Artificial intelligence (AI) is more human than we think. The development of many systems for image recognition, text analysis, sound manipulation, etc. requires the work of essential “small hands”. The young sociologists Maxime Cornet, doctoral student at the Interdisciplinary Institute of Innovation, and Clément Le Ludec (Centre for studies and research in administrative and political sciences, in Paris), who defended his doctorate in March, tried to understand their role by interviewing, since 2021, around twenty companies in this sector in France. This led them to study seven of their contractors in Madagascar, as well as around two hundred of their employees. In the newspaper Big Data & Society, they published, in 2023, with Antonio Casilli, “The problem of annotation. Human work and outsourcing between France and Madagascar”.
Why do artificial intelligence systems need small hands?
Clément Le Ludec: These techniques are used to classify, detect, etc., according to learning principles. Large quantities of so-called training data – images, videos, texts, etc. – are used for their development, in order to be able to generalize the responses to new data. Humans are therefore essential to train AI, either to generate data, for example by filming themselves walking in front of a camera, or to verify that the model’s predictions are correct. But the main activity consists of annotating texts or images, in order to build the learning corpus, for example by indicating on the photo of an intersection which road signs are there, or by identifying traces of rust on photos of electrical poles, or by spotting whether a customer is stealing from a store. Even so-called generative AI is affected. ChatGPT required a lot of annotation to teach the program what is an acceptable response or not, according to a certain scale of values. In our database of companies using these human tasks, a third belongs to the automatic language processing sector.
Maxime Cornet : In this host of human activities, we have even seen a fourth activity, the most “extreme”, which is to hire people to replace the software and make the client believe that there is artificial intelligence behind it.
How is this invisible work organized?
M. C. : Some companies keep these tasks in-house, especially if the data is sensitive. But many tell us that for this repetitive and arduous work, which can consist of viewing several hundred images per day, they cannot find anyone in France. Hence the outsourcing that we have seen to specialized companies in Madagascar. To our knowledge, no quantitative study exists to estimate the share of this outsourcing, but in our database of around twenty companies, two thirds use this subcontracting for this work on data. We also estimate that the latter represents 5% to 10% of the cost of AI software. The development of artificial intelligence does not mean job losses due to automation, as some argue, but rather their displacement to developing countries.
You have 63.83% of this article left to read. The rest is reserved for subscribers.
