Born sixty years ago, artificial intelligence, more and more efficient,is now an integral part of our daily life. While some see it as a promise of emancipation, others point to the lack of ethics and morality of the programs. Between illusion and reality, investigation at the heart of algorithms…
You’ve certainly had to run a test – called reCAPTCHA – during an online query or purchase to certify that you’re not a robot. This is a way to protect websites from spam and other abuses. What you may not know is that by checking these images of bridges, trucks or tricolor lights, you are helping to improve algorithms, especially that of Google Street View, Google being the owner of reCAPTCHA. Similarly, when you tag a friend on Facebook or you like a video on YouTube, and more generally,with each of your clicks, you enrich an artificial intelligence.
Ditto, when you use any of your connected objects – soon 6 per person on average according to the forecasts of the firm Gartner – Smartphone, watch,speaker, TV or even car and even refrigerator. We are all artificial intelligence trainers that permeates our lives and yet we have a hard time understanding this technology born a little over sixty years ago. How is AI manufactured and what challenges does it pose? Diving into the other side of the scene.
If the horizon of researchers is to recreate the mechanisms of the human brain in machines, we en are still far from it. Because machines are incapable of reasoning. “They just bring out what they’ve been taught,” says Luc Julia, vicepresident of innovation at Samsung, also known as one of the creators of Siri, Apple’s voice assistant. The author of Artificial Intelligence does not exist (First Editions, 2019) prefers to talk about “enhanced intelligence.” AI is simply the latest wave of digital technology,” confirms Bruno Sportisse, president of INRIA, a research institute for digital science and technology, which brings together 200 research teams, half of which work in the field of this discipline.
Long confined to universities and the military before being invested by web giants – Google, Facebook, Amazon, Microsoft or Apple in the West; Baidu, Alibaba, Tencent in China , AI now permeates all sectors of the economy to the point of being dubbed “the electricity of tomorrow”. Businesses find a way to delegate difficult and repetitive tasks to the machine. They discover virtues in optimizing the customer relationship en by creating more and more natural interfaces.. They also see it as a valuable decision-making tool, in banking, insurance, healthcare… Not to mention the field of “smartobjects”, from the autonomous car to the connected refrigerator, which invade our daily lives..
Why such a tel success? Because in the last ten years each of the major components of AI has seen tremendous progress. Algorithms, the “kitchen recipe” based on statistics and probabilities that allows the machine to imitate human behavior, are becoming more and more sophisticated. While the available data (the ingredients)and the computing power needed to grind the whole thing have grown exponentially. “AI l’on is software bricks that are integrated with other software to make them more efficient,” says Alban Leveau-Vallier,a professor at Sciences-Po who is preparing a philosophy thesis on artificial intelligence and intuition at the University of Paris VIII.
Take the business of customer relations. They mainly incorporate language en processing tools to create their chatbots. Those in finance, who shuffle a lot of encrypted data, use machine learning to develop decision models. Industrial manufacturers, the other hand, mainly use robotics and artificial vision to design their machines. AI is everywhere, but it disappears by integrating into the digital environment.
At the root of this meteoric progress is deep learning. Built on neural layers, this machine learning technology has revolutionized AI by allowing software to “recognize”images, automatically translate documents, hear and decipher human speech by reconstructing, in their own nos way, our perceptions. It is these deep learning-educated algorithms that are, for example, able to drive autonomous cars today.
But to be operational, these little jewels, especially those that rely on deep learning, need thousands of hours of training. Because, unlike humans,the machine learns only by multiplying observations on a large scale. “To recognize a cat with 95% accuracy, it needs about 100 million images of that animal, where only two are enough for a child,” says Luc Julia. The more est abundant and relevant the raw material, the more artificial intelligence improves. “While algorithms are easily accessible in open source, data accounts for 80% of the work required to develop an AI,” says Alban Leveau-Vallier. If AI is the electricity of tomorrow, en the data, they, are the oil..
There en are several sources: online data, the Google Search database contains for example several hundred million images; the actual data, in the case of the autonomous car, it is collected by sensors arranged on all types of roads; or customer data with an advantage to companies with large files. “This is the case for insurance companies, l’IA banks, hospitals, transport or distribution specialists,” says consultant Olivier Ezratty in his opus on AI uses in 2019. It is still necessary toil extract, label (i.e.qualify) and validate this learning data. .
Who does this ant work? It’s us! “Training an algorithm still requires a strong human contribution,” admits Aurélie Jean, a doctor of science and entrepreneur who recently published On the Other Side of the Machine (Observatory Editions). ). That’s not much to say. While in many companies, the work of collecting and annotating content is carried out in-house by business experts, digital platforms, they, mobilize a lot of small external hands. . Sometimes, as we have seen, it is the Internet users who do the job for free through their l’a online activity; ; sometimes micro-self-employed, paid to the task,for example to surround images or to compare two videos.. They are recruited remotely through specialized companies such as Amazon Mechanical Turk, Clickworker or Figure Eight. Antonio Casilli, a French sociologist who specializes in social networks, denounces in the first case the hidden work of the”produsagers”and in the second case that, underpaid,of the “digital proletariat”.
