Lesson 1

AI Can Tell Stories

Story generation at scale — mechanics, limitations, and the narrative uncanny valley.

When OpenAI released GPT-3 in 2020, researchers and writers immediately began experimenting with long-form narrative generation. The results were striking in their local coherence and unsettling in their long-range incoherence. The model could write beautiful prose for paragraphs, then contradict established facts, forget character names, or introduce inexplicable plot elements. Researchers called this the "narrative coherence cliff" — the point where local statistical fluency could no longer mask the absence of a governing intent.

The Architecture of AI Narrative

Language models generate narrative by predicting token sequences over a finite context window. Three structural properties shape their narrative output:

Local coherence vs. global coherence: Models maintain sentence-to-sentence fluency well. Paragraph-to-paragraph logic degrades. Chapter-scale structure requires external scaffolding.
Recency bias: The model attends more strongly to recent context than distant context, causing earlier established facts to fade.
Distributional averaging: Output reflects the statistical center of training data — common plot structures, frequent character archetypes, averaged prose rhythms.

The Narrative Uncanny Valley

Readers often report a sense of unease with AI-generated narrative — something that sounds right but feels wrong. This may reflect the model's lack of a governing narrative intent: the text is pursuing statistical likelihood, not meaning. Plot events occur because they are likely given prior text, not because they serve a larger purpose.

Key Tension

The same property that makes AI narrative convincing at the sentence level — statistical fluency — makes it hollow at the structural level. Good narrative requires intention, not just probability.

Back Take Quiz

Quiz 1

AI Can Tell Stories

5 questions — free, untracked, retake anytime.

causes the 'narrative coherence cliff' in AI-generated stories?

✓ Correct — ✅ AI maintains local fluency but lacks governing intent. Over long distances, the absence of a central purpose produces incoherence.

❌ The coherence cliff: statistical fluency produces good sentences but can't substitute for a governing narrative intent that shapes meaning at scale.

is 'recency bias' in language model narrative generation?

✓ Correct — ✅ Recency bias: the model attends more to recent context, so earlier established facts, characters, and plot details fade — causing continuity errors.

❌ Recency bias: the model weights recent context more than distant context, causing earlier established facts to fade and producing continuity errors.

does 'distributional averaging' mean for AI narrative voice?

✓ Correct — ✅ Distributional averaging: AI output gravitates toward what's most common in training data — frequent plot structures, character archetypes, averaged prose rhythms. Distinctive voice requires pushing against this.

❌ Distributional averaging: AI generates what's statistically central to its training data — common patterns, frequent archetypes, averaged rhythms rather than distinctive voice.

do readers sometimes feel the 'narrative uncanny valley' with AI-generated text?

✓ Correct — ✅ The uncanny valley: statistically fluent sentences that feel wrong because there's no governing purpose behind them — events happen because they're likely, not because they mean something.

❌ The uncanny valley: text that sounds right but feels hollow because it's pursuing statistical likelihood, not meaning. Events happen because they're probable, not purposeful.

maintain long-range narrative coherence in AI-generated stories, writers need:

✓ Correct — ✅ Long-range coherence requires external scaffolding: structured outlines, periodic world-state summaries, explicit character tracking — things that compensate for the model's recency bias.

❌ Long-range coherence requires external scaffolding — outlines, world-state reminders, chapter summaries — that compensate for recency bias and provide the governing structure the model can't maintain internally.

Back to Lesson Start Lab 1

Lab 1

Narrative Architecture

Develop scaffolding strategies for AI-assisted long-form narrative.

Lab 1 — Narrative Architecture Analysis

Analyze the structural limitations of AI narrative and develop strategies for overcoming them.

The AI opens with a specific example of the narrative coherence cliff.
Develop a scaffolding approach for a long-form AI-assisted narrative project.
Address: how do you prevent distributional averaging from erasing your distinctive voice?

Consider: outline structure, context management, world-state documentation, and voice injection strategies.

🎯 AI GuideLab 1

Back to Quiz Next: Lesson 2

Lesson 2

We Tell Stories Together

Co-authorship theory, emergent narrative, and the problem of creative credit.

Researcher Ross Goodwin drove from New York to New Orleans in 2017 with a camera, GPS, microphone, and clock all feeding real-time data into a neural network trained on road trip literature. The network generated a novel in real time as he drove. The result — "1 the Road" — was published as a book. The question of who wrote it has no clean answer: Goodwin designed the system, drove the route, curated the inputs. The network generated the text. The training data supplied the aesthetic vocabulary.

Emergent Narrative

Some of the most interesting AI-collaborative narrative happens when the output is not fully controlled — where the system generates something the human didn't anticipate and couldn't have written. This "emergent narrative" is distinct from prompted generation: the AI is not executing a directive but producing unexpected material that the human then responds to, curates, or incorporates.

Curation as authorship: Selecting, arranging, and editing AI output is a creative act.
Constraint design: Designing the system that generates the output is a creative act.
Responsive authorship: Reacting to unexpected AI output and building on it is a creative act.

The Creative Credit Problem

When is a work "AI-generated" vs. "AI-assisted" vs. "human-authored with AI tools"? The field has no consensus. Practical distinctions that have emerged:

Was the human the primary conceiver of the work's meaning?
Did the human make the significant curatorial decisions?
Would a different human with the same tools produce a recognizably different work?

No Clean Answer

The Goodwin case shows why: the human conceived, drove, curated, and published. The AI generated. The training data supplied. Credit is distributed and none of the traditional categories fully fit.

Back Take Quiz

Quiz 2

We Tell Stories Together

5 questions — free, untracked, retake anytime.

makes Ross Goodwin's '1 the Road' project an example of emergent narrative?

✓ Correct — ✅ The system generated unexpected text in real time from live data inputs — the narrative wasn't dictated or fully anticipated. Goodwin designed the system and curated the result.

❌ Emergent narrative: the system generated unexpected content from real-time inputs that Goodwin hadn't scripted. He designed the system and curated — the text emerged.

is curation considered a creative act in AI-collaborative work?

✓ Correct — ✅ Curation involves aesthetic judgment: what to include, exclude, arrange, and emphasize. These decisions shape meaning — and that's a creative act.

❌ Curation involves aesthetic judgment — what to include, what to discard, how to arrange — that shapes meaning. That's creative authorship.

question best distinguishes 'AI-assisted' from 'AI-generated' work?

✓ Correct — ✅ The most meaningful distinction: was the human the primary author of the work's meaning? If yes, it's AI-assisted. If meaning emerged primarily from unguided AI output, it's closer to AI-generated.

❌ The best question: was the human the primary conceiver of the work's meaning? If yes, it's AI-assisted. If meaning emerged from AI output without strong human direction, it's closer to AI-generated.

the Goodwin case, who or what contributed to authorship of '1 the Road'?

✓ Correct — ✅ Authorship was genuinely distributed: Goodwin designed, drove, and curated; the AI generated; the training data supplied aesthetic vocabulary. No single party fully accounts for the work.

❌ Authorship was distributed across Goodwin (design, driving, curation), the AI (text generation), and the training data (aesthetic vocabulary). None of the traditional categories fully fit.

does 'responsive authorship' mean in AI-collaborative creative work?

✓ Correct — ✅ Responsive authorship: rather than directing AI toward a predetermined outcome, the human responds to unexpected AI output — letting surprises generate new creative directions.

❌ Responsive authorship: the human responds creatively to unexpected AI output, letting the AI's surprises generate new directions rather than directing toward a predetermined outcome.

Back to Lesson Start Lab 2

Lab 2

Emergent Narrative

Experience and analyze the authorship in AI-collaborative storytelling.

Lab 2 — Emergent Narrative

Experience and analyze emergent co-authorship.

The AI opens with a highly specific, unexpected story sentence from an unusual context.
Respond with a sentence, and keep building turn by turn for 3 exchanges.
After 3 exchanges, analyze: what decisions did you make that shaped the work's meaning? What would have been different with a different human?

Pay attention to the moments when you accepted AI's direction vs. redirected it. Those choices are where your authorship lives.

🎯 AI GuideLab 2

Back to Quiz Next: Lesson 3

Lesson 3

AI Makes Pictures and Music

Diffusion architecture, GANs, training data politics, and the commodification of style.

In 2023, artist Greg Rutkowski discovered that his name had become one of the most-used style prompts in Stable Diffusion — "in the style of Greg Rutkowski" appeared in millions of AI-generated images. Rutkowski, a living Polish fantasy artist, had not consented to his style being used as a training signal. "I'm afraid to search my own name," he told a reporter. The Stable Diffusion model had been trained on his work scraped from the internet without permission or payment.

Diffusion Architecture

Diffusion models work through a two-phase process: a forward process that progressively adds noise to training images until they become pure noise, and a reverse process trained to remove that noise step by step, guided by a conditioning signal (the text prompt). What the model learns is the mapping between text descriptions and visual statistical patterns.

CLIP embeddings: Text and images are mapped into a shared embedding space, allowing text prompts to guide image generation.
Latent diffusion: Modern models (Stable Diffusion) operate in a compressed latent space, making generation faster.
LoRA fine-tuning: Small model adaptations that inject a specific style or subject with minimal training data — enabling style theft at low cost.

The Commodification of Style

Style is not copyrightable under current law — you cannot copyright "impressionism" or "the style of Vermeer." But the ability to generate convincing approximations of a living artist's distinctive style at near-zero cost has commercial implications that existing copyright law was not designed to address.

The Rutkowski Problem

An artist's style is their commercial identity — the thing clients pay for. A tool that can generate "a Greg Rutkowski painting" on demand doesn't violate his copyright but does compete directly with his livelihood. Current law has no mechanism to address this.

Back Take Quiz

Quiz 3

AI Makes Pictures and Music

5 questions — free, untracked, retake anytime.

diffusion models, what does the 'reverse process' learn to do?

✓ Correct — ✅ The reverse process learns to remove noise step by step, guided by the conditioning signal (text prompt) — progressively refining noise into a coherent image.

❌ Reverse process: remove noise step by step, guided by the text conditioning signal, until a coherent image emerges from noise.

does CLIP do in text-to-image generation?

✓ Correct — ✅ CLIP creates a shared embedding space for text and images — enabling text prompts to guide image generation by finding the visual patterns that match the textual description.

❌ CLIP maps text and images into a shared embedding space — enabling the model to find visual patterns that match a text description.

does the Greg Rutkowski case represent a gap in current copyright law?

✓ Correct — ✅ Style isn't copyrightable. The model can generate 'a Rutkowski painting' on demand, competing directly with his commercial work, without technically violating any law. Current law has no mechanism for this.

❌ The gap: style isn't copyrightable, so generating 'in the style of' a living artist is currently legal — but it directly competes with their livelihood in ways current copyright law wasn't designed to address.

is LoRA fine-tuning in the context of image generation?

✓ Correct — ✅ LoRA fine-tuning: a small, efficient model adaptation that can inject a specific style or subject (including a living artist's distinctive style) with minimal training data — making style replication very cheap.

❌ LoRA: a small model adaptation that injects a specific style or subject with minimal training data. It makes replicating a specific artist's style fast, cheap, and easy.

phrase 'commodification of style' in AI art means:

✓ Correct — ✅ Commodification of style: what was previously a commercial differentiator (a distinctive artistic style built over years) can now be approximated by anyone with a text prompt, at near-zero cost.

❌ Commodification of style: a distinctive artistic identity that took years to develop can now be approximated by anyone with a prompt — eliminating its commercial scarcity value.

Back to Lesson Start Lab 3

Lab 3

Style, Consent, and Commerce

Develop a policy position on AI art and artists' rights.

Lab 3 — Style, Consent, and Commerce

Analyze the Rutkowski case and develop a policy position on AI art and living artists.

The AI opens with the tension: style isn't copyrightable, but style replication competes directly with an artist's livelihood. Should the law change?
Develop your policy position — what obligations should AI companies have toward living artists?
Address the LoRA problem specifically: targeted style replication from minimal data.

Consider: opt-in vs opt-out, compensation models, attribution requirements, and whether "style" should become legally protectable.

🎯 AI GuideLab 3

Back to Quiz Next: Lesson 4

Lesson 4

Is It Real?

Deepfake architecture, detection arms races, and the epistemology of synthetic media.

In 2024, a finance worker in Hong Kong transferred $25 million after attending a video conference with what appeared to be his company's CFO and other executives — all of whom were, in fact, AI-generated deepfakes. The attacker had used publicly available photos and video of the real executives to create convincing synthetic avatars. No technical detection was used by the victim; the deception was social and visual.

GAN Architecture and Deepfake Generation

Deepfake video typically uses GANs (Generative Adversarial Networks): a generator that produces synthetic images and a discriminator that tries to detect fakes. They train against each other — the generator improves by fooling the discriminator; the discriminator improves by catching the generator. After sufficient training, the generator produces output the discriminator (and often humans) cannot reliably distinguish from real.

Face-swap deepfakes: Replace one person's face in video with another's, matching lighting, angle, and expression.
Full synthesis: Generate a person who doesn't exist from scratch, or a real person performing actions they never performed.
Voice cloning: Synthesize a person's voice from minimal samples; combine with synthetic video for full-avatar deepfakes.

The Detection Arms Race

Deepfake detection tools have consistently lagged behind generation quality. Detection relies on artifacts — compression anomalies, facial inconsistencies, blinking patterns — that disappear as generation improves. Researchers have moved toward content provenance as a more durable approach: cryptographic signatures embedded at creation that verify authenticity, rather than trying to detect forgery after the fact.

C2PA Standard

The Coalition for Content Provenance and Authenticity (C2PA) is developing an open standard for embedding cryptographic provenance information into media files at creation — a kind of digital chain of custody. Adopted by Adobe, Microsoft, and major camera manufacturers. The challenge: forged media doesn't carry this signature, so absence of provenance becomes suspicious but not definitive.

Back Take Quiz

Quiz 4

Is It Real?

5 questions — free, untracked, retake anytime.

does GAN training produce increasingly convincing deepfakes?

✓ Correct — ✅ Adversarial training: generator improves by fooling the discriminator; discriminator improves by catching the generator. After sufficient iterations, the generator produces output that humans can't reliably distinguish from real.

❌ GANs: generator tries to fool discriminator, discriminator tries to catch generator. They improve against each other until the generator produces output indistinguishable from real.

do deepfake detection tools consistently lag behind generation quality?

✓ Correct — ✅ Detection relies on artifacts (inconsistencies, anomalies) that improve as generation quality improves. It's a perpetual arms race with the generator always one step ahead.

❌ Detection relies on artifacts that disappear as generation improves. Every advance in detection is followed by advances in generation that eliminate those artifacts.

is the C2PA provenance approach, and why is it more durable than detection?

✓ Correct — ✅ C2PA embeds cryptographic provenance at creation — a chain of custody that travels with the file. Rather than detecting fakes, it verifies real content. More durable because it doesn't depend on detecting artifacts.

❌ C2PA: cryptographic signatures at creation time that travel with the file. Verifies authentic content rather than detecting forgery — more durable because it doesn't rely on artifact detection.

the Hong Kong $25M deepfake case, what made the attack successful?

✓ Correct — ✅ The attack succeeded socially, not technically: the victim trusted what appeared to be familiar colleagues in a familiar context. No technical verification was attempted.

❌ The attack succeeded through social engineering — exploiting trust in familiar faces and a familiar context. Technical detection was never attempted.

does 'absence of C2PA provenance' not definitively prove a media file is synthetic?

✓ Correct — ✅ Provenance absence is suspicious but not definitive: large amounts of legitimate older media simply predate C2PA adoption and also lack signatures.

❌ Many legitimate older files predate C2PA adoption and have no signature. Missing provenance means 'possibly synthetic' — not 'definitely synthetic.'

Back to Lesson Start Lab 4

Lab 4

Deepfake Architecture and Policy

Develop technical and policy responses to synthetic media threats.

Lab 4 — Deepfake Architecture and Policy

Analyze the deepfake problem from both technical and policy perspectives.

The AI opens with the Hong Kong case and asks what systematic safeguards would have prevented it.
Develop both technical and procedural safeguards against deepfake attacks.
Address C2PA: is cryptographic provenance sufficient, and what are its limits?

Consider: technical detection limitations, social engineering vulnerabilities, provenance standards, and organizational protocols.

🎯 AI GuideLab 4

Back to Quiz Next: Lesson 5

Lesson 5

Who Owns the Story?

Copyright doctrine, fair use, authorship theory, and the legal frontier of AI creative output.

The legal landscape in 2023-24: The US Copyright Office published guidance stating that AI-generated content lacks human authorship and cannot be copyrighted, while human creative selection and arrangement of AI output can be. The Authors Guild sued OpenAI. Getty Images sued Stability AI. Sarah Silverman sued Meta and OpenAI. The New York Times sued OpenAI and Microsoft, alleging that verbatim memorization of their content constituted infringement — producing in discovery evidence that the model could reproduce NYT articles verbatim when prompted. No cases had reached final verdict as of early 2025.

Copyright Doctrine and AI

Three distinct legal questions have emerged:

Output copyrightability: Can AI-generated content be copyrighted? Current US position: not without meaningful human authorship. The "minimal human authorship" threshold is undefined.
Training data as infringement: Does training on copyrighted works constitute infringement? The fair use defense (transformativeness, market effect) is being tested. The NYT case introduced evidence of verbatim memorization — potentially undermining the transformativeness argument.
Style vs. expression: Style is not copyrightable. Specific creative expression is. "In the style of Hemingway" is legal; reproducing Hemingway's actual sentences is not.

Authorship Theory

Copyright law was built on Romantic authorship theory — the idea of a singular creative genius producing original work. AI challenges this at every level: training (no single author), output (no human author), style (drawn from many humans). Legal philosopher Shyamkrishna Balganesh has argued that AI output requires a new authorship category that can accommodate distributed, human-machine creative processes.

The Unresolved Question

If a human writes detailed prompts, curates AI output extensively, edits and arranges the results — how much human creative input is sufficient for copyright protection? The Copyright Office has not answered this question, and the courts haven't either.

Back Take Quiz

Quiz 5

Who Owns the Story?

5 questions — free, untracked, retake anytime.

did NYT v. OpenAI introduce as evidence that potentially undermines the 'transformativeness' fair use argument?

✓ Correct — ✅ Verbatim memorization of protected content undermines transformativeness — if the model can reproduce articles word for word, it arguably did not transform them.

❌ Verbatim reproduction on demand undermines transformativeness: if the model memorized and can reproduce the original expression, it's harder to argue the training 'transformed' the work.

current US Copyright Office guidance, what determines whether AI-assisted work can be copyrighted?

✓ Correct — ✅ Current guidance: human creative authorship determines copyrightability. Pure AI output is not protectable; work with sufficient human creative contribution may be. The threshold for 'sufficient' is undefined.

❌ Current US position: meaningful human creative authorship determines copyrightability. Pure AI output isn't protectable. The threshold for 'sufficient' human contribution is currently undefined.

does 'style is not copyrightable' create a gap in legal protection for AI and living artists?

✓ Correct — ✅ Style isn't protected, so generating 'in the style of' a living artist is currently legal — even when it directly competes with their commercial work. Existing copyright law wasn't designed for this.

❌ Style = not copyrightable. AI can generate 'in the style of' any living artist, competing directly with their commercial work, without infringing any current legal protection.

does Shyamkrishna Balganesh's argument about AI authorship propose?

✓ Correct — ✅ Balganesh argues that Romantic authorship theory — singular human genius — can't accommodate AI. A new authorship category is needed for distributed human-machine creative processes.

❌ Balganesh: copyright was built on Romantic authorship (singular human genius) and can't accommodate AI. A new legal category is needed for distributed human-machine creativity.

'fair use' defense for AI training on copyrighted works primarily rests on which argument?

✓ Correct — ✅ The primary fair use argument: training is transformative — it produces a new kind of tool, not a reproduction of the original works. The NYT case challenged this by showing verbatim memorization.

❌ The fair use defense for training primarily rests on transformativeness — training produces a new tool, not a reproduction. The NYT verbatim memorization evidence challenges this argument.

Back to Lesson Start Lