Advances in computer vision are focused on developing mass surveillance technologies.

The decisive ball in the final tennis match between Sinner and Alcaraz . The umpteenth offside called on any opponent of Hansi Flick's Barça team. Or the automatic correction of postures to avoid injuries like the one that kept Carolina Marín from winning gold at the Paris Games. Computer vision is already being used for all of this. This branch of artificial intelligence (AI) is also behind the automatic identification of friends in photos, self-driving cars , and other emerging revolutions, such as advanced robotics or the discovery of new proteins. But the coveted object of artificial vision (AV) is humans: according to a study published in Nature , the leading scientific journal, the vast majority of the studies ( papers , in the jargon) and derived patents focus on people, on identifying their different parts, what they do, and the environments in which they move.

Researchers from several universities in the United States and Europe have compiled thousands of papers and also thousands of patents filed or registered since 1990, when research in computer vision was still bordering on science fiction. Back then, extracting information from an image or video was a challenge for machines. It still is, as demonstrated by the use of visual CAPTCHAs to confirm that we are human. However, AI has advanced both in its ability to analyze data and its ability to interpret it. There are many fields in which computer vision is being used . It's helping humans . But it also has a downside, and it's huge.

Using around thirty keywords, some as explicit as surveillance or facial recognition and others less obvious, such as iris or airport , the researchers analyzed 19,000 papers on AV presented at a major annual industry conference held for more than 30 years. They did the same with 23,000 registered patents derived from those works. The former disclose the advances achieved, the latter protect them so that only their authors can profit from them.

The authors summarize the results in their Nature article: "We found that 90% of the articles and 86% of the derived patents extracted data related to human subjects." The majority of the papers, more than two-thirds, focused on obtaining information from the human body as a whole or from specific parts of it, especially the face. Another portion of the articles and patents (18% and 16%, respectively) extracted data from human spaces such as work, home, or transit areas. A smaller, but not insignificant, 1% (5% for patents) falls into the category that the researchers called "socially relevant human data," which would seek to enable machines to understand mental states, economic status, cultural affiliation, etc.

It is by putting these results on a timeline that we detect the emergence of an entire ecosystem of automated mass surveillance technologies. In the early 1990s, they found only a few hundred papers and derivative patents in the field of AI, and barely half were focused on analyzing human data. Two decades later, the annual production of papers and patents had tripled, and 78% were now focused on human data. Combining the absolute increase in machine vision research and the relative increase in human-centered work, the science of AI surveillance has grown fivefold.

“It is difficult to establish a clear cause for the significant increase in the production of surveillance-related articles and patents,” acknowledges Abeba Birhane, a researcher at Trinity College Dublin (Ireland) and senior author of this study, in an email. However, she points to several factors: “The technical advancement of the field of computer vision itself, the lack of a critical and socially conscious culture within the field, institutional structures that reward surveillance, or the magical thinking surrounding technology that assumes surveillance technologies are the solution to political and cultural problems.” She also mentions the significant increase in funding due to the interest invested by powerful organizations that benefit from surveillance, whether for power, control, or financial gain.

Another data the authors have revealed is the concentration of research in the field of AI. Leading the way is the United States, producing more research than all the others combined. Somewhat distant, but far ahead of the third-placed country (the United Kingdom), is China. This shouldn't be surprising. On the one hand, they are the two technological superpowers, regardless of the purpose or use of the technology in question. Furthermore, the authoritarian nature of the Chinese regime and the connection between the US military and defense industry and cutting-edge science would explain the rest. In fact, two of the most relevant players in next-generation AI (not necessarily related to computer vision), Palantir and Anduril, are collaborating with US national security . The former is already developing a system for the Trump administration targeting immigrants. And OpenAI, the creator of ChatGPT, recently signed an agreement with the US Department of Defense .

In a commentary also published in Nature , Monash University (Australia) researcher Jathan Sadowski, who was not involved in this study, discusses one of its findings. The authors analyzed the evolution of language used in papers and patents, detecting a tendency to obscure their purpose with increasingly neutral words or to eliminate them altogether. For example, references to humans appear less and less frequently, even though they are the subject of the studies, and are being replaced by terms like "objects ."

“Referring to humans as objects could be considered simple engineering shorthand,” Sadowski writes. It would fit into a more general trend in technological language. But for this scientist, these dehumanizing practices cannot be lightly ignored: “These technologies are created in a political and economic landscape in which the interests of large corporations and military and law enforcement institutions have enormous influence on the design and use of AI systems.”