These en “click-throughs” as the author of Waiting for robots (Threshold , 2019) are estimated to be between 40 million and several hundred million, mainly in Asia and Africa. The holy grail? Let the machine learn on its own, less supervised and less data-intensive way. Researchers are working on the issue. They test several avenues and, among them, reinforcement learning. . Thus,in the game of go, rather than forming on a history of games, the algorithm tests several scenarios by playing against itself.. Promising..
In the meantime, our lives are increasingly en governed by artificial intelligence, whether it’s getting a loan, getting a job or diagnosing our health. But how can we be sure that an algorithm is impartial? To the extent that it is designed and educated by a human, it can be expected to be perfectable. Examples of cognitive biases (created by the programmer himself) or statistics (generated by the data used) regularly make headlines. Amazon had to correct its recruitment algorithm based on the hundreds of thousands of resumes received over ten years. It selected mostly men who had made up the overwhelming majority of hiresin the past. .
In the United States, recidivism risk assessment software in the prison population was deemed unconstitutional because it overwhelmed black people. The parade against these sexist or racist algorithms? “Integrating ethics from the moment they are conceived,” replies Telecom Paris’s professor-researchers, David Bounie and Winston Maxwell, who wrote a report in March 2019 on the issue of bias. “Ethics, will inevitably be local as the cultures differ.”
But there is still a safeguard on which almost all countries agree, that of”explicability”. This neologism refers to the art of breaking down the steps of training an algorithm to better identify biases and correct them or, failing that, explain them. It’s not that simple. Because, based on probabilities and statistics,an algorithm works a bit like a black box: you enter a data and it comes out a result.. But we don’t know exactly what’s going on between the two… A bit like in en our brain: we know the processes and modes of operation.. However, it remains almost impossible to anticipate for sure an individual decision. In the United States, this virtuous approach to “explaining” algorithms is self-regulated.. ” For the past two years, all the big boxes, and among them the GAFA, have set up an internal ethics committee,” observe the two researchers..
In Europe and in particular in France, the law regulates artificial intelligences concerning the general public, especially in the field of education and health. “The result of an algorithm must be able to be understood and challenged,” insists David Bounie. In the private sector, the concern for transparency should lead “within two years”, predicts Alban Leveau-Vallier,to a certification of algorithms, granted by an independent body on the model of ISO standards. What about privacy? Europe is at the forefront with its en General Data Protection Regulation (GDPR) which came into force in 2018, requiring digital platforms to inform consumers about the collection of their personal data and its use. .
This text has just inspired California, whose Consumer Privacy Act est has been in operation since the beginning of the year. The aim is not to restrict companies but “to avoid rejections towards this technology”, explains Bruno Sportisse, promoting a so-called “trustworthy” AI. “. Bathed in algorithms since childhood, the younger generations will be all the more sensitive!
What threats to jobs?
The number of studies trying to assess the impact of AI on employment is no longer counted, threatening not particularly workers but employees and even managers. And according to these forecasts, the gaps are vast. From the most pessimistic, that of two researchers from the University of Oxford named Frey and Osborne who estimated in 2013 at 47% the number of jobs technically “robot is able” in the United States by 2030, to the most optimistic,that of the Forrester Institute in 2016, estimating at 6% net job losses. In the meantime, an OECD study recommended accounting for tasks that could be automated, rather than jobs. Repetitive chores would be delegated to the machine as a priority to direct employees to more rewarding tasks. .
In such a case, the challenge is to re-qualify these people by accompanying their rise in skills with training. The darkest scenarios fear the appearance of”a useless class,” as the historian Yuval Harari calls it. They advocate the allocation of a universal income to alleviate tensions. In investissement contrast, experts such as Taiwanese Kai-Fu Lee, who was Google’s president in China before founding his own private equity firm, are betting on creating new jobs, especially in the service sector.
The AI’s ecological challenge
It is difficult to assess the carbon footprint of AI, drowned in the mass of energy consumption of all digital uses. Some experts have attacked the subject by the small end of the eyeglass. In her Anatomy of an AI System, Kate Crawford, a Microsoft researcher and founder of the AI Now Institute in New York, dissects the entire manufacturing process for Amazon’s Echo connected speaker. To conclude the ecological aberration of this small object. In the same vein, researchers at the University of Massachusetts calculated that training a deep learning model for four to seven days to educate a personal assistant to recognize your voice and interpret a voice command consumed the equivalent of 5 cars during their lifetime!
We also know that AI fuels big data. The proliferation of data centers – several thousand worldwide and more than 130 in France – is expected to increase their weight in global energy consumption, currently estimated at 4%. The parade? Run these with renewable energy. And above all, deploy low-energy AI learning. “There is an active area of research on this subject,”says Alban Leveau-Vallier..