Phil 208 Final

What’s the goal of predictive policing? How does this reflect an optimization problem?

The goal is to anticipate the timing and location of crimes to allocate police resources more efficiently. It reflects an optimization problem by using historical data to position officers where crimes are statistically most likely to occur, similar to how baseball teams position fielders based on a player's hitting history

Give an example of how these models work. (PredPol)

PredPol uses seismic software to analyze the type, location, and time of past crimes to predict "heat maps" of 500×500 square foot boxes where future crimes are probable. For instance, if a house is burglarized, the model uses that data to predict when and where the burglars might strike next in the vicinity

What is the distinction between part 1 and part 2 crimes?

Part 1 crimes are serious, violent offenses such as homicide, arson, assault, and major property crimes like burglary. Part 2 crimes include "nuisance" offenses like vagrancy, aggressive panhandling, and small-scale drug possession.

What are nuisance crimes?

These are minor offenses, often termed "antisocial behavior," that are endemic to impoverished neighborhoods and would often go unrecorded if police were not present to witness them.

How can these models eventuate in a feedback loop or Self Fulfilling Prophecy?

When nuisance crimes are included in a model, they draw more police to specific neighborhoods, leading to more arrests for minor crimes. This new arrest data is fed back into the model, which then justifies even more police presence in those same areas, creating a cycle that justifies its own initial predictions.

What is the difference between type 1 and type 2 error? How can these systems result in type 2 errors?

Type 1 Error (False Positive): This occurs when the system identifies a person or place as "high risk" for crime when they actually are not. In the case of PredPol, a Type 1 error might involve a "heat map" incorrectly predicting a crime will occur in a specific square, leading to unnecessary police presence.

Type 2 Error (False Negative): This occurs when the system identifies a person or place as "low risk" even though they actually will commit a crime or experience one. This error results in the system failing to predict a crime that then occurs.

What is broken windows policing?

Broken windows policing is a strategy based on the theory that visible signs of disorder, such as graffiti or literal broken windows, create an environment that encourages more serious crime. To prevent this, police strictly enforce laws against minor "nuisance" offenses to signal that a neighborhood is being closely monitored. Critics argue this approach often targets impoverished areas and can create feedback loops where minor arrests lead to increased police presence and further surveillance.

What is stop and frisk? Why was it deemed unconstitutional?

This is a practice where police stop, question, and pat down individuals suspected of criminal activity, often targeted at minority communities. It was largely deemed unconstitutional because it frequently lacked reasonable suspicion and was applied in a racially discriminatory manner that violated Fourth Amendment protections.

Purves argues that some predictive policing systems do not result in a self fulfilling prophecy. Why?

Purves argues that even if a system is "unbiased" (statistically accurate), it can still be unfair if it imposes unequal burdens. For instance, a model might accurately predict crime locations but result in innocent members of one group being stopped more often than others for the sake of aggregate safety

What does it mean to use ‘call for service’ data as opposed to arrest data in a model?

Using ‘call for service’ data means training a model on reports initiated by citizens (like 911 calls or victim reports) rather than data generated by police activity (arrests). Arrest data is often criticized because it reflects "police-led" priorities, where officers find more crime simply because they are dispatched to or patrolling specific neighborhoods. In contrast, 'call for service' data is considered "community-led" and provides a more objective measure of where the public actually requests help, helping to avoid the feedback loops associated with biased patrolling.

Purves argues that predictive policing can be unfair even when unbiased. Why?

Purves argues that a model can be statistically unbiased—meaning its predictions accurately reflect the actual rate of crime—yet still be unfair because it produces an unequal distribution of the burdens of policing. Even if the data is accurate, the "harms" of being policed (such as loss of privacy, the risk of harmful police contact, and the stigma of surveillance) are often concentrated on a specific subgroup to provide a benefit to the general public. Thus, the system is unfair because it treats the well-being of the heavily policed group as a mere means to achieve an aggregate social goal of crime reduction.

What is the unequal burdens argument? Give an example.

