AI Surveillance in China, Network Studies, and Game Theory
Link: https://www.nytimes.com/2018/07/08/business/china-surveillance-technology.html
Many previous studies on large networks were conducted using preexisting networks, such as Facebook or other forms of social media. These networks reveal very useful trends such as the six degrees of separations. But they fail to model real life in certain aspects, such as how Facebook allows for passive engagement, in which one can be connected with another individual via the news feed without direct communication. Online connections can also be formed with much less effort, and do not fade with time as in real life. The reason these networks are used as opposed to real life relationship networks is that it is infeasible to collect such tremendous amounts of network data by hand; one cannot ask millions of people to list out all of their friends.
In China, however, this non-feasibility may soon become feasible thanks to a massive, developing surveillance system consisting of millions of cameras, designed to track 1.4 billion people. In offices, for example, these systems have been used to track employee movement throughout the building for the entire day. Such data could be used to track which employees are spending time together and have a connection. The system is also capable of recognizing and matching faces with national I.D. numbers, although it requires the target to stand still for several seconds. This limitation means that China probably won’t be able to gather data to model massive networks yet. But AI is developing at a rapid rate, and it will most likely not be too long before China is collecting massive amounts of data on interpersonal data. It is both interesting and somewhat unsettling to see what will come out of such wide-scale implementation of government surveillance.
The new AI surveillance system in China also reveals a very interesting result of game theory. First, consider the real life implications of an AI surveillance system. Police officers should investigate people deemed suspicious more often with the AI system than without the AI system. The reasoning is that, using human judgment to see who is suspicious is less accurate than using human judgment plus a massive AI surveillance system to see who is suspicious. Thus with the implementation of AI, the payoff for investigating someone suspicious is increased for the police officer, because they will catch criminals more often. Consider a game between a citizen and a police officer. The strategy for the citizen is to either commit a crime or don’t commit a crime. The strategy for the police officer is to investigate or not investigate a citizen they deem suspicious.
Consider the first payoff matrix without AI surveillance. Although the two strategies by citizen and policeman are not done at the same time in real life, let’s say that they are for simplicity’s sake.
If the police investigates someone who committed a crime, the police gets payoff +X for doing a good job, and the citizen gets payoff -X for getting arrested.
If the police does not investigate someone who committed a crime, the police gets payoff -X because there is still a criminal loose, and the citizen gets payoff +X because they gained from the crime.
If the police investigates someone who did not commit a crime, the police gets payoff -X because they lost time that could have been used to investigate an actual criminal. The citizen gets payoff 0 because they do not lose or gain much from being investigated.
If the police does not investigates someone who did not commit a crime, the police gets payoff +X because they get time to use to investigate someone else, and the citizen gets 0 payoff because they do not lose or gain much from being investigated.
There is no pure Nash equilibrium. The mixed-strategy Nash equilibrium for the without-AI-surveillance matrix is p=1/2 and 1/2. There are some flaws with this result, due to many simplifications. But the point of all of this simplified modeling is to see what happens when the payoff for the policeman investigating is increased. Represent this in the payout matrix by turning the payoff of (Investigate, Commit Crime) into (+2X,-X) as shown in the with-AI-surveillance payout matrix. Keep the other payouts the same. The mixed-strategy Nash equilibrium for this new with-AI-surveillance payout matrix is q=2/5, p=1/2. In other words, criminals should commit crime less. And this is exactly what happened via the installation of cameras in China. In real life, the addition of these systems has led to reports of drastically decreased jaywalking and bike theft.