very good at chess, who operated the automaton. Behind mechanical automation, there was nothing other than the work of a man, in flesh and blood.
In 2005, Amazon was inspired by this story to name its crowdsourcing platform, Amazon Mechanical Turk. Crowdsourcing platforms make it possible to organize and direct tasks carried out by thousands of humans around the world, and are used by many companies to train and collect data used by AIs.
The underpaid army that keeps AI alive
In summary, Amazon's platform makes it possible to break down millions of repetitive tasks, such as image classification or recognition of violent social content, into small actions executed simultaneously by thousands of workers. These actions are remunerated from a few cents each, reducing the cost of developing technology such as artificial intelligence for the thousands of workers who live in low-income countries or who do not have other solution to survive.
Companies that tout themselves as “FullAI,” and fully digital, are in reality supported by thousands of workers forced by social circumstances to perform repetitive, poorly paid tasks. These workers do not have the complete freedom to choose another job, as in the case of an imprisoned person, and can in some cases earn less than 2 euros per hour, while on the other side of the world , digital business leaders receive investments worth millions.