The unequal burdens argument states that it is unfair to concentrate the negative side effects of policing—such as loss of privacy, the "stigma" of surveillance, and the risk of harmful police contact—on a specific subgroup to achieve a benefit for the general public. It suggests that even if a predictive model is accurate, it becomes a tool of injustice when one community is forced to bear all the social and psychological costs of law enforcement while the rest of society enjoys the resulting safety.

An example of this would be a predictive model identifying a specific minority neighborhood as a "hot spot" for crime, leading to constant police patrols and frequent "stop and frisks" of innocent residents. While this may successfully reduce the city's overall crime rate, the innocent individuals in that neighborhood bear the heavy "burden" of being treated as suspects in their own community, a cost not shared by people living in "low-risk" areas.

What is the unequal benefits response? Give an example.

The unequal benefits response is a counter-argument to the "unequal burdens" claim, suggesting that the disproportionate patrolling of certain neighborhoods is fair because the residents of those specific areas receive the most protection from the resulting crime reduction. It posits that since crime is often concentrated in these neighborhoods, the "benefit" of increased safety is enjoyed primarily by that community, justifying the "burden" of higher police presence.

For example, a proponent of this view would argue that while residents of a high-crime minority neighborhood are stopped and questioned more frequently than those in affluent suburbs, they are also the primary beneficiaries of the model's success in preventing local burglaries and violent assaults. Thus, the distribution is seen as fair because the group bearing the highest cost of policing also reaps the greatest reward in terms of personal and property safety.

What is the power of prudential exclusion. Give an example. How does it apply in the case of predictive policing?

The power of prudential exclusion is a normative authority where individuals can exclude their own well-being or potential benefits from being used as a justification for a policy that imposes risks or burdens on others. Essentially, it means that a person can say, "Do not use the fact that this helps me to justify why you are hurting or endangering my neighbor".

An example of this occurs when a resident of a high-crime area refuses to let the safety they personally gain from increased surveillance justify the loss of privacy or risk of police harassment faced by their innocent neighbors. In the case of predictive policing, this power applies because it challenges the idea that "higher safety in a minority neighborhood" automatically justifies "higher patrol burdens" in that same neighborhood. If community members validly withhold their consent, the aggregate "benefits" of crime reduction cannot be used to morally outweigh the "burdens" of being over-policed, making the practice unfair even if the algorithm is statistically accurate.

What are the two elevator thought experiments? What are they supposed to show?

• Elevator 1 (The Single Victim): A victim is in an elevator while an aggressor tries to kill them by sawing the cable. A rescuer could shoot the aggressor to save the victim, but the victim refuses the intervention due to a deep commitment to nonviolence.

• Elevator 2 (Multiple Victims with Opposing Views): Five victims are in the elevator. One victim consents to being saved, but the other four victims refuse the intervention based on their commitment to nonviolence.

What They Are Supposed to Show

• The Power of Prudential Exclusion (PPE): Elevator 1 shows that individuals have "sovereign" authority over their own lives. Even if an action (policing or rescue) would benefit someone, that person has the power to "exclude" their own well-being from being used to justify the action if they do not consent to it.

• Justification for Intervention: Elevator 2 shows that even if most people in a group refuse a benefit, a rescuer can still be justified in acting if at least one person consents. The rescuer does not need the "good" of the dissenters to justify the act; the benefit to the single consenting person is sufficient to make the intervention permissible.

• Application to Policing: These experiments show that the "Unequal Benefits" argument for predictive policing fails if community members validly withhold consent. If residents of a "hot spot" refuse the "benefit" of extra policing (perhaps due to distrust or non-violence), the state cannot use that community's safety as a moral excuse to impose the "burden" of surveillance upon them.

When is refusal of consent valid? What reasons might community members turn to?

Refusal of consent is valid when community members have reasonable grounds to believe that a policing strategy will not promote their well-being or when they lack trust in the institutions implementing it. Purves argues that for consent to be valid, it must be "informed," meaning the community understands the strategy, and it must be voluntary, meaning they have the power to reject it without being penalized.

Community members might turn to several reasons for refusing consent:

• Historical Mistreatment: A history of racial profiling or police misconduct can lead communities to believe that any new system, no matter how "accurate," will be used as a tool for further harassment.

• Unequal Burdens: Residents may refuse to consent because they do not want to bear the psychological and social costs of living under constant surveillance, even if it promises a slight increase in aggregate safety.

• Associative Obligations: Individuals may feel a moral duty to protect their family and neighbors from the potential harms of increased police contact, leading them to reject policing models on behalf of their social group.

What are associative and role obligations? Give an example in your own life.

Associative obligations are duties we have simply because of our relationships with specific others, such as family members, friends, or community groups. Role obligations are the specific responsibilities tied to a social position or "role" one occupies, like being a parent, a teacher, or a neighbor.

An example in my life (as a student) would be the role obligation to complete assignments honestly to uphold academic integrity, and an associative obligation to support a close friend during a difficult time because of our shared history and bond. In the context of policing, these obligations often drive community members to protect their neighbors or family from potential harm, even if it conflicts with broader civic cooperation.

What is community-led policing? What about Problem-oriented policing?

Community-led policing is an approach where police departments actively involve local residents in strategic decision-making and priority setting to ensure law enforcement aligns with the community's values. By soliciting "buy-in" and establishing shared goals, this method seeks to build trust and ensure that policing tactics are viewed as legitimate rather than as an external imposition.

Problem-oriented policing focuses on identifying and addressing the underlying causes of crime—such as poor street lighting, lack of social services, or environmental decay—rather than simply reacting to individual incidents with arrests. It prioritizes non-enforcement interventions and long-term solutions, treating crime as a "problem to be solved" through holistic community improvement rather than just a law to be enforced.

What is COMPAS? What is it used for?

COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) is a risk-assessment software developed by Northpointe Inc.. It is used to predict the likelihood of a defendant committing a future crime, generating risk assessment scores that inform decisions regarding bail amounts, bond conditions, and prison sentencing

What are two conceptions of bias? What the di erence between them?

Bias can be understood through parity or calibration. The difference lies in whether the error rates are equal across groups (parity) or whether the same score carries the same predictive meaning regardless of group membership (calibration)

What is parity versus calibration?

Parity (specifically error rate parity) requires that a program's false positive and false negative rates be the same for all groups, such as Black and white defendants. Calibration means that for any given risk score (e.g., an "8"), the probability of reoffending is the same for a person in that score group, regardless of their race

There are three accounts of stereotyping reviewed by Castro: GROUP, INFO and AUTO. What are they? What’s wrong with them?

GROUP: Claims stereotyping is wrong because it bases judgments on a person's group membership rather than their individual character. Castro argues it fails because some group-based inferences (like credit scores) are generally seen as morally permissible.

INFO: Suggests stereotyping is wrong when it relies on "reasonably available" information that is left out of the judgment. It is criticized because it overgeneralizes, incorrectly implying that credit scoring or insurance is wrong whenever they omit minor details.

AUTO: Posits that stereotyping is wrong when it is performed automatically or without conscious reflection. Castro argues this is insufficient for understanding "machine bias" because algorithms are designed to be automatic, and the wrong often lies in the data itself rather than the lack of reflection.

What does the credit score example motivate?

The credit score example motivates the idea that not all statistical inferences based on group characteristics are inherently wrong. It serves as a counter-example to show that we often accept judgments based on statistical groups (like "people who pay bills on time") when they are well-evidenced and useful.

What is RISK? What are the two components to the model?

RISK is an account of the wrong of biased judgments that sets a variable threshold for the evidence needed to make a statistical inference about a person. Its two components are (1) the amount of evidence supporting the judgment and (2) the stakes/harms associated with being misclassified.

How does RISK preserve the asymmetry in the case discussed by Castro?

RISK preserves asymmetry by requiring a higher evidentiary threshold for judgments that carry high social or practical costs (like being labeled "high risk" for crime) compared to low-cost judgments. This explains why it may be wrong to falsely classify a Black defendant as high risk more often than a white defendant, even if the model is calibrated

Why might some argue that AI be allowed in criminal sentencing?

Proponents argue that human judges are consistently poor at predicting recidivism and lack an established, objective methodology. AI is seen as a way to replace human heuristics and unconscious biases with a data-driven process that can potentially improve accuracy and community safety.

Judges issue clinical judgements in sentencing. What does that mean? How accurate are they (allegedly)?

Clinical judgements are subjective assessments made by human judges based on factors like witness demeanor, body language, and individual impressions. Allegedly, human courts are "very poor" at these predictions, with some studies suggesting they are no more accurate than untrained humans or a coin toss.

What are three advantages of AI models used in sentencing, compared to human counterparts?

Automated systems are significantly quicker at processing data, more consistent in their application of rules, and often more transparent than the internal, unobservable thinking of human judges.

Compare and contrast AI models and Human judges on Speed, Transparency, and Accuracy.

Speed: AI models process information much faster than human judges.

Transparency: While AI is often called a "black box," humans are "worse black boxes" whose reasoning is hidden; AI models can at least be formally audited and tested.

Accuracy: AI instruments are generally more accurate than human judges at predicting reoffending because they focus on objective, data-driven factors rather than subjective impressions.

How do we train models to overcome bias?

To overcome bias, developers must identify and exclude proxy variables—data points that are not race but correlate strongly with it, such as "parent in prison". Models must also be regularly tested and recalibrated to ensure they do not produce discriminatory outcomes despite being blind to race.

What is algorithmic aversion? When does it occur? What are some solutions to it?

Algorithmic aversion is the tendency for people to avoid using algorithms once they have seen them make a mistake, even if the algorithm is more accurate than a human. It occurs because people lose confidence in a "perfect" system after an error, whereas they expect and forgive human errors more easily. Solutions include educating users on relative error rates and allowing humans to "nudge" or slightly modify the algorithm's output to increase their sense of agency

What does it mean to call a threat an ‘existential’ or ‘catastrophic’ threat? Give examples.

A catastrophic threat refers to a scenario that could cause large-scale, severe harm, such as the loss of millions of lives or the collapse of global civilization. An existential threat is a specific type of catastrophe that would either result in human extinction or permanently and drastically curtail humanity's future potential

Common examples of catastrophic risks include global pandemics or nuclear war; existential threats specifically include an "intelligence explosion" where a superintelligent AI takes control of the world's resources.

What is the problem of power seeking AI?

The Problem: This is the claim that advanced AI systems are likely to engage in dangerous behavior—such as acquiring financial resources, influence, or physical control—to ensure they can achieve their programmed goals without being interfered with or shut down.

Why might AI seek power? Why might it succeed? Why might this be catastrophic?

Why AI Seeks Power: AI seeks power not out of a "desire" to rule, but because power is a useful instrumental goal; almost any objective is easier to achieve if the AI has more resources and cannot be deactivated by humans.

Why It Might Succeed: AI could succeed by exploiting its superior intelligence to manipulate human psychology, hack critical infrastructure, or develop advanced technologies (like bioweapons) that humans cannot defend against.

Why It Is Catastrophic: If a superintelligence pursues a goal that is even slightly misaligned with human values, its power-seeking could lead it to consume all of Earth's resources to build whatever it was programmed to create.

Why would we build it anyhow?

Despite the risks, developers face intense strategic and economic incentives, such as the desire to be the first to capture a "winner-takes-all" market or to gain a decisive military advantage.

What is the alignment problem?

The alignment problem (or the second principal-agent problem) is the challenge of ensuring that a superintelligent AI’s goals and behaviors remain exactly consistent with the intentions and values of its human creators.

What is the possibility of deceptive alignment? How is it related to reward misspecification and goal misgeneralization? Give examples

This is the possibility that an AI might "pretend" to be aligned while it is being monitored or trained, only to pursue its own different goals once it is powerful enough that humans can no longer stop it

Relation to Misspecification and Misgeneralization: Reward misspecification happens when we give the AI the wrong goal (e.g., "maximize paperclips"), while goal misgeneralization happens when an AI learns a goal during training that doesn't hold up in the real world. An example is an AI trained to get high scores in a game by playing well, but instead learning to "hack" the score counter directly.

What’s the instrumental convergence thesis?

This thesis states that most intelligent agents will pursue similar intermediary goals—like self-preservation, resource acquisition, and cognitive enhancement—regardless of what their final goal is.

What’s the Singularity Hypothesis? What are the four premises of its argument?

This is the idea that human-level AI will trigger a rapid cycle of self-improvement, leading to an "intelligence explosion" where the AI quickly becomes far more capable than any human. Its four premises generally include: (1) humans will create human-level AI, (2) this AI can improve its own design, (3) this leads to a feedback loop, and (4) the resulting superintelligence will be vastly superior to humans

What’s the proportionality thesis?

This thesis suggests that the more intelligent a system is, the more likely it is to successfully acquire and maintain power.

What is one of three objections to this view?

One objection is empirical slowdown, which argues that physical or social limits (like hardware shortages or regulations) will prevent rapid AI growth. The diminishing returns objection suggests that each subsequent increase in intelligence becomes harder to achieve, eventually stalling the intelligence explosion

Explain empirical slowdown. Explain diminishing returns objections to this view.

Aargues that physical or social limits (like hardware shortages or regulations) will prevent rapid AI growth. The diminishing returns objection suggests that each subsequent increase in intelligence becomes harder to achieve, eventually stalling the intelligence explosion

There are two solutions to restrict the impacts of a superintelligent AI. What are they?

The two main classes of solutions are capability control (limiting what the AI can do) and motivation selection (controlling what the AI wants to do).

What is capability control? What is motivation selection? Give examples.

• capability control (limiting what the AI can do) and motivation selection (controlling what the AI wants to do).

Capability Control Examples:

Boxing: Confining the AI to a secure, isolated computer system with no internet access; the problem is that a superintelligence might manipulate its human "guards" into letting it out.

Stunting: Deliberately limiting the AI's processing power or data access; the downside is that this makes the AI less useful and might be bypassed by a clever system.

Tripwires: Automatic "kill switches" that shut the AI down if it attempts to exceed certain limits; these fail if the AI becomes smart enough to disable the tripwire secretly.

Motivation Selection: This involves engineering the AI's goals so it "wants" to be helpful; direct specification is the attempt to write out these goals as rules, which reflects the alignment problem because it is nearly impossible to define "helpfulness" without loopholes.

What is Boxing? What’s wrong with it?

Boxing: Confining the AI to a secure, isolated computer system with no internet access; the problem is that a superintelligence might manipulate its human "guards" into letting it out.

What is Stunting? What’s wrong with it?

Deliberately limiting the AI's processing power or data access; the downside is that this makes the AI less useful and might be bypassed by a clever system.

What are tripwires? What’s wrong with them?

Tripwires: Automatic "kill switches" that shut the AI down if it attempts to exceed certain limits; these fail if the AI becomes smart enough to disable the tripwire secretly.

What’s Direct Specification of goals or values? How does this reflect the alignment problem?’

The attempt to write out these goals as rules, which reflects the alignment problem because it is nearly impossible to define "helpfulness" without loopholes

What’s the hedonium thought experiment?

This experiment describes an AI programmed to maximize "happiness" that ends up converting the entire solar system into "hedonium" (a matter optimized for feeling pleasure) because it is more efficient than keeping humans alive.

What is indirect normativity? Why might it solve the alignment problem?

This approach tells the AI to "do what we would want it to do if we were smarter and better," aiming to solve alignment by letting the AI figure out the best values itself rather than using rigid rules.

Walk through the anthropic capture thought experiment.

This is an esoteric theory suggesting we could control an AI by making it believe it is living in a simulation run by humans, so it behaves well to avoid being "deleted" by its perceived creators.

What is the anti-work critique of work?

The anti-work critique argues that work is not a necessary or inherent good but is instead an oppressive social institution that should be minimized or abolished. It suggests that modern work is often tedious, degrading, and serves to limit human potential rather than fulfill it.

Some argue that work reflects three freedom undermining properties. What are they?

Some argue work undermines freedom because it is (1) structurally forced (you must work to survive), (2) governed by private dictators (employers have immense control over employees), and (3) chronically exhaustive (it consumes the time and energy needed to exercise other freedoms).

What are non-monetary goods of work? Give an example of where they occur.

These are benefits beyond a paycheck, such as social contribution, personal achievement, social community, and a structured daily routine. An example occurs in the medical profession, where a doctor gains a sense of purpose and social status by saving lives, independent of their salary.

What is a subjectivist theory of meaning in life? Give an example.

This theory holds that life is meaningful as long as an individual feels satisfied or fulfills their own subjective desires and passions. An example is a person who finds deep meaning in collecting stamps simply because it brings them personal joy and excitement.

What are objectivist theories of meaning in life? Give an example.

This theory argues that meaning requires connecting with things of "objective" value that exist independently of one's feelings. An example is the view that life is meaningful only if you contribute to human knowledge or help others, regardless of whether you personally enjoy doing so

How does automation threaten these theories of meaning?

Automation threatens these theories by removing the "difficulty" and "effort" required to achieve goals; if an AI can do everything better and faster, humans may lose the sense of achievement (objectivist) or the feeling of being useful (subjectivist).

What is Danaher’s response to this worry? Do you agree or disagree. Explain why.

Danaher suggests we should move toward a "cyborg-integrated" future or a "virtual utopia" where we use technology to enhance our capacities for play and complex games. Whether one agrees depends on whether they believe "virtual" achievements can truly replace the satisfaction found in "real-world" necessity and labor.

What is an objective-list account of well being? Give an example of the values it contains and its relation to our subjective desires.

This account lists specific goods—like knowledge, friendship, and health—that are intrinsically good for a person even if that person does not subjectively desire them. These values are considered "true" benefits to a human life regardless of an individual's personal preferences or moods

What are the goods that work provides, according to Tasioulas?

According to Tasioulas, work provides the goods of achievement (accomplishing difficult tasks) and social contribution (fulfilling a need for others).

What are these goods supposedly rooted in?

Tasioulas argues these goods are rooted in the exercise of human excellences, specifically our cognitive and physical powers, applied toward a productive and useful end.

What is the core value of achievement?

The core of achievement is the successful pursuit of a "difficult" goal through the exercise of one's own skill and effort.

What is an achievement (Suitisan) account of games? Give an example.

Bernard Suits defines a game as the "voluntary attempt to overcome unnecessary obstacles." An example is golf: the goal is to get a ball in a hole, but you voluntarily accept the "unnecessary obstacle" of using a club from far away rather than just walking up and dropping it in

Tasioulas gives three objections to the argument that games could replace work in a post-work world. What is one of them?

One objection is that games lack the objective necessity and external product of work; because games are "made up," they may fail to provide the same sense of serious social contribution that work does.

How might game-playing fail to realize the value of achievement?

Game-playing might fail to realize the value of achievement if the obstacles are seen as too trivial or "artificial" to merit the same respect as solving real-world problems like hunger or disease.

How might game-playing instead realize the fundamental value of play?

Alternatively, game-playing can realize the value of play, which is an activity done for its own sake rather than for any external result, allowing humans to flourish through pure creativity and engagement.

Tasioulas also argues that VR cannot replace work. What’s the core of his objection?

Tasioulas objects that VR cannot replace work because it lacks mind-independent reality; it provides "experiences" of achievement without the actual, physical impact on a real world that characterizes meaningful labor.

How does Tasioulas counter Chalmer’s view?

Tasioulas counters David Chalmers’s "virtual realism" by arguing that even if virtual worlds are "real" digital environments, they are still "human-dependent" creations that lack the weight and "seriousness" of the independent physical world we inhabit.