PHIL 123 · AI & Ethics · Chapter 15 of 18

Offline reading copy · canonical web edition: https://ethicsandai.your-digital-life.org/chapter/creativity-innovation-and-ai/

Creativity, Innovation, and AI: Authorship, Judgment, and the Ethics of Generated Work

In 2022, an image titled Théâtre D’opéra Spatial won first place in a digital arts category at the Colorado State Fair. The image looked like a dramatic scene from a strange opera house: robed figures, an ornate stage, and light pouring through a circular opening. The controversy began when people learned that Jason Allen had made the image using Midjourney, a generative AI image system. Allen did not simply type one sentence and accept the first result. In the U.S. Copyright Office record, he said he entered prompts and revisions at least 624 times, selected an image, then used Photoshop and Gigapixel AI for later editing and upscaling. When he tried to register the work with the U.S. Copyright Office, the Office refused registration as submitted because the work contained more than a minimal amount of AI-generated material that Allen would not disclaim.

The case is useful because every simple reaction misses something. One person sees a striking image and says, “That is art.” Another sees a contest entry made with a system other entrants may not have expected and says, “That is unfair.” Allen sees his prompting, selection, revision, and editing as creative labor. The Copyright Office Review Board sees a legal authorship problem: U.S. copyright protection requires human authorship, and the Midjourney-generated expressive content could not be claimed as Allen’s human-authored expression without disclaimer. The Office did not say that a human can never make copyrightable work with AI assistance. It said this work, as submitted, still contained AI-generated material that had to be excluded from the copyright claim.

AI ethics has to hold those layers together. The image can be visually interesting, the human effort can be real, the contest fairness question can remain unsettled, the copyright claim can fail as submitted, and the training-data question can still be ethically contested. The same case contains judgments about product, process, authorship, labor, audience trust, and field-level effects.

This chapter argues that AI creativity cannot be judged by asking once and for all whether machines are creative. Responsible judgment asks what kind of practice a tool helps build. It looks at the output, the process, the human contribution, the prior labor and data, the audience’s trust, the field-level effects, and the system’s ability to respond when harm appears.

That is why this chapter belongs in PHIL 123. To build like a philosopher in a creative or innovative field is to design a practice with visible assumptions, clear standards, accountable judgment, and responsibility to affected people. Students, designers, clinics, publishers, and scientific teams all face the same broad demand when AI enters the work: describe the practice before defending it.

What Are We Judging?

AI creativity debates often get stuck on questions like “Can AI make art?” or “Is AI creative?” Those questions catch only part of the problem. Students now face a wider range of cases: chatbots drafting poems, image models generating scenes, design systems producing logo variations, music models composing songs, coding agents building games, and scientific AI systems predicting protein structures. These cases raise different ethical questions.

The first philosophical move is to separate the layers of judgment. The Stanford Encyclopedia of Philosophy entry on creativity notes that philosophers ask about creative products, creative processes, and creative persons. For AI ethics, we need one more layer: creative practices. A practice includes the surrounding tools, data, institutions, audiences, incentives, and responsibilities.

Four layers for judging AI-assisted creativity and innovation
Layer Main question
Product What was made, predicted, designed, or changed, and what standard should it meet?
Process How was the output produced, selected, revised, checked, or validated?
Person or contributor Where did human judgment, skill, intention, disclosure, or responsibility appear?
Practice What kind of creative field does this use help build?

Those four layers keep us away from two bad shortcuts. “AI made it, so it is fake” ignores careful human uses of AI. “The output is impressive, so the practice is creative and responsible” ignores authorship, disclosure, prior labor, audience trust, and field-level harms.

Return to Théâtre D’opéra Spatial. Product judgment asks whether the image is visually compelling. Process judgment asks what Allen did and how much control he had over the expressive elements. Person-level judgment asks what intention, revision, and skill belonged to Allen rather than to the model. Practice-level judgment asks about contest disclosure, audience expectations, Midjourney’s training data, future contest rules, and the effect on digital artists. Careful ethical analysis allows those questions to produce different answers.

Key Point: The Output Is Not Enough

An AI-generated or AI-assisted output can be impressive while the practice behind it remains ethically thin. A modest output can also come from a careful, disclosed, responsible practice. In this chapter, creativity and innovation are judged across product, process, person, prior labor, audience, field, and accountability.

The chapter gives a conceptual genealogy of the debate. Older discussions of creativity often focused on inspiration, genius, originality, or the artwork itself. Contemporary philosophy of creativity asks more carefully whether creativity belongs to a product, process, person, or practice. Margaret Boden gives a process vocabulary. Berys Gaut gives a value-and-agency vocabulary. Copyright and authorship debates ask what kind of human judgment belongs in a work. Labor and training-data debates ask whose prior work made the system useful. Scientific innovation cases such as AlphaFold ask how AI can extend inquiry inside a validated practice. Responsible innovation asks how new capabilities should be steered.

The chapter follows that chain: sameness and slop, Boden and Gaut, human judgment, labor and copyright, AlphaFold as a positive innovation case, and a practical framework students can use in their own fields.

When Novelty Becomes Sameness or Slop

Generative AI systems can make more content than human audiences can evaluate. They can produce images, articles, books, comments, songs, ads, summaries, and fake expertise at a scale that makes novelty feel cheap. The result may be more output with less freshness, care, originality, or value. A creative field can become faster and less alive at the same time.

The term “AI slop” has become a public label for low-quality AI-generated material that fills online spaces: bizarre images on social platforms, search-optimized articles with little expertise, fake or shallow books, spam submissions to magazines, AI-reworked Reddit posts, synthetic music, junk websites, and low-care summaries. The term is messy, and careless use can turn it into an insult for every AI-assisted work. Some AI-assisted work is careful, disclosed, revised, checked, and useful. Slop names a failure mode: production becomes cheap while judgment, verification, context, care, and responsibility become scarce.

Slop connects directly to the philosophy of creativity because novelty alone is cheap. A generated article about mushroom foraging may contain newly arranged sentences, but if it is inaccurate, it creates danger. A generated book may contain original wording in the weak sense that no one wrote those exact sentences before, but if it misleads buyers, crowds out better books, or impersonates expertise, the novelty has little ethical value. A Reddit post may be entertaining, but if it is fabricated to harvest attention, the community now has to spend energy deciding what can be trusted.

Slop appears in different ways across different fields.

In culture and aesthetics, slop can narrow what people see. Platforms reward quantity, speed, and engagement. If generated images, stories, songs, and videos fill the feed, audiences may have to sort through more material to find work with care behind it. Artists and writers may face a market where cheap imitation travels faster than craft.

In education, slop can weaken learning. If students use AI to produce a polished paragraph before they have struggled with the idea, the output may hide what they do not understand. Students should receive help, but some help short-circuits the slower work of forming judgment, vocabulary, and taste. A student who only accepts the first fluent output may never learn how to decide whether the output is any good.

In professional fields, slop can shift costs onto others. Editors, moderators, librarians, teachers, hiring committees, clients, clinicians, and readers may now spend more time filtering plausible but low-care material. A workplace may celebrate “productivity” because more drafts, posts, reports, or designs appear, while the review burden quietly grows. A field can become faster and worse at the same time.

Caution: Slop Is a Failure Mode, Not a Synonym for AI Use

Slop is low-care synthetic material that shifts judgment and cleanup costs onto others. Responsible AI-assisted work can move in the opposite direction: it can be disclosed, revised, checked, situated, and accountable to a real audience.

Slop also changes the social environment around creative work. In magazine publishing, editors have reported being overwhelmed by machine-generated submissions. In online communities, moderators describe a new burden of deciding whether posts are authentic, fabricated, copied, or AI-reworked. In book marketplaces, authors and readers worry that low-quality generated books can crowd search results, borrow the credibility of real authors, or mislead buyers who assume a human expert stands behind the work. The Authors Guild is a stakeholder advocacy source rather than a neutral measurement of the whole marketplace, but the concern is still ethically important because writers, publishers, booksellers, and readers are among the people who bear the costs of synthetic flooding.

The Amazon KDP disclosure policy shows how platforms are beginning to separate AI-generated and AI-assisted work. Amazon asks publishers to disclose AI-generated text, images, or translations. It distinguishes those cases from AI-assisted brainstorming, editing, refinement, or error-checking where the human created the underlying content. That distinction is imperfect, but it shows the policy problem. A platform needs more than a yes-or-no AI label. It needs a way to ask how much of the finished material came from the system, how much human review occurred, and whether the reader is being misled about expertise, quality, or authorship.

Language and idea homogenization are related risks. Research by Anderson, Shah, and Kreminski found that, in a 36-participant comparative study, people using ChatGPT for creative ideation produced ideas that were less semantically distinct from one another than people using an alternative creativity-support tool. The same study found that ChatGPT users generated more detailed ideas and felt less responsible for them. That does not prove that AI always makes creativity homogeneous. Other work, such as Ashkinaze and colleagues’ large dynamic experiment on AI ideas, found a more complicated pattern: high exposure to AI-generated examples increased collective idea diversity without improving individual creativity. The evidence is still developing. The cautious conclusion is enough for this chapter: common AI systems can shape the pattern pool from which many people draw, and that can affect voice, variety, responsibility, and the norms of a field.

The homogenization problem begins when many users draw from the same high-probability language, visual, musical, or design patterns. A model trained to produce likely continuations tends to return patterns that fit what has already been common or rewarded. That can be useful when the goal is clarity, convention, or rapid drafting. It can be harmful when the goal is local voice, difficult thought, experimental form, or work that resists the obvious next phrase. Students should notice the difference between using AI to test alternatives and using AI to replace the slow formation of taste.

Slop is a practice problem as much as a content-quality problem. Easy production raises the value of judgment. Uncertain authenticity raises the value of disclosure. Shared model defaults raise the value of local voice and situated purpose. Marketplaces that reward volume create a need for field-level governance.

A useful comparison comes from literary prizes. In 2026, the Commonwealth Foundation issued an update after serious concerns were raised about alleged AI use in the Commonwealth Short Story Prize. The Foundation said it did not use AI tools as conclusive evidence because of concerns about artistic ownership, consent, and the limits of detection tools. It instead reviewed process evidence, including discussions with winners, working drafts, time-stamped documents, and notes. After that review, the Foundation said it was satisfied that AI was not used to write the winning stories. That case does not show that detectors are useless or that contests should ignore AI concerns. It shows why process evidence matters. A fair practice must protect readers and entrants from deception, but it must also protect writers from unsupported accusations.

Generative systems can support creativity when they help a human explore, revise, test, translate, prototype, or communicate with care. The same kind of system can degrade creativity when it floods a space with plausible material that no one has really judged. The ethical question is whether the practice preserves or weakens the conditions under which creativity and knowledge can be trusted.

What Kind of Creativity Is Happening?

The slop problem gives us the stakes. Boden gives us the first set of concepts. To preserve creativity and innovation, we first need to say what kind of creative activity is happening.

Margaret Boden’s account of creativity, developed across works such as The Creative Mind and her writing on creativity and artificial intelligence, gives us a helpful starting point because she asks how creative processes work. Her framework is usually summarized through three kinds of creativity: combinatorial, exploratory, and transformational. These categories are not moral verdicts. They classify the activity before we decide whether the activity is responsible.

Boden also distinguishes psychological creativity from historical creativity. The SEP discussion of creativity explains this through P-creativity and H-creativity. A work is psychologically creative if it is new to the person who produced it. It is historically creative if it is new in all of history. Students need both categories. An AI-assisted idea may help a student see a connection they had not seen before, and that can be educationally valuable even when the idea is not historically original. A classroom may care partly about P-creativity because learning involves making something newly one’s own. A contest, publication, patent system, scientific field, or artistic movement may care much more about H-creativity because the claim is field-level originality.

Combinatorial creativity brings familiar ideas together in a new combination. A joke connects two frames that normally stay apart. A designer combines retro typography with a modern interface. A student pairs a personal story with a philosophical framework. Generative AI is often good at this kind of work because large models can draw connections across patterns in language, images, code, music, and design. A chatbot can suggest unexpected analogies. An image model can blend visual styles. A coding assistant can combine familiar libraries in a quick prototype.

Combinatorial creativity is valuable when the combination has purpose. It is thin when it is only collage. A generated image described as “cyberpunk Van Gogh in the style of stained glass” may be visually striking, but it may also be shallow if it only stacks recognizable cues. A student using AI to compare Aristotle’s virtue ethics with a nursing scenario may produce a useful combination if the student understands both sides and revises the result. The same process can be learning or slop depending on the judgment around it.

Exploratory creativity moves inside an existing space of rules or possibilities. A jazz musician improvises within a harmonic structure. A chess player explores legal moves. A game designer experiments within the constraints of a genre. A scientist explores a space of possible molecules or protein structures. Many AI tools fit this category well. They can generate variations quickly, expose alternatives, and help a human search a space more widely than the human could manage alone. A designer asking for twenty layout variations may not want the tool to invent a new theory of design. The designer may want to explore options inside an existing design space, then choose and revise with human judgment.

Exploratory creativity can support judgment or replace it. It supports judgment when it expands the set of options a human can evaluate. It replaces judgment when the human treats the tool’s list as if the list were already the answer. A student asking a chatbot for five possible objections to an argument may learn if they compare, reject, refine, and explain the objections. A student who simply copies the strongest-sounding objection without understanding it has outsourced the very activity the assignment was meant to develop.

Transformational creativity is harder because it changes the space itself. Boden’s phrase “conceptual space” is important here. A conceptual space is the generative system that defines a domain’s range of possibilities: chess moves, jazz harmonies, architectural constraints, painting styles, molecular structures, research methods, or genres of writing. Exploratory creativity moves through that space. Transformational creativity changes the rules or constraints of the space so that outcomes become possible that were previously outside the field’s ordinary ways of thinking.

That is why transformational creativity is difficult to judge in the moment. In art, it might involve a movement that changes what counts as painting, music, or performance. In science, it might involve a model or method that changes what researchers can ask. In technology, it might involve a platform or technique that reorganizes a field. AI systems may help people explore an existing space more quickly, and a smaller number may alter a field’s space of possibilities. That stronger claim takes time to evaluate because a field must decide whether the change produced understanding, value, better practice, and responsible use.

Boden’s categories shift the AI discussion from output status to process. A prompt that asks for a familiar style mashup is often combinatorial. A designer using AI to explore hundreds of packaging variations is doing exploratory work. A scientific system that changes how researchers approach a long-standing problem may belong closer to the transformational end, though we should be careful about attributing the transformation to the software alone. The categories let us discuss the process without pretending every generated output has the same status.

Pause: Boden Classifies the Process

Boden helps us ask what kind of creative operation is happening: combining, exploring, or transforming. She does not decide whether the operation is fair, disclosed, legally protected, educationally useful, or socially valuable. A process can be combinatorial, exploratory, or even field-changing while still raising ethical problems.

Boden is especially useful for AI ethics because she moves the question from mystery to process. We can ask whether the AI combined familiar patterns, helped a person search a possibility space, changed a field’s methods, supported a student’s understanding, produced historically new work, or created a plausible imitation of work that already existed. Those process questions are more precise than “is it creative?”

Creativity, Value, Skill, and Judgment

Process gets us only part of the way. A system can combine ideas, explore variations, or transform a conceptual space in ways that are useless, harmful, manipulative, or careless. We also need value.

Berys Gaut’s work on the philosophy of creativity is helpful because it treats creativity as broader than art. Creativity appears in science, craft, business, education, and everyday problem solving. Many AI creativity questions will not look like museum or music debates. A nurse may use AI to redesign a patient-education handout. A small business owner may use AI to test marketing ideas. A biology lab may use AI predictions as part of a discovery process. A student may use AI to clarify an argument. These are creative practices because people are trying to make something new and valuable under constraints.

Gaut also keeps originality, value, agency, and skill in the conversation. Philosophers often say creativity requires novelty and value. That phrase sounds simple, but each word needs interpretation. Novelty asks whether the output or idea is new in some relevant sense. Value asks whether the novelty serves a real purpose or standard. Agency asks whether the result came from a person or process that can count as more than accident. Skill asks whether the achievement involved competence, control, judgment, or understanding.

Originality without value gives us noise. A string of random words may be new, but that does not make it creative in a meaningful sense. A bizarre image may be novel and still empty. A generated business idea may be different and still useless. A design may look unusual and still fail the user. Value without agency gives us a different problem. A diamond may be valuable and unusual, but geological pressure is not creative in the same way a jeweler is. Pure luck is also different from creativity. If someone accidentally spills paint and the result looks beautiful, the canvas may be interesting, but the accident does not show the same kind of creative achievement as a painter who understands what they are doing.

AI makes these distinctions visible. Generative systems can produce novelty faster than humans can judge value. They can produce outputs that look skillful without showing human skill in the ordinary way. They can create plausible variations that feel like achievement even when the user did little more than request them. This does not mean AI-assisted work cannot be creative. It means the human contribution has to be described carefully.

Value also has to be assessed inside a practice. A design that looks clever may fail if it confuses users. A generated story may have vivid sentences and still feel empty because the plot, voice, and point of view do not hold together. A scientific prediction may be valuable even if it is not beautiful, because it helps researchers identify the next experiment or line of inquiry. A classroom use of AI may be valuable when it helps a student notice a weak assumption, and harmful when it hides that weakness under fluent wording. Creativity is a judgment about novelty, value, process, agency, skill, and use.

This also means creativity can be ethically mixed. A manipulative phishing email can be creative as deception. A deepfake used to humiliate someone can be technically inventive. A scam book can show efficient use of tools. A persuasive ad campaign can be ingenious while exploiting insecurity. A surveillance interface can be elegantly designed while helping an institution treat people as objects. Creativity does not automatically make a practice good.

The phrase “good of its kind” helps here. A horror story can be good as a horror story even if it is not comforting. A medical handout can be good if it is accurate, readable, culturally appropriate, and clinically reviewed, even if it is not artistically daring. A phishing email can be good as a phishing email because it deceives effectively, but that only shows technical or strategic success within a bad practice. Students should not confuse local success with moral justification. Ask what kind of thing the work is, what standard it is supposed to meet, and whether the practice itself deserves support.

For PHIL 123, Gaut’s lesson can be turned into four questions:

  1. What is new here, and new to whom?
  2. What is valuable here, and by what field standard?
  3. Who exercised judgment, skill, selection, revision, or responsibility?
  4. What is the practice for, and who bears its costs?

Return again to Théâtre D’opéra Spatial. The image may appear original to many viewers. The process involved human prompting, selection, post-processing, and machine generation. Allen did more than click once, but the expressive features generated by Midjourney were not under his control in the way a painter’s brush strokes would be. The contest fairness question depends partly on disclosure and rules. The copyright question turns on human authorship under current law. Calling the image creative does not settle any of those questions.

This template helps students avoid a common mistake. A person defending AI-assisted work may point to effort: “I spent hours prompting.” Effort matters, but effort is not the same as authorship, value, responsibility, or fairness. A person criticizing AI-assisted work may point to the model: “The machine made it.” Machine contribution matters, but it is not the same in every case. Some human uses involve deep revision, validation, context, and responsibility. Others involve almost none. The ethical analysis starts when we reconstruct the practice.

Human Judgment Is More Than Tool Use

People often defend AI-assisted work by saying that artists, writers, scientists, and designers have always used tools. The claim is true and too quick. A tool can extend human action in many different ways. A pencil records a mark the hand makes. A camera captures light through a designed apparatus. A digital brush responds to a user’s movement. A spreadsheet helps calculate. A simulation helps explore possible outcomes. A generative model can produce detailed expressive content from a prompt in ways the user did not specify and may not be able to predict. All of these can be tools, but the human contribution differs.

The U.S. Copyright Office has drawn a distinction that helps clarify this issue even outside law. In its Part 2 report on copyrightability, the Office says current U.S. copyright protection requires human authorship. AI assistance can still leave room for protectable human expression through selection, coordination, arrangement, editing, modification, or other choices. The Office also says that material generated by AI without sufficient human control over expressive elements is not protected as human-authored expression. Under the current generally available technology it reviewed, prompting alone is usually not enough to give the user sufficient control over the expressive result.

That legal distinction trains ethical attention. The question is what the human contributed: expressive content, selection among outputs, revision, arrangement, source checking, disclosure, or responsibility for errors and effects on others.

Authorship, credit, and responsibility can come apart. A student may use AI to generate possible titles for a presentation. The student may have little authorship over each suggested phrase while remaining responsible for choosing an accurate and appropriate title. A designer may use AI to create rough mockups, then rebuild the final design by hand. The final contribution may include significant human judgment, even if AI helped the exploration phase. A person may also generate an entire story with minimal editing, submit it under a human name, and hide the process. That person may be responsible for the submission, but the authorship claim is much weaker.

Searle’s Chinese Room argument gives a compact philosophical pressure point. Searle argued that a system can manipulate symbols according to rules without understanding what the symbols mean. You do not need to accept every part of Searle’s argument to see why it remains useful in AI creativity debates. A model can produce text that looks meaningful without intending the meaning. It can produce an image that looks expressive without having a point of view. It can generate a design that looks purposeful without caring about the user. The human user may supply purpose, interpretation, selection, and revision. Or the human may simply pass along a fluent output.

Searle does not settle the AI creativity debate. He does not prove that AI can never be useful, creative, or innovative. He gives us a question: does producing appropriate symbols show understanding, or could the system be manipulating form without grasping meaning? In educational, clinical, legal, scientific, or artistic contexts, that question matters because understanding often belongs to the responsibility structure. Someone has to know what the work means, why it is acceptable, what its limits are, and when to refuse it.

The distinction becomes important in education. If a student asks a chatbot to generate an argument and submits it with little revision, the student has not shown much philosophical judgment. If a student uses a chatbot to test objections, notices where the chatbot used the wrong vocabulary, revises the argument in course terms, and explains what was accepted or rejected, the human judgment is visible. The same surface tool can support learning or replace it. The ethical difference lies in the practice.

For Théâtre D’opéra Spatial, the Copyright Office accepted in the administrative history that Allen’s Photoshop edits could contain human-authored expression, but it treated the Midjourney-generated image content differently. Students can acknowledge Allen’s effort while still asking whether that effort gave him authorship over the expressive details. They can acknowledge that the image impressed people while still asking whether other artists in the contest were competing under the same expectations. They can acknowledge that AI can help make art while still asking whether the process was disclosed clearly enough.

Caution: “Tool” Is Too Broad

Calling AI a tool may be accurate, but it rarely settles the question. Ask what the tool did, what the human did, what the audience was told, and who remains responsible for the result.

Disclosure should be tied to the audience’s real trust question. A contest judge may need to know which parts of an image were generated and which parts were human-made. A teacher may need to know how a student used AI in order to evaluate learning. A reader of a health guide may need to know whether a qualified person checked the information. A client may need to know whether a logo concept was generated using systems that imitate living artists. A researcher may need to describe the model, data, uncertainty, and validation process. A blanket label such as “AI used” may be too vague. A blanket denial may be deceptive if AI shaped the final work in a way the audience reasonably cares about.

The Commonwealth Short Story Prize case shows the other side of disclosure and evidence. People worried about AI use, but the review also had to protect writers from unfair accusation. The Foundation examined drafts, time-stamped documents, notes, and conversations with authors. That is the right kind of evidence for process questions. A detector score alone cannot carry a serious authorship judgment. A responsible field needs procedures that can investigate concerns without treating stylistic difference, second-language writing, translation, accessibility tools, or unusual revision patterns as proof of misconduct.

For students, the immediate practice is to make judgment visible. If AI helped you brainstorm, say how. If it drafted a passage, say what you changed and why. If it gave you sources, verify them. If it generated an image, explain the human contribution and the disclosure expectations of the context. Work affecting people’s health, money, education, safety, or legal standing also needs validation, expertise, and responsibility.

The Labor Behind the Output

Preserving creativity also means asking what prior labor made a tool possible. Many generative AI systems were trained on large datasets containing images, books, code, songs, websites, captions, translations, and other materials made by people. The model does not usually reproduce a training work exactly, though it sometimes can. More often, the model learns patterns from many works and generates new outputs from those patterns. That technical description leaves the ethical question open: did the system benefit from human creative labor in ways that require consent, credit, compensation, limits, or new institutions?

Trystan Goetze’s argument about AI art and labor treats artist objections as philosophical claims with reasons that can be reconstructed and tested. The objection is roughly this: some AI image systems extract value from artists’ work by using their labor to train models, then produce outputs that can compete for attention, commissions, and market opportunities against those same artists. The problem reaches beyond the user’s prompt. It concerns the social relation between prior creative labor, model training, platform profit, and downstream competition.

Put as an argument, the objection looks like this:

This objection moves the debate away from the surface of the output. A generated image might not copy any single artist. It may still depend on a training infrastructure that drew value from creative communities. A generated book may not plagiarize one author sentence by sentence. It may still enter a market already shaped by the labor of writers, editors, publishers, reviewers, booksellers, and readers. A generated song may not reproduce one track exactly. It may still imitate a style closely enough to affect a musician’s livelihood or reputation.

A defender of generative AI can reply that culture has always involved learning from previous work. Human artists study paintings they did not make. Writers imitate rhythms and genres before they find their own voices. Students learn by reading models. Musicians absorb styles from scenes and traditions. Scientists build on prior research. If learning from prior work were automatically exploitation, then nearly all education, science, and artistic development would become suspect. That reply identifies a real problem for overly broad versions of the labor objection.

The reply does not settle the commercial-scale training question. Human learning is embodied, limited, situated, and socially accountable in ways large-scale machine training usually is not. A student who studies an artist cannot instantly produce thousands of marketable variations in that artist’s style. A painter who learns from a tradition does not usually build a platform that charges users to generate substitutes for living artists. A writer who absorbs a genre still has to write, revise, publish, and answer for the work. Generative AI changes scale, opacity, speed, and market relation. Those changes explain why the labor objection deserves a serious answer even if human creativity has always depended on influence.

The strongest version of the objection does not need to say that every use of generative AI is illegal theft. That would flatten the debate. It asks more careful questions. What materials were used for training? Were they lawfully acquired? Were creators able to opt out? Did the system allow style imitation or market substitution? Who profited? Were affected communities consulted? Did users disclose AI assistance? Did the platform create ways to compensate or credit creators? Did the use support human work or replace it with cheaper outputs that trade on earlier labor?

Micaela Mantegna’s work helps explain why copyright law alone cannot carry all of these ethical concerns. Copyright is important. It shapes what can be protected, licensed, infringed, registered, or defended in court. But copyright does not map perfectly onto fairness. Mantegna argues that generative AI raises concerns about consent, attribution, compensation, labor displacement, and the broader ethics of AI development, while copyright doctrine is a blunt and sometimes distorting policy tool for those concerns.

For example, copyright protection often protects owners of copyright interests, who may be publishers, labels, studios, platforms, or companies, not necessarily the creators whose labor made the work possible. Expanding copyright protection for AI-generated outputs could also strengthen large rights-holders and platforms while doing little for individual artists, writers, voice actors, labelers, moderators, or other workers in the AI supply chain. Mantegna also warns that copyright expansion can create new hierarchies of labor: some visible intellectual work may receive protection while other work needed for AI systems, including data work and moderation, remains hidden. This is why a practice can be legally unresolved, or even partly legal, while still raising ethical questions about who benefits, who loses bargaining power, who gets credited, and who has a realistic path to compensation.

The U.S. Copyright Office’s AI materials help separate questions that often get tangled. Output copyrightability asks whether a finished work contains protectable human-authored expression. In the Théâtre D’opéra Spatial case, the Office refused registration as submitted because Allen would not disclaim the AI-generated material. Human contribution may have existed, especially in selection and later editing. The Office still declined to treat the Midjourney-generated expressive material as human-authored material Allen could register without disclaimer. The 2025 D.C. Circuit decision in Thaler v. Perlmutter also reinforces the current legal baseline that U.S. copyright authorship requires a human author.

Training-data fairness asks a different question: whether using copyrighted works to train AI systems is lawful and fair. The Copyright Office’s Part 3 pre-publication report on generative AI training treats training as a fact-specific issue rather than automatically permitted or automatically forbidden. The report also warns against treating creative works as mere neutral data; copyrighted works embody protected human expression even when they become part of a dataset.

Recent cases show why details matter. Thomson Reuters v. Ross Intelligence involved legal-research materials and a competitor’s AI-assisted legal search tool, not a general-purpose image or chatbot model. Bartz v. Anthropic produced a mixed June 2025 ruling: the court treated training on lawfully acquired books differently from claims involving pirated copies and retained libraries. According to Reuters reporting in May 2026, a proposed $1.5 billion settlement of the authors’ claims was still under judicial review and had not yet received final approval. Kadrey v. Meta granted Meta summary judgment on the record before that court, but the opinion itself warned against reading the decision as a general rule that using copyrighted works to train language models is lawful in all cases.

Third, ethical responsibility asks what a person, company, institution, or field should do while the legal answer remains unsettled. A college, publisher, design firm, lab, or student can ask whether AI use should be disclosed, whether creators should be compensated, whether certain styles or living artists should be off-limits, whether a field needs licensing norms, whether audiences should know when they are reading generated content, whether a company should profit from a tool trained on communities it now competes against, and whether a student should avoid a tool because it conflicts with the learning goals of the assignment.

These questions return us to Théâtre D’opéra Spatial with more precision. The case asks how much human control counts as authorship, what contest entrants should disclose, whether other entrants were competing under the same expectations, what the training data contributed, and which parts of the image came from Allen’s judgment rather than a model trained on prior images. Some of those questions belong to copyright. Some belong to contest fairness. Some belong to a broader ethics of creative labor.

The labor objection also has a limit. Human creativity has always involved influence, imitation, education, apprenticeship, quotation, remix, genre, tradition, and borrowing. Artists learn by studying other artists. Writers absorb styles. Musicians imitate before they innovate. Scientific fields advance by building on prior work. A world where no one could learn from prior human creativity would destroy creativity. The ethical problem comes from the scale, opacity, commercial structure, lack of consent, and potential market substitution involved in many AI systems.

A slogan cannot sort those cases. A responsible field has to distinguish training, citation, licensing, imitation, analysis, fair use, and substitution. It also has to distinguish a student’s small experiment from a company’s commercial deployment. The same ethical vocabulary cannot do all the work at once. Students should learn to ask which part of the practice is under judgment.

Innovation Is More Than New Stuff

Creativity and innovation overlap, but they are not identical. Creativity often focuses on producing something new and valuable. Innovation usually emphasizes putting something new into use: a method, product, tool, workflow, or form of organization. AI makes this distinction important because many AI systems are marketed as innovative simply because they are new, fast, or powerful. Novelty and speed are not enough. The ethical question is whether the innovation helps a field pursue valuable goals responsibly.

AlphaFold gives us a strong positive case. Proteins are made from chains of amino acids, and their three-dimensional structures affect what they do in living systems. For decades, predicting protein structures from amino acid sequences was one of biology’s hardest computational problems. In 2024, the Nobel Prize in Chemistry was awarded partly to Demis Hassabis and John Jumper for developing an AI model for protein structure prediction, and partly to David Baker for computational protein design. The AlphaFold Protein Structure Database, developed with EMBL-EBI, provides open access to predictions for over 200 million protein structures.

The mechanism matters. AlphaFold does not look at a protein through a microscope. It starts from the protein’s amino-acid sequence, then uses related biological information, including aligned sequences from evolutionarily related proteins, to infer which parts of the chain are likely to be near each other in three-dimensional space. In the AlphaFold Nature paper, the system is described as predicting three-dimensional atomic coordinates from the primary sequence and homologous sequence alignments. Its architecture represents multiple sequence alignments and pairwise residue relationships, updates those representations, and produces structure predictions with confidence estimates.

This is an AI success story, but it should not be told as machine genius replacing scientists. AlphaFold depends on prior scientific work, known protein structures, sequence databases, engineering teams, benchmarks, and researchers who know how to interpret predictions. A predicted structure is not the same as complete biological understanding. It can guide research, narrow a search space, suggest hypotheses, and accelerate parts of a workflow. It still requires domain expertise, experimental follow-up, uncertainty communication, and responsible use.

AlphaFold helps show why innovation ethics needs a different frame from simple AI art debates. The question is not whether a protein prediction is authentic in the way an artwork might be. The questions are more like these: Does the system support reliable scientific inquiry? What uncertainty remains? Who has access? Who gets credit? What public value does the tool create? What happens when researchers, companies, or governments use the predictions for different purposes? How should the field communicate the limits of the tool? What governance practices are needed as AI systems become more central to scientific discovery?

Philosophy of science helps here because discovery is a practice involving instruments, models, data, teams, institutions, criticism, replication, and interpretation. AI systems can change where some of those activities occur. They may help generate hypotheses, identify patterns, predict structures, design molecules, or sort through enormous spaces of possibility. But a prediction still needs a scientific community capable of testing, explaining, using, and challenging it.

AlphaFold keeps the chapter from treating AI creativity and innovation only through slop and copyright. AI can extend human inquiry in ways that appear genuinely valuable. It can help researchers approach problems that would otherwise remain painfully slow. It can support medical research, biology, chemistry, and other fields. It can give more people access to tools that were once limited to specialized labs.

At the same time, AlphaFold should not become a blanket defense of AI. A high-value scientific system does not make every generative content system responsible. Protein-structure prediction and spam publishing are different practices. An open scientific database and a closed commercial image generator raise different questions. A lab prediction requiring expert interpretation and a public-facing health article written by a chatbot have different risk profiles. Responsible judgment depends on the field, the use case, the validation process, the affected people, and the available governance.

AlphaFold also raises questions about credit. The Nobel Prize went to human scientists. The system itself did not receive the prize as an autonomous discoverer. Students need this distinction because AI systems often become characters in public stories. We say “AlphaFold discovered” or “AI solved” because it sounds dramatic. More careful language reveals the actual practice: teams built systems, systems generated predictions, databases made predictions available, researchers used them, and institutions evaluated the contribution. Credit and responsibility both belong to that distributed practice.

The case also raises questions about openness. AlphaFold DB’s open access has been part of its public value, and that access shapes ethical judgment. A tool that helps many researchers work on important problems may deserve different evaluation from a tool that mainly captures value for a platform while shifting costs onto users, workers, or audiences. Innovation should be judged by the goods it helps people pursue and by the social arrangements through which it pursues them.

AlphaFold also shows why prediction and understanding should be separated. A prediction can be extremely useful before scientists fully understand every causal pathway involved. That does not make the prediction worthless. It also does not make the prediction self-interpreting. If researchers treat an AI system as an oracle, they may overtrust outputs, ignore uncertainty, or lose sight of the experimental work that gives predictions meaning. If they treat the system as a powerful instrument inside a scientific practice, they can ask concrete questions: What confidence should attach to this prediction? What experiment would test it? What domain knowledge is needed? What use would be irresponsible? The innovation lies partly in the model, partly in the database, partly in the scientific community that learns how to use the predictions well.

Key Point: Prediction Is Not Understanding

AlphaFold shows that AI can be scientifically valuable without being an autonomous knower. A useful prediction can transform a research workflow, but scientists still have to interpret, test, contextualize, and govern the use of that prediction.

The separation between prediction and understanding applies outside science too. A chatbot can predict a plausible answer without understanding the client’s legal situation. An image model can generate a plausible medical illustration without knowing whether the anatomy is right. A design model can produce a convincing interface without understanding the disabled user’s experience. A writing model can produce a polished paragraph without knowing whether the argument is true. Prediction-like fluency can be useful, but it needs human judgment inside the relevant practice.

Responsible Innovation Across Uneven Fields

Responsible innovation is a useful framework because it asks how new capabilities should be steered. Jack Stilgoe, Richard Owen, and Phil Macnaghten are associated with a framework often summarized through four dimensions: anticipation, inclusion, reflexivity, and responsiveness. For students, the framework can be translated into plain questions.

Anticipation asks what futures a technology might create. What could happen if this tool became common in a field? What benefits could it produce? What failures, shortcuts, incentives, and harms might follow? Anticipation is disciplined imagination, directed beyond the immediate convenience of a tool.

Inclusion asks who should be part of the conversation. AI creativity and innovation affect more than the user. Artists, writers, students, patients, researchers, workers, audiences, platform moderators, and future practitioners may all bear costs or receive benefits. Inclusion gives affected people a voice without giving every person veto power over every tool. It prevents designers, companies, and early adopters from capturing all the language of progress while affected people become invisible.

Reflexivity asks people and institutions to examine their own assumptions. Why do we call this innovation? Are we optimizing speed, profit, access, learning, care, recognition, discovery, fairness, or public benefit? What are we tempted to ignore because the tool is impressive? Reflexivity is especially important with AI because fluency can hide weakness. A generated report, image, or model output may feel complete before anyone has asked what standard it should meet.

Responsiveness asks whether people can change course when evidence, harm, or stakeholder feedback appears. A responsible practice includes ways to revise, slow down, disclose, validate, compensate, or stop. If a company, school, lab, or creator cannot respond to foreseeable problems, then the innovation is brittle. It may look agile while becoming ethically stubborn.

Responsible innovation connects directly to building like a philosopher. A philosopher asks what assumptions a tool carries, what kind of human practice it forms, who gets power, who loses visibility, what objections matter, and what would count as responsible revision. In design terms, this resembles the broader philosophy-of-technology claim that technologies shape possibilities, roles, values, and institutions.

The unevenness across fields is central. AI-generated marketing copy, AI-assisted architecture, AI-supported drug discovery, AI-generated fiction, AI design mockups, and AI tutoring do not carry identical duties. A field with high safety risks needs stronger validation. A field with vulnerable users needs stronger care. A field built on personal expression needs clearer authorship norms. A field where workers’ prior labor trains the system needs stronger attention to consent and compensation. A field where audiences cannot easily verify claims needs clearer disclosure.

This field-specific approach avoids two lazy positions. One says AI is creative and innovative, so resistance is backward. The other says AI is extractive and fake, so responsible use is impossible. Both positions avoid the slower work of judgment. The better practice asks what the tool is doing inside a specific activity: extending human capability, hiding labor, lowering standards, supporting access, replacing learning, flooding a field with low-value output, creating public goods, concentrating power, or preserving human accountability.

Consider several examples.

In writing education, an AI tool might help a student see possible structures, test an objection, or revise a confusing sentence. It can also produce fluent paragraphs that prevent the student from learning how to think through the argument. Responsible use would focus on visible judgment: what the student asked, what the tool supplied, what the student checked, what the student rejected, and what the student can now explain.

In design, an AI system might help a small nonprofit produce usable visuals without hiring a full design team. It can also create generic branding, imitate living artists, or flood clients with plausible mockups no one has examined carefully. Responsible use would focus on context, disclosure, human selection, accessibility, audience needs, and respect for creative labor.

In health communication, an AI system might help a community clinic create patient handouts at appropriate reading levels in multiple languages. It can also generate medical misinformation, erase culturally relevant details, or produce translations that sound fluent but mislead patients. Responsible use would require clinical review, translation review, source checking, update dates, and clear responsibility.

In science, an AI system might help researchers search a vast problem space, generate predictions, or find patterns no human could see unaided. It can also produce overconfident results, obscure uncertainty, or concentrate scientific infrastructure in a few companies. Responsible use would focus on validation, openness, credit, governance, uncertainty, and public benefit.

In entertainment, an AI system might help a small game studio prototype characters, dialogue, sound effects, and environments. It can also imitate living artists, replace entry-level learning opportunities, or generate assets whose provenance is unclear. Responsible use would focus on licensing, credit, disclosure, labor development, and the difference between prototyping and publishing.

These examples show why the chapter began with practice. The ethical questions change because the practice changes. A single AI policy or slogan cannot judge all creative and innovative work. Students need questions that travel across fields while still forcing attention to the concrete case.

A Creativity and Innovation Judgment Audit

Every AI-assisted creative or innovative project should be designed around judgment. A person building like a philosopher asks what the tool can make and what kind of practice the tool normalizes.

Begin with description before verdict. Identify the AI-assisted practice and the field standard it should meet. Then trace the process. Did AI combine familiar materials, explore variations, transform a possibility space, validate a claim, or mostly generate volume? Locate the visible human judgment in goal-setting, selection, revision, source checking, disclosure, validation, or refusal. Then widen the frame to prior labor, affected people, trust, field-level effects, and who can respond when something goes wrong.

Those questions should not produce one universal AI rule. They should produce a field-specific judgment.

Suppose a community clinic wants to use AI to draft a one-page diabetes-management handout for patients. The product standard is accuracy, readability, cultural fit, accessibility, and clinical usefulness. The process might involve AI generating a first draft, but that does not make the practice responsible by itself. Human judgment needs to appear in clinical review, source checking, language access review, and final approval by qualified staff. Prior labor appears in medical research, public-health writing, translation work, and the training data that made the model fluent. The audience trust question is more specific than “was AI used?” A patient needs to know whether the handout was clinically reviewed, when it was last updated, and who is responsible for the information. The field-level risk is that clinics may publish fluent but inaccurate materials faster than staff can review them. The field-level benefit is that small clinics may gain capacity to produce clearer, more accessible resources. Responsible innovation would require clinicians and patients in the review process, attention to care rather than mere speed, and a way to revise the material when errors or access problems appear.

The same analysis works for a student project. In graphic design, evaluate the creative process the generated logo supports, whether it imitates living artists, what the client understands, whether the designer can explain the choices, and how the field changes when clients expect dozens of instant options. In education, evaluate whether the tutor supports learning or replaces practice, what students disclose, what teachers can evaluate, and what skills might atrophy. In science, evaluate the validation, access, credit, and uncertainty practices that make a useful prediction part of responsible inquiry.

By the end of that analysis, the student claim should connect the process the tool supports, the human judgment that remains visible, and the field conditions that make the use responsible or irresponsible. A claim that lacks one of those pieces is too thin for PHIL 123.

That argument will not settle every case, but it blocks the most common shortcuts. Students cannot say “AI made it, so it is fake” without looking at the human practice. They cannot say “a human prompted it, so it is authored” without looking at control, disclosure, and contribution. They cannot call a system innovative when it mainly produces low-value content at scale. They cannot treat legality as if it exhausts fairness. They also cannot treat a high-value case like AlphaFold as proof that every AI tool deserves the same trust.

AI will continue to affect creative and innovative fields. The hard cases will rarely announce themselves as hard. They will appear as convenient tools, impressive outputs, new markets, cheaper workflows, faster prototypes, and exciting breakthroughs. The task is to ask what kind of practice is being built around those tools. Creativity and innovation remain human concerns because humans decide what to value, disclose, protect, validate, refuse, and help create.

A responsible account can say both sides plainly. AI can extend creativity and innovation when it helps people explore, test, validate, or communicate within a responsible practice. It can degrade a field when it hides labor, weakens learning, floods attention systems, or produces outputs no one has properly judged. In a project, students should identify the tool and field standard, show where human judgment appears, and explain who is affected and how the practice can respond.

References

The legal section should not be read as legal advice. Its purpose is to show why copyrightability, training-data fairness, and ethical responsibility are distinct questions.