Camera systems use artificial intelligence and promise to anticipate crimes. Surveillance cameras are among the most common solutions to try to discourage criminals. Now, instead of just recording what happens on site, this equipment is also used to predict robberies and other incidents. ? So-called predictive policing tries to identify non-standard situations before they result in crimes. With the support of artificial intelligence, this technology has been applied in increasingly complex scenarios. Images from cameras, previously analyzed by people, are now examined by systems capable of issuing alerts when someone, for example, remains standing still for a long time in the same place, in behavior considered suspicious. This type of tool attracts the interest of companies seeking to reinforce security at their facilities, but it worries experts, who warn of the lack of regulation and the risk of a state of permanent surveillance over people who are not committing crimes. How it works Monitoring systems use AI to identify suspicious movements Disclosure/NoLeak "The system connects to the camera, transforms the image into metadata and recognizes patterns. If the image pattern changes, it issues an alert", explains Nicolau Ramalho, founder of NoLeak, a company that developed a predictive policing system for private security. ? In this context, metadata is the result of converting video elements into numerical sequences. The systems are trained with images of situations considered normal and are then able to identify when something deviates from this standard. (learn more below) In the case of NoLeak, for example, the system does not take any action on its own when identifying a suspicious situation. Instead, it recommends that images be reviewed by responsible parties, such as the company's security chiefs. "The alerts go to a video management system, and the operator takes whatever action he deems appropriate, whether directly calling a police force or the private security team", he states. The company has more than 5 thousand connected cameras and 100 customers, including industries and condominiums. According to NoLeak, the system reduces operators' eye strain and allows them to monitor up to 10 times more cameras with the same number of employees. Other companies in the sector include Actuate, AxxonSoft and Coram, which promise to transform conventional security cameras into systems capable of predicting possible crimes. But experts point out that this use of technology can restrict citizens' right to come and go. "We are not only talking about criminals who will be effectively reached, but also about ordinary citizens who, on their way to work, may pass through a system that led to an error", warns Fernanda Rodrigues, research coordinator at the Internet and Society Reference Institute (IRIS-BH). For Fernanda, tools of this type demonstrate the expansion of the use of artificial intelligence in security, despite doubts about their effectiveness and impacts. "What we have seen is an increase in the search for tools and technologies that will serve as a solution to a public security problem, as if there were a silver bullet to solve something historic in our society." AI monitoring systems can identify employees without protective equipment Disclosure/NoLeak How crime is 'predicted' The most advanced predictive policing systems use machine learning models trained to recognize patterns in images. Instead of just following previously programmed rules, they analyze large volumes of videos to learn which behaviors are considered normal and then issue alerts when they detect situations outside this standard. When trained with images of a public road, the system learns to identify where the road is and which lanes vehicles should travel on. If a car passes on the wrong side of the road or on the sidewalk, for example, it can identify this change and indicate that something is out of order. "The system needs around two weeks of analysis of this metadata to form a database sufficient to recognize the image pattern", says Ramalho, from NoLeak. To identify possible crimes, the tool only analyzes suspicious movements of people, and not their physical characteristics, he explains. "It doesn't identify the face to analyze behavior. What it does is transform the image into metadata and recognize patterns." The company's founder highlighted that, in addition to reinforcing property security, the solution is capable of checking whether there are errors in a production line and whether employees are away from work. In these cases, the idea is to avoid unscheduled downtime and increase company efficiency. Systems can use artificial intelligence to monitor occupancy levels on industrial conveyors Disclosure/NoLeak Warning about lack of transparency For Fernanda, from IRIS-BH, evaluating the accuracy of a predictive policing system and the absence of bias depends on independent studies on its operation. "Opacity is one of the main problems, mainly because we don't know how efficient these tools are. In fact, what we know is that there are studies that indicate they are significantly flawed", he says. These limitations have already been identified in solutions used in other countries. "There are many examples in the United States of technologies that sought to use data about places with the highest crime rates, but ended up enhancing surveillance," he says. One of the best-known cases is that of PredPol, which gathered historical data, such as date, time and location of crimes, to guide where police officers should carry out patrols during non-occurring moments. The system led police officers to patrol even more neighborhoods with lower family incomes and a lower proportion of white residents. In practice,