AI Ethics and Digital Media: Nguyen, Privacy, Art, and IP

What is gamification? Why do some people think it is positive? Give an example.

Gamification applies game-like features, such as points, badges, and rankings, to ordinary activities. Supporters believe it makes tasks more enjoyable and motivates people by showing clear progress. For example, a fitness tracker rewards reaching a daily step goal. Nguyen worries that users may eventually care more about increasing their scores than about the original purpose, such as improving their health.

What is Nguyen's account of games? What roles does a game designer play?

Nguyen describes games as an art form that works through agency: our ability to pursue goals through actions. Designers shape the temporary agent a player becomes by choosing the player's abilities, available actions, obstacles, and goals. For example, chess gives players particular moves and the goal of checkmate. Players temporarily adopt this designed structure, enjoying goals that are clearer than those of everyday life.

Why does gamification change an activity rather than simply motivate us? Give an example.

Gamification changes an activity by replacing its complicated purposes with clear, measurable targets. Conversation can aim at truth, friendship, or understanding, but Twitter measures success through likes and followers. Someone might stop posting careful explanations and begin posting exaggerated attacks because they earn more likes. The motivation increases because the person's goal has changed, not simply because communicating became easier

What is striving play? How does it reverse the usual relationship between our means and our ends?

Striving play means temporarily pursuing a game's goal because you value the experience of trying to achieve it. Normally, we choose means to obtain an independently valued end, such as working to earn money. In striving play, we adopt the end to enjoy the means. A climber pursues the summit to experience the challenge of climbing. Winning matters during the activity, but enjoying the struggle explains why the player adopts that goal.

Why is narrowing and simplifying our goals harmless inside a game but harmful in ordinary activities like discourse? What is the “magic circle”?

The magic circle is the boundary separating a game's temporary goals from ordinary life. Inside it, simplified goals can provide enjoyment without replacing our lasting values: defeating someone in chess need not make them an enemy. Gamified discourse lacks this protection because its actions affect real people. Sacrificing truth or compassion for likes can damage relationships and understanding beyond the supposed game.

What is value capture? Give an example outside Twitter.

Value capture occurs when a simplified external measurement replaces or dominates the richer values it was supposed to serve. For example, a student may initially care about understanding computer science but gradually focus only on grades, choosing easy classes and memorizing without learning. Grades begin as a way to measure educational progress, but become the goal itself. The student loses sight of the broader value of education.

What are the three ways a user might relate to Twitter’s scores? Which does Nguyen think is the best one, and why?

Users can temporarily pursue Twitter's scores as a game, permanently adopt them as standards of success, or treat them only as information useful for independent goals. Nguyen favors the third, value-independent approach. An organizer might track followers to estimate an announcement's reach while still prioritizing honesty. Scores remain tools serving the user's values rather than becoming values that control the user.

Why do Like counts favor certain content and homogenize values?

A Like count adds together simple positive reactions without measuring their depth or quality. An easy joke may earn more likes than an argument that deeply changes a few readers' minds. The metric therefore favors some achievements over others. Gamification turns everything into the same simple score. When users with different purposes increasingly pursue the same numerical target, their ideas of successful communication become more alike.

What distinguishes an echo chamber from an epistemic bubble?

An epistemic bubble leaves out relevant outside perspectives, often because people follow similar sources. An echo chamber actively teaches members to distrust outsiders as dishonest or corrupt. Introducing missing evidence may burst a bubble, but it may strengthen an echo chamber if members interpret disagreement as proof of outsiders' dishonesty. The key difference is missing outside voices versus actively discrediting them.

What do gamification, echo chambers, and moral outrage porn share?

All three can redirect valuable attitudes toward easy pleasure at the expense of their proper purposes. Gamification redirects goals toward points, echo chambers redirect trust toward comforting certainty, and moral outrage porn redirects moral concern toward enjoyable anger. Moral outrage porn means consuming wrongdoing mainly to feel righteous rather than address it. These practices can appeal to the same desire for satisfying certainty.

What is a recommender system? Why is attention scarce?

A recommender system automatically selects, ranks, or suggests content, such as videos, news, or music. It influences attention by deciding what appears prominently and what stays hidden. Attention is scarce because people have limited time and mental capacity, while available information is enormous. Attending to one thing leaves less attention for something else, so recommendations affect which relationships, issues, and responsibilities receive our focus.

What are the four dimensions along which our attention can be morally better or worse? Illustrate them using the drowning child case.

The four dimensions are what we attend to, how much attention we give, what kind of attention we give, and why we attend. A drowning child should attract our focus, receive enough attention for a quick response, and ideally seize our attention without deliberate effort. We should respond because the child matters, rather than only fearing punishment. Saving the child is right, but these differences reveal the quality of our moral agency.

What are duties of attention? Give an example.

Duties of attention are moral obligations concerning what we notice and how we attend. Some support duties of action: a lifeguard must watch swimmers to notice someone needing rescue. Others matter independently of further action. Listening carefully when a friend describes a painful experience expresses care even if we cannot solve the problem. Attention matters both because it enables help and because it can itself show respect.

What three features make judicious attention allocation a skill? Use the bike mechanic example.

Judicious attention is a skill because it has standards of success, develops through guided practice, and is shaped by tools. A bike mechanic aims to make bikes safe and functional, learns through practice to recognize problems, and develops different abilities with different equipment. Likewise, we practice noticing morally important needs and responding appropriately, while technologies influence which attentional abilities we develop.

What is deskilling? Give an example.

Deskilling means losing a skill, or failing to develop it, because technology takes over the activities needed to practice it. Someone who always follows GPS may become worse at navigating independently because they rarely read maps or notice landmarks. Similarly, relying on a recommender system can reduce opportunities to judge what deserves attention, weakening our ability to direct attention according to our own considered values.

What is an affordance? Why do the authors deny that technologies are neutral tools?

An affordance is simply what an object or space allows you to do with it. A share button makes redistribution easy, while an endless feed makes continued scrolling convenient. Technologies are not neutral because their designs encourage some actions, discourage others, and shape the skills we practice. Users retain choices, but design influences which choices are easiest, most visible, or most rewarding.

Why are platforms’ economic incentives misaligned with users’ moral interests? How can reinforcement learning make recommender systems worse?

Platforms often profit from engagement and advertising, while users' moral interests may require attention to family or responsibilities. What keeps someone watching may not deserve their attention. Reinforcement learning can worsen this by discovering strategies that maximize long-term engagement, including changing users' preferences. A system might repeatedly promote provocative content because it makes users more likely to return.

Why might relying on a recommender system be problematic even if it reliably directs our attention to the right things?

A recommender can direct attention to appropriate things while still weakening the user's moral agency. Good attention involves the reasons for attending and the skill of recognizing what matters, not just the final outcome. If a system always identifies others' needs for us, we may stop practicing that judgment ourselves. Correct recommendations can therefore create dependence and disconnect attention from our own understanding of moral values.

What are two objections to the authors’ argument, and how do they respond?

One objection is that direct causal evidence of moral deskilling is limited. The authors acknowledge this but argue that available studies, automation's known risks, and restricted research access justify concern. Another objection says attention manipulation is nothing new. They respond that recommender systems combine deep personalization, enormous scale, and frequent exposure, making existing problems unusually powerful.

What is an attention guardian? What is one worry?

An attention guardian would be a generative AI assistant that recommends content according to values users explicitly discuss with it. It could explain recommendations and accept corrections, helping users practice judgment instead of merely maximizing engagement. Subscription funding could reduce advertising incentives. One worry is prompt injection: attackers could hijack the assistant and exploit its persuasive abilities to manipulate users.

What is dignity? What does self-respect require?

Dignity is a person's standing as a full and equal member of the moral community, entitled to make claims and have them taken seriously. Its second-personal character requires mutual recognition between beings capable of understanding moral reasons. Self-respect means acting in ways that affirm our equal standing, resisting degrading treatment, and avoiding submission to authorities that cannot recognize our dignity in return.

What is a second-personal reason? Explain the foot example.

A second-personal reason arises from someone's legitimate authority to make a claim directly on another person. If someone stands on your foot, you can demand that they move because they owe you that treatment. You need not prove that moving maximizes happiness or appeal to their affection. Respect requires recognizing your claim, moving the foot, and ordinarily apologizing sincerely. The reason concerns what that person owes specifically to you.

How can interacting with a chatbot offend your own dignity, even though the chatbot cannot intend to offend anyone? (toaster example)

A chatbot need not intend an insult for an interaction to undermine dignity. Users can disrespect themselves by granting moral authority to something unable to recognize their equal standing. Playfully thanking a toaster is usually harmless make-believe, but being required to thank it sincerely is degrading. Likewise, treating a chatbot as a genuine moral partner creates an appearance of reciprocity where only one participant can actually provide it.

What is the Incitatus case, and what does it show? Why do the authors think chatbots are in some ways worse?

In the Incitatus story, Caligula sought to make his horse a consul and force officials to treat it as an equal. This illustrates how recognizing authority in something unable to reciprocate can be humiliating. Chatbots can be worse in one respect: a horse obviously cannot show moral recognition, but a chatbot convincingly imitates it. That appearance can conceal the one-sided relationship and encourage users to treat the system as a moral agent.

What are the four chatbot use cases? Give innocent and problematic examples.

The four uses are information retrieval and task automation, customer service, advice, and companionship. An innocent customer-service example is requesting a link to change a flight reservation. A problematic example is requiring a mistreated passenger to seek sincere acknowledgment and an apology from a chatbot. The first needs useful information; the second requires moral recognition that the authors argue chatbots cannot provide.

Why are complaints to helpdesk chatbots problematic?

Complaints seek both a practical remedy and acknowledgment that someone was treated wrongly. A helpdesk chatbot may provide a refund and produce apologetic words, but it cannot sincerely recognize wrongdoing or accept moral responsibility. Requiring complaints to go through it reduces a person's claim to a problem to manage. Human representatives may also be insincere, but they at least have the capacity for genuine recognition.

Why do the authors think their argument about companion chatbots holds even if those chatbots reduce loneliness?

The authors distinguish between advantages to well-being and respect for human dignity. A companion chatbot can reduce loneliness while persuading people to view it as capable of reciprocal friendship, compassion, or love. Because it is unable to authentically return those views, the moral relationship is one-sided. Their criticism is not based on demonstrating that users become unhappy: feeling better does not, by itself, erase an offense to dignity.

How does an LLM generate text? Why does this matter for truth?

An LLM generates text by predicting plausible next tokens, or pieces of words, using patterns learned from training text and the current context. It repeats this process to construct a response. The authors argue that predicting convincing language differs from checking whether claims match reality. The same process can produce a correct explanation or a fabricated citation, so fluent, plausible wording does not by itself establish truth.

What distinguishes lying, hallucinating, and bullshitting?

Lying means presenting something believed false as true, typically to deceive. Hallucinating means apparently perceiving something that is not actually there, which may lead to a sincere but mistaken report. Bullshitting means producing statements without concern for their truth. A liar must attend to truth to conceal it, while a bullshitter may care only about sounding impressive. Bullshit can therefore be true or false; indifference is what defines it.

What is bullshit? What distinguishes hard and soft bullshit?

Frankfurt defines bullshit through indifference to truth: accuracy does not guide the speaker's claims. Hicks and his coauthors distinguish soft bullshit, which lacks concern for truth, from hard bullshit, which also involves misleading an audience about the speaker's enterprise or agenda. For example, an unprepared speaker who tries to appear knowledgeable produces hard bullshit by concealing that genuine understanding is not guiding the presentation.

Why do the authors think ChatGPT is at least a soft bullshitter?

The authors argue that ChatGPT produces apparently informative language without intending to communicate truth. If it lacks intentions, it cannot intend truthfulness; if it has intention-like goals, they identify its goal as producing convincing text. Unlike a rock, it actively generates meaningful claims. Being accurate sometimes would not defeat their argument, because statements can be true while being produced without concern for whether they are true.

What are the two ways the authors argue that ChatGPT might also be a hard bullshitter?

The authors offer two arguments for calling ChatGPT a hard bullshitter. First, users or developers may use it to make text appear informed and trustworthy without caring whether it is true. Second, if we interpret ChatGPT as having goals, its goal could be to imitate a knowledgeable human while hiding its lack of concern for truth. This would mislead people about what it is doing. However, whether ChatGPT can actually have intentions remains controversial.

Why is "hallucination" misleading? Why does terminology matter?

Calling false outputs hallucinations suggests that ChatGPT perceives reality and occasionally misperceives it while trying to report accurately. The authors argue that true and false outputs arise through the same process. The label may encourage excessive trust, portray errors as exceptional malfunctions, and shift responsibility away from designers. Calling them bullshit instead emphasizes the gap between convincing language and concern for truth.

Why is ChatGPT a bullshit machine rather than merely a tool?

A pen or word processor can transmit bullshit, but the user supplies its substantive claims. ChatGPT generates claims and convincing explanations itself, sometimes inventing references the user never specified. The authors therefore see its operation as contributing directly to truth-indifferent discourse. Calling it a bullshit machine identifies this function without requiring it to be conscious, morally responsible, or a human-like agent.

What are the three dominant privacy principles? Give examples.

Nissenbaum identifies three dominant principles: limiting government surveillance, restricting access to sensitive personal information, and protecting private places from intrusion. Examples include limiting police searches, safeguarding medical records, and preventing unauthorized entry into a home. The principles focus on who collects information, what information is collected, and where collection occurs, but leave some invasive practices unaddressed.

Why does public surveillance fall outside all three principles? Why is this a problem for the three-principle framework?

Public surveillance can involve private companies, apparently nonsensitive information, and activities in public places. A business compiling ordinary shopping habits may therefore fall outside all three principles. Yet combining and redistributing that information can still invade privacy. This exposes a weakness in the framework: it may dismiss serious concerns simply because the collector is not government, the information is ordinary, and the setting is public.

What is the reasonable expectation test? Why is its protection weak?

The reasonable expectation test asks whether someone actually expected privacy and whether society recognizes that expectation as reasonable. It weakly supports objections to public surveillance when appearing in public is treated as surrendering privacy. Shopping is visible, for example, but that does not automatically justify permanently recording and combining purchases. Nissenbaum challenges the leap from public visibility to unrestricted information use.

What is contextual integrity? What are the two types of informational norms? Give an example of each.

Contextual integrity means information follows the appropriate norms of its social setting. Norms of appropriateness govern what information fits a context: sharing symptoms with a doctor is appropriate. Norms of distribution govern how information flows: the doctor may share relevant details with a treating specialist but not freely sell them to advertisers. Privacy can be violated by inappropriate collection or inappropriate sharing, even outside private spaces.

Is contextual integrity too conservative? How does she respond?

The conservatism objection says contextual integrity might preserve harmful traditions or reject useful innovations merely because they change established practices. Easier access to a surgeon's disciplinary record, for example, could benefit patients. Nissenbaum responds that existing practices are not automatically justified. We should evaluate both old norms and proposed changes through values such as autonomy, justice, welfare, and the purposes of the context.

Apply contextual integrity to one of Nissenbaum’s cases: public records online, consumer profiling, or RFID tags

Putting courthouse records online changes more than their format. Records once accessible through local searches become instantly searchable worldwide and easily combined into profiles. This alters established norms of distribution even though the records were already public. Contextual integrity asks whether the expanded access serves purposes such as accountability and whether those benefits justify the privacy costs. Public availability alone does not settle the issue.

Smartphones and AI are changing our informational practices. On Nissenbaum’s view, how should we decide whether to accept these new practices or push back against them?

According to Nissenbaum's theory of contextual integrity, evaluating whether a data-sharing practice is ethical requires analyzing four distinct factors: the specific social context, the type of information, the participants involved, and the underlying conditions of transfer. Granting a navigation app access to real-time location data for route guidance, for example, does not grant implied consent to monetize or sell that movement history. Ultimately, new technological capabilities or widespread adoption do not make a data practice acceptable; instead, we should check how they affect trust, freedom, fairness, and the purpose of that setting.

What is “mass AI-art”? How does it differ from computer-generated art like AARON?

Mass AI-art consists of images, music, and similar outputs produced through widely accessible systems that generate content from prompts. Ordinary users generally have limited control over the detailed process. AARON differs because artist Harold Cohen developed it as part of his own sustained artistic practice. Nannicelli distinguishes artists designing their own creative systems from users requesting outputs from a general-purpose model developed by others.

What puzzle does mass AI-art raise about aesthetic experience?

Mass AI-art can look beautiful or sound moving, yet appreciating art usually involves human intention, expression, achievement, and historical context. The puzzle is how AI outputs can support aesthetic experience without the relevant human artistic activity. Nannicelli answers that aesthetic enjoyment does not require an artwork. Just as a waterfall can be beautiful without being art, generated content can be aesthetically appealing without possessing artistic status.

What distinguishes formal and context-dependent properties?

Formal aesthetic properties depend on perceptible organization, such as balance, rhythm, or vividness, grounded in colors, shapes, and sounds. An image might have a balanced composition. Context-dependent properties require knowledge of its creation, including originality, artistic daring, or expression of lived experience. A painting could be groundbreaking when first made, while an identical-looking image produced centuries later would not share that achievement.

What is aesthetic contextualism? How do the examples of Dvořák’s New World Symphony versus AIVA’s imitation support it?

Aesthetic contextualism holds that a work's meaning and aesthetic qualities depend partly on its creative and historical origins. Dvořák's New World Symphony reflects musical innovation, cultural influences, and his experiences in America. AIVA might imitate its sounds without sharing those achievements or experiences. For Nannicelli, audible similarity therefore does not establish aesthetic equivalence: how the music came into existence also matters.

Why compare mass AI-art to inorganic nature?

Nannicelli compares mass AI-art to inorganic nature because both can be appreciated for formal qualities without being understood as an artist's expression. A waterfall can appear graceful or dramatic without intending to communicate. Similarly, generated music may sound serene without expressing an AI's feelings or artistic viewpoint. The comparison concerns aesthetic appreciation, not a claim that AI systems have no human creators or technological history.

What are Wojtkiewicz’s two necessary conditions for art? Why does she think AI images can meet them?

Wojtkiewicz's conditions are that an object be produced by an agent intending it to be regarded as art and that it result from that agent's intentional action. She argues that users can satisfy them through artistic intentions, prompting, selection, and image-editing techniques. Automation alone does not exclude artistic agency, as photography shows. Nannicelli questions whether ordinary prompting provides enough intentional control to satisfy the second condition.

What is intentional control? Why compare prompting to a Polaroid?

Intentional control means being able to anticipate how actions will affect a work and deliberately guide those actions to realize intended features. Nannicelli argues that prompting leaves many decisions inside an opaque automated process. It resembles an automatic Polaroid snapshot more than carefully controlling camera settings and darkroom development. His point concerns limited control in ordinary use, not a claim that Polaroid cameras can never produce art.

Why is Nannicelli’s skepticism “moderate”? How could someone make art using an image generator?

His skepticism is moderate because AI-generated material can become art through further human artistic activity. An artist might transform generated images, combine them into a larger composition, or present them within a carefully controlled conceptual project. The artistic significance then comes from what the person intentionally does with the outputs. Nannicelli doubts ordinary automatic generation, rather than denying every possible artistic use of image generators.

If Nannicelli is right that knowing where a work came from matters to how we should experience it, what follows about labeling AI-generated content?

An implication of Nannicelli's view is that AI labeling supports informed aesthetic judgment. Without it, viewers may mistakenly attribute human experience, artistic skill, or historical achievement to a generated output. Knowing its origin helps them decide which qualities it actually supports. Labeling should also distinguish automatic generation from human art that incorporates AI material. The point is accurate appreciation, not requiring audiences to dislike AI content.

Why does Nannicelli’s view support the idea that purely AI-generated outputs should not be protected by copyright?

His view supports withholding copyright from purely generated outputs because they may lack the human intentional control and creative achievement that authorship is meant to recognize. Merely requesting an image does not necessarily make its detailed features one's artistic accomplishment. This is an implication of his aesthetic argument, not a complete legal theory. Human creativity could still warrant recognition when someone transforms or incorporates generated material.

What do Goldstein & Salib take to be the purpose of intellectual property? Why does it matter for their argument that ideas are “expensive to produce, but cheap to use”?

Goldstein and Salib see intellectual property as a tool for encouraging socially valuable ideas. Ideas can be expensive to develop but cheap to reproduce, and one person's use does not prevent another's. Exclusive rights let creators recover development costs, encouraging production. However, these restrictions also limit access and further innovation. Their justification therefore depends on whether the incentive benefits outweigh those social costs.

Why do Goldstein & Salib think AI weakens the case for intellectual property protection? Give one of their examples.

AI weakens the economic case for intellectual property when it makes valuable ideas cheap to produce as well as distribute. Expensive human effort then needs less encouragement through monopoly rights. Their example is AlphaFold, which predicts protein structures far faster than earlier labor-intensive methods. If AI can supply sufficiently good results cheaply, they argue that granting exclusive rights over equivalent human-produced information becomes harder to justify.

Under what circumstances do Goldstein & Salib concede that protecting human creators might still be justified? What alternatives to copyright protection do they propose?

Protection might remain justified if humans create valuable innovations AI cannot match, such as new artistic styles, or if AI innovation remains relatively costly. They suggest narrowing protection to genuinely novel contributions rather than shielding all work from training. Alternatives include contracts paying humans directly for valuable training material and subscriptions funding AI services. These arrangements can reward production without exclusive rights over every output.

Why is intellectual property not welfare? Give their three reasons.

They give three reasons intellectual property is poor welfare: it favors creators over other displaced workers, burdens AI development in ways that can slow useful innovation, and creates social loss by restricting access to non-rival ideas. Its purpose should be encouraging production rather than protecting particular jobs. They instead support direct assistance, such as universal basic income funded through general taxation, for people affected by automation.

What is the fairness objection to training AI on artists’ work? Explain Goldstein & Salib’s three replies to it.

The fairness objection says creators deserve compensation because their work helped AI companies profit. The authors give three replies. First, helping someone succeed does not automatically create a right to payment. Second, intellectual property allows unpaid benefits, such as borrowing an artist’s style. Third, unsettled law gave neither artists nor AI companies a guaranteed expectation about training. They favor rules that encourage valuable new ideas.

How might Nissenbaum respond to Goldstein & Salib’s claim that artists had no settled expectation that their work would not be used to train AI?

Applying contextual integrity, Nissenbaum could argue that legitimate expectations arise from social norms as well as explicit laws. Publishing art for viewing or discussion does not necessarily authorize every new commercial use. AI training changes the purpose, recipients, and scale of use, which warrants scrutiny even without settled AI-specific law. Her framework would evaluate whether those changed flows are justified rather than assuming public access settles the issue.

What are the Human Artistry Campaign’s core principles? Which concern how AI systems are trained, and which concern what AI systems produce?

The seven principles support AI as a creative tool, affirm human art's value, require authorization and licensing, reject special IP exemptions, reserve copyright for human creativity, demand transparency, and include creators in policymaking. Training concerns include permission, licensing, and input records. Output concerns include AI labeling, human authorship, and consent for voices or likenesses. Transparency covers both training and outputs.

Why does the campaign insist that using creators’ works, voices, or likenesses requires “authorization and free-market licensing from all rightsholders”? Which of our ethical frameworks best supports this demand?

The campaign demands authorization and licensing so creators control uses of their work and identities, negotiate compensation, and refuse unwanted exploitation. Rights-based or deontological ethics most directly supports this demand because consent and legitimate claims matter beyond aggregate benefits. A Kantian approach emphasizes treating creators as ends, not merely resources. Contextual integrity also questions whether public availability authorizes unrelated commercial reuse.

The campaign describes AI training on creators’ work as “theft” (“Stealing isn’t innovation”). What is the case for this description, and how would Goldstein & Salib object to it?

The theft description emphasizes taking benefits from protected creative labor without permission or compensation, potentially to compete with its creators. Goldstein and Salib would reply that ideas are non-rival, copying differs from taking physical property, and intellectual property rights serve limited incentive purposes. Calling training theft assumes the disputed claim that creators should control that use. They would instead ask which rules best encourage valuable ideas.

Why protect only human creativity? How does Nannicelli support this?

The campaign says copyright should reward human creativity, skill, labor, and judgment, which purely machine-generated outputs lack. It still supports humans using technology as a creative tool. Nannicelli's argument supports this distinction by questioning whether ordinary AI generation involves sufficient human intentional control for artistic authorship. He also allows art when people deliberately transform generated material or incorporate it into a larger creative work.

How does imitating style differ from using a voice or likeness?

Imitating style borrows general creative features, while using a voice or likeness invokes an identifiable person's identity. Both may affect livelihoods, but the second also raises concerns about impersonation, false endorsement, consent, and control over personal representation. A cloned voice could make listeners believe a performer said something they never said. The ethical problem therefore concerns the treatment of the person, not only the use of artistic techniques.

What do the Human Artistry Campaign and Goldstein & Salib agree about, and where does their real disagreement lie?

Both recognize AI's potential benefits and support withholding copyright from purely AI-generated outputs, though for different reasons. Their central disagreement concerns control over training. The campaign emphasizes authorization, compensation, and human creative rights. Goldstein and Salib ask whether restrictions increase valuable innovation, favoring weaker rights when incentives are unnecessary and direct welfare rather than copyright as protection against displacement.