Class# 2: Frameworks for Understanding Law’s Role in Society + Autonomous Weapons
Reconciling Impossible Expectations
By Michael Fischer & Shreyas Parab
As with any subject, it is important to limit the playing field. Too often the scope of expectations is just so high that meeting those expectations is not only difficult to achieve, it is impossible. In a diverse class composed of law school students, PhD students in computer science, political science, education, economics, medical students and undergraduate students is it important to define what will be taken from the class and the impact of the material.
The class will be focused on appeasing both the lawyers’ desire to know about the current and future states of AI and how it will impact the regulation and cases they will go on to work with and the non-lawyers’ desire to understand legal systems and how governmental institutions go about regulating the fields they work in and perhaps see artificial intelligence being used on a daily basis. At the end of the day, both groups will walk away understanding artificial intelligence and just how exigent and important questions facing regulating artificial intelligence are and just how interdisciplinary it needs to be.
How will regulation impact the people taking this class. Perhaps it is the doctor that becomes an ophthalmologist who uses computer vision to diagnose progression of diseases. Or is it the law student who goes on to work in criminal law who encounters algorithmically determined risk assessment tools daily (these are real motivations from students who are in this class at the moment), and needs to understand the “political economy” that surrounds how artificial intelligence is used and its limitations in order to understand why the law works that way or why it doesn’t work at all.
Each group of students bring with them a diversity of experiences and backgrounds that contribute different perspectives which will cumulatively result in a thorough understanding of the subject matter which is regulating artificial intelligence. Professor Céllular posits that this is the only solution going forward: that the nature of these discussions necessitate cross disciplinary collaboration and that the results will manifest differently, but what matters is that these frameworks are precipitated in a systematic way and abided by.
Bridging psychology, politics, and economics
How do we reconcile the views of different groups of students? How do we reconcile the view different views of different intelligent systems? While psychologists, politicians, and economists work on similar problems, they each have a different lens by which they spot issues and develop solution. In addition, when they work together, they each don’t fully understand the others. Each reason about their field differently. What is interesting is how each is able to work together and persuade the others of their ideas. How do these different groups talk within their group and how do they talk to other groups? With the advent of the internet the way in which different groups interact and make decisions is more important than ever.
One way to understand these interactions is through a game theoretic lens. Game theory gives a framework whereby individual agents are modeled as rational decision makers. While the agents can be simple, the interactions between these agents quickly becomes complex. Game theory though does not model everything and we see a divergence in what game theory predicts as optimal and human psychology? Take for example the Prisoner’s dilemma. It shows why two rational individuals might not come to an to an optimal outcome even if it is in both their best interests. If decision making is complicated for individuals, it becomes exponentially more complicated when taking place between large organizations,and governments. Only so much is able to be modeled and the model of such systems will always be falling short. How do all of these organizations actually work, mathematical or not. Is there a unified theory that takes into account human cognition?
Case Study: AI in the Government
After finishing law school, Professor Céllular took his talents to the Treasury Department, specifically working with the undersecretary of Treasury. The undersecretary of the Treasury also manages the Secret Service, which is responsible for investigating financial crimes in the United States dealing with counterfeiting, money laundering, and fraud. Professor Céllular became interested in a simple problem at the Treasury. The Treasury has a program called the Currency Transaction Report that collects data on all transactions over $10,000. They then use an algorithm to determine which of these are suspicious. In addition they have another program called the Suspicious Activity Report that banks need to submit to the Treasury for all transactions that appear suspicious.
From these two programs they are able to collect massive numbers of transactions from banks. He became interested in what are the risks of such reporting systems as they related to how they could be biased against people of certain races or genders. But much of that data was being warehoused and not being acted upon thus being underutilized. Professor Céllular noted how easy it was to write high level speeches for the under secretary about how collecting that data and creating Suspicious Activity Reports were good, but when it came to making policy and implementing, it became much harder to determine what “good” was and how it could be achieved.
Before he was able to solve the problem, he moved from the Treasury to the White House where he was Special Assistant to the President for Justice and Regulatory Policy and leading the Domestic Policy council staff on a variety of agency issues. Here he was able to get first hand knowledge of how difficult it was to reconcile political, social, and technical needs. He was one of the nexuses between senior executive staff (including the Chief of Staff and President) and the heads of the agencies to help them understand the implications and political economies around their actions. Several agencies came to him with creative and innovative ways of using artificial intelligence. For example, the Department of Agriculture wanted to maps agricultural patterns using AI. Or the National Oceanic and Atmospheric Administration wanted to boost clam production while making clams less expensive. Unfortunately, he had to work with them to marry the technical goals with the realities of the political economy and social landscape. For example, what the one agency might not know is that their proposal for innovating might be greatly beneficial to them, it could hinder progress on a completely unrelated initiative because we live in a world where not all factors are independent of each other and the interrelation of innovations and progress have unsuspected connections.
It is not possible for a single agency to try to know the rest of the landscape because that detracts from their essential role of finding the technical implementation needed to get policy enacted and rolling, yet it is also relevant to them to get things done. The onus on these agencies almost always result in subpar performance compared to the high expectations set upon them.Humans sometimes struggle to find these connections and relationships that are obfuscated to the everyday viewer, but computers on the other hand are extremely good at it… in fact there’s an entire chunk of machine learning which focuses on finding relationships in data.
Enter the concept of how intelligence is partly “social” and how in order to make realistic expectations of AI, we need to help instill this contextual and multi-faceted stream of inputs with carefully defined and designed objectives that are not simply limited to what humans WANT, but what humans don’t even know they want. For example, if you tried to create an artificial intelligence system that provides recommendations to open-ended prompts and asked it to find the best solution to stop climate change, it would probably tell you to kill all humans. Of course, humans would protest and say, “Well, we didn’t mean that. We can’t kill humans.”’ Then, the system might recommend, “Stop humans from reproducing.” Of course the system understands that it is not killing humans directly per se, just preventing births from happening. This would keep going as humans constantly try looking for what they understand to be “the right answer”.
Social AI
In order to create realistic expectations, we have to go beyond the individual and understand how groups make decisions. Decision making is heavily influenced by social psychology. Individuals make decisions very different than how groups make decisions. Groups make decisions based on varying interests and emotions that are present.
How groups make decisions has been studied by many academics and has different schools of thought. John Dewey, an American social theorist, found that when people discuss their ideas through a process known as deliberative democracy, it not only shapes the decision, but it also shapes people’s interests. He considered participation, not representation the essence of democracy. When people deliberate about something the outcome becomes more refined. Each person is able to learn from the other.
Another theory of decision making was developed by Marquis de Condorect and is called Condorcet’s Jury Theorem. It is a theorem about the probability of a group of people arriving at the correct decision based on the size of the group. However, it assumes all people are acting independently which is often a false assumption.
Another framework is Max Webers’ Theory of Organizations seeks to find an ideal framework through which decisions are made. He identifies hierarchy of command, impersonality, written rules of conduct, advancement based on achievement, specialized division of labor, and efficiency as being the keys to good decision making. However, some criticize these rules as being overly simplistic and analytical and don’t take into account the numances of people and their emotions.
How people make these decisions though also has a darker side. In group settings, we run the risk of social conformity. Groupthink can take over, and a group of people can partake in an irrational or dysfunctional decision-making outcome.
Deliberative democracy, is a form of democracy where deliberation is central to the decision making process. Deliberation is key to how humans and computers make decisions. Within a neural network, neighboring nodes deliberate with each other through a process known as backpropagation. The result of this that by the end of training, each of the nodes attains a final weight and the group of nodes is able to accomplish a task that none of the node would be able to individually.
Deliberative democracy happens within AI neural networks and also across AI neural networks. It is not enough to just achieve a singular goal, the emphasis is on adapting to the constantly changing environment and landscape that exists in order to achieve optimal outcomes for the entire group. Although humans can try explicitly setting the rules of the game, when there are multiple AI agents pitted against each other, emergent behavior happens that seem to “break the rules” or find new workarounds the humans could not have possibly thought of.
Case Study: OpenAI’s Hide ‘N Seek
Recently, OpenAI created a virtual hide and seek game. The system uses a machine learning process called reinforcement learning. Reinforcement learning uses individual agents within a model world. The agents explore the world and take different actions to achieve an optimal goal. In this specific example, the goal of the “hider” agents is to protect themselves from the “seekers”. While each of these two agents has a simple objective, the emergent behavior of how these two groups develops into surprising outcomes as shown in the video.
We find collaboration among the two hiders appear quickly, then coordination of defense mechanisms including evasion techniques, while simultaneously, the seekers are developing targeting methods given their nearby surroundings including the ramp.
This system will result in the most efficient outcome but can come at the cost of breaking the rules. Of course in this innocent example of hide and seek it means no worries, but in the context of autonomous weapons, for example or genetic engineering, we find that this emergent behavior could have deadly consequences. This computer simulation gives us a ready example of how when multiple groups make decisions, the results can lead to unexpected outcomes.
Case Study: Darknet Shopper
The attacks of the hide and seek example are a metaphor for the real world. What if reinforcement learning were used to train an autonomous agent to advance the general goals of the user. Say for example, the user asked it to “find valuable investment opportunities”, “make my family healthier”, “improve my happiness”. To fund the agent, the human would give the agent a certain amount of Bitcoin that it could use to accomplish the goals. The agent could then get things on Amazon to accomplish the stated goal. Over time though, because this agent has full autonomy and learns dynamically, it discovers that it can better accomplish its goals by purchasing items on the dark web. (See for example this art project dark web). On the dark web it can buy food, fake shoes, drugs, and even weapons. Is the user of the agent responsible for everything that it does, even though they took no-part in training it? Is the operator knowingly using the bot to avoid civil or criminal liability by intentionally keeping themselves unaware that would otherwise implicate them in a crime. In the case of United States v. Jewell, it was held that willful blindness is the same as positive knowledge.
When things go badly, and people become angry, legal arrangements come into play to try and resolve the matter. There are different types of laws that we have to resolve these issues.
Types of Law
American law has developed economic, social, immigration, labor, and divisions. This is a tall order to fill. There are different intellectual traditions about law: classical, progressive, process, realist. These intellectual traditions have shaped the different bodies of law that we have in the United States, namely public law, private law, and international law.
There are three main types of public law. Public law is about how the federal or state government interacts with private individuals.
Constitutional Law defines defines and limits the power of the government. It sets for a structure of the governed, and the different branches: the executive, legislative, and judiciary. It is the supreme law. If another law conflicts with the Constitution, it is considered void.
Administrative Law are created by administrative agencies such as the FCC, SEC, FDA, and many, many others. Administrative agencies are created by Congress and run by the executive. Administrative agencies are very powerful. In a year, Congress, typically thought of as the main law maker, passes about 300 laws a year. Administrative agency pass about 10,000 laws a year. Congress likes to give administrative agencies power because they have domain expertise in crafting laws. Additionally, if a law is not liked by people, they can blame the agency as opposed to taking responsibility personally.
Criminal Law involves cases where the prosecutor is the state, and the defendant is a person. With criminal law, crimes are offenses against the state. So if we go back to the example of murder, murder is considered an offense not just against one person, but an offense to everyone in society. As such, the state, which represents everyone in society, is the prosecutor of the crime. The differences in the burden of proof are a result of American’s being historically wary of a government that is too powerful and that “it is better that 100 guilty persons should escape than that one innocent Person should suffer”.
The same conduct can result in both civil and criminal liabilities. An act does not have to fall under just area of law. For example, if someone is murdered, the state will purse murder charges. In addition, the family can pursue civil charges against the accused murderer. Different bodies of law also have different standards by which someone is convicted.
As opposed to Public Law, Private Law is for wrongs that are committed against private individuals by other private individuals. One person’s freedom ends when they harm another person. There are three main branches of private law.
Property law is the area of law establishes ownership over immovable (land, house,
property rights) and immovable (personal property, intellectual property, patent, copyright).
Contract Law is the establishment of a legally binding agreement between individuals. An agreement typically involves the exchanges of goods, services, or money. If the contract is broken, the injured party is entitled to damages as specified in the contract or other legal remedies as provided by civil law.
Tort Law is like the catch-all of civil law. With contract law, the contractual obligations are chosen by the parties, but in tort law, the obligation is imposed by the state. Claims under Tort law are typically not “planned” for like in contract law. For example, if a person’s dog bites another person, there was no contract in place to deal with this event, and it is not criminal, so it falls under Tort Law.
Lastly, we have International Law which has humanitarian law, for example when nations are at war, they restrict hurting civilians. Then there is also international criminal law whereby people can be convicted of violating these laws. Such as aiding and abetting and command responsibility.
Burden Of Proof
In a criminal trial, the government appointed prosecutor bears the burden of proof to present to a jury that there could be “no reasonable doubt” in a “reasonable person” that they defendant is guilty. Some have classified it as someone being 95% certain about the verdict Criminal cases have the highest standard of proof in the United States. Different burdens of proof are apparent in the number of jurors needed to come to a verdict. In the federal courts and in the state, all jurors are required to come to the same verdict, whereas in civil trials, it varies by state and is lower.
In comparison, civil cases have a lower standard of proof, such as a “preponderance of evidence”. Put another way, it means that it was more likely than not that an event happened in one way in comparison to another way. So if one party is 51% more guilty than they win the trial. The difference in standards are from the legal duties and responsibilities we have to one another.
There is a third burden of proof that is called clear and convincing, that is used in some patent cases. Here, the person should be around 75% sure that the verdict is correct.
Damages
In civil trials, if the defendant is found guilty, they are required to pay some amount of money to the person that is suing them. THere are two types of damages that can be awarded. The first is compensatory, which are “to make the plaintiff whole”. How much money does the plaintiff need to put them back in a position before they were harmed. These can be expenses for lost wages, damages to property, medical bills, and pain and suffering.
The other type of damages are punitive damages which are meant to punish the defendant and deter future action. The theory is that the defendant is not going to held liable for 100% of the time so there needs to be an additional disincentive to not be negligent in the first place.
For criminal cases, the person usually has to pay with prison time. Additionally criminal cases can order the guilty party to pay restitution to the injured party. Usually a civil judgment is decreased by the amount of restitution that the injured party has already received.
Timeline of a Civil and Criminal Trial
For a criminal case there are several steps that need to be followed. In practice, only a subset of them are followed because most of the time, the parties agree to settle before all steps are completed. First we’ll go through a criminal case, then a civil case.
First there is a grand jury is assembled. A grand jury is needed to indict someone by demonstrating to a jury that there is enough evidence to have a trial. The defendant here has not representation, and the saying goes that most prosecutors “can indict a ham sandwich”. Secondly, there is a process known as discovery. Here the prosecutor is required to give all evidence they have to the defendant. Next there are pre-trial motions, whereby lawyers from each side shape the course of the trial by advocating to the judge which evidence and arguments should be allowed in court. Additionally, the defense will try and get the case dismissed because of lack of evidence. Then there is jury selection. The attorneys from each side and the judge find acceptable jures. Each side tries to remove people that have a bias toward their client. Each is allowed to remove a certain number of people from the pool for cause (they are based) and without cause, without giving a reason. The main part is then the
trial. Each side makes the argument for there case. Next is the verdict. The jury comes to a verdict based on a certain standard of proof. If the jury does not reach a verdict, it is a called a hung jury, and the trial is considered a mis-trial. The prosecution can then either drop the case or try the case again Lastly, sentencing. If the defendant is guilty, the judge then decided the punishment based on minimum and maximum sentencing guidelines for certain crimes.
In a civil trial the steps are reduced to discovery, pre-trial motion, jury selection, trial, and verdict.
Plea Bargain And Settlements
The court system would be overrun with cases if everytime someone sued someone else, there was a court trial case. Additionally the cost of having a trial for the plaintiff and defendant can be very high in both legal and emotional costs. So the court system puts emphasis on trying to reach deals before the suits get to trial. In criminal cases, deals are called plea bargains and in civil cases they are called settlements. There is always uncertainty in how a case will turn out, so rather than risk it, people sometimes like to have certainty. In criminal cases, the prosecution will offer the defense a deal where they plead guilty to a lesser crime and not go to trial, sometimes also in exchange for testimony that could help convict a more important person. Plea bargain need to be approved by the judge. In a civil case, one side will pay the other side a certain amount of money in exchange for them dropping the case.
Legal Adaptation
“For the reformer has enemies in all those who profit by the old order, and only lukewarm defenders in all those who would profit from the new order; this lukewarmness arising partly from fear of their adversaries, who have the laws in their favour; and partly from the incredulity of mankind…” — Niccolo Machiavelli
Law is slow to change. It was not until July 2003 when Delaware stopped offering the option of being hung to death row inmates. In fact, it was not the law that changed, but that Delaware had dismantled its gallows. It was not until February 7th, 2013 that Mississippi officially ratified the 13th amendment to the Constitution… a mere 148 years later due to a clerical issue.
We could think that the legal system is designed to help America through society and economic change, but we have to recognize the less laudable factors. Factors like the initial distribution of interests and power, friction with existing arrangements, institutional inertia, and path dependence prevent the legal system from moving as fast as it might need to to adapt to the constant change. The best legislation and regulation occur when it is timeless in a sense, a framework that molds and adapts itself to the time it’s in… for example the Constitution. Of course, much is left unanswered in the Constitution and it has many shortcomings, but it has built itself to be a living structure.
Unfortunately we are trying to create perfect regulation for an imperfect system. The improved processes of political accountability and increased civic awareness seem like they will change “business as usual” in the legal system, but unfortunately the imperfection in regulating the real world still exists. There are a variety of pressures, variables, and interests that cannot be feasibly modeled and accounted for because they are so interconnected, often act not in their best interests, and they are rapid to change. Thus, we ask a lot from the legal system and should not be totally surprised when it doesn’t live up to what we hope it does.
How Law Changes?
So when the legal system does not live up to what we expect, how do we change it? We rely on a framework that we touched in the last notes, which differentiates how we govern, economic & social behavior, & political process.
In all cases of legal change, all three parts of the triangle have to somehow be affected in order to catalyze legal change. For example, there are certain stimuli that beget rapid legal change such as geopolitics, political upheaval, economic crisis, social/cultural change, and finally technological change.
In all of these kinds of catalysts, our economic & social behavior changes at the grassroots level, which gets those in the legislative and executive branch listening in fear of losing votes, which then informs the mechanism that we employ to respond to that change.
We can see legal change happening from our Constitution or statutes, but oftentimes that takes much longer to change and adapt. In fact, the Equal Rights Amendment which codifies into law the equal proection from discrimination for all people regardless of sex has been in the ratification process since 1923 and still remains unratified by 13 states. Delaware did not adapt their state constitution to offer equal rights to all regardless of sex till 2019. Largely considered a symbolic gesture, it was an inordinate amount of political capital to affirm even just something that was symbolic.
We see it much easier to see regulation change via common law, the decisions of federal and state agencies, and social norms. When in America, the so-called “Muslim ban” went out in the form of an executive order, it was the courts that initially deemed it unconstitutional before being upheld by the Supreme Court. It was the Department of Homeland Security that implemented the executive order and could help enforce it or just let it wither away without impacting the day to day travel plans of thousands of people. America experienced just how fast the law could move and sometimes the dangers of law moving so quickly. That within the first month of having a new Commander in Chief, America experienced rapid implementation of regulation that quickly took place and received backlash. The backlash is the social/cultural change that defines norms in this country that have an equally powerful role in shaping how the legal system works. Millions of Americans protested at airports, effectively shutting down an entire industry across the country in its busiest hubs.
Of course, this rapid evolution of the legal system was met with controversy and perhaps points to the reason law can be slow to change. We as a society have difficulty understanding whether particular signals from the markets or society actually describe the desires of where the group wants to go.
Imagine taking a group of friends to pick a restaurant. Each person might want something different and their “moods” or tastes for a specific kind of cuisine are relatively arbitrary and oftentimes lack rationale, but simply they are convicted that they are “craving Chinese food”. We must balance out the actual dietary needs of the group (i.e the vegetarians, the various allergies) with the capricious wants of the group. How reactionary should we be? Is taking one person’s reaction at one particular instance fair to generalize? If Dave wants Chinese food right now, is that what he will want everytime we go out for dinner?
Perhaps the pessimistic view would suggest as Machiavelli does that legal systems are slow to change in order to protect the power of the “incumbents,” those who have concentrated massive amounts of power and influence disproportionate to the rest of the population. Perhaps, however, the legal system is slow to change because of lengthy feedback loops (it could take years to start seeing how the laws/regulation system are affecting a statistically significant group) and because of the process in which laws are changed.
The best things in life take time and in order to have the Constitution, we needed the Articles of Confederation, decades of social upheaval, an entire revolutionary war, and so much more in order to be able to build a legal infrastructure that would work.
Professor Céllular recounted his experience working with the various US Federal Agencies at his time in the White House, focusing on the role of agencies and their unique position to enact change, but the eventual struggle they face. Federal agencies were created to bring marry the science and technical knowledge with the political process. The agencies act as a way to more directly influence the execution of legislation and focus on a limited scope of tasks rather than the operations of the entire country as the President does. If one were to look for experts in the government, they would find them in the agencies. These are the ecologists who determine the impacts of air pollution on native species who reside in the EPA or the biochemist who protects the health of the American patient at the FDA and unfortunately they have limited resources to enact the change they are often expected to.
The agencies were created to not just combine the technical needs with the political process, but to also listen to the people directly impacted by legislation and create feedback loops so the President can be informed on how to better run the country.
The federal agencies should like a good idea and they can be, but oftentimes there is a difference between what agencies should and want to do and what agencies can do.
Professor Céllular mentioned how easy it might be to write talking points of the agency’s work, but in practice they have lots of conflict with the political process in which concessions must be made to appease all the stakeholders and interests.
Case Study: Goldberg v Kelly
The enormous task facing government agencies is most clear when looking at a court case like Goldberg v Kelly. Although this case doesn’t focus on US federal agencies, state agencies operate under a similar scope with similar roles and expectations. The defendant John Kelly was receiving social welfare through the federally-assisted state program under New York’s home social relief program and one day suddenly stopped receiving them. He was infuriated that the government could just take away his aid without due process. He challenged that he was constitutionally afforded a procedure for notice and termination of such kind of aid. The agency protested saying that this additional requirement would overburden their already stringent resources. This due process can get expensive and although it would be great to have the ability to hear each appeal of social welfare aid put in from a citizen, they were not equipped to handle such a myriad of requests. They argued that the more money that is used in overhead costs meant that the less money could be disbursed, creating a vicious cycle. Yet, the courts ruled for them to provide every appeal the due process it requested. This put an enormous burden on the agency. Property was not what you owned, but also what you were owed. These expectations put on the agency were incredibly impossible to live up to and thus results in the massive backlog of requests and inefficiencies we currently experience in the social welfare system. By forcing agencies to better seal up the cracks in the system by giving each case due process, it also unknowingly fractured the capabilities of the agency to perform its vital functions as well.
Surprisingly, the very act of reforming an agency for the better necessitated the need for new regulation to fix the problems created by the initial regulation. People are frustrated with the system, but by fixing those problems we are left in a cycle of always moving the goal posts further and further away for those agencies. Just as we can’t expect agencies to be the golden bullet to implement and enforce the technical and logistics of regulation, we cannot expect artificial intelligence to eliminate the problems we see, but perhaps assuage them and create new ones simultaneously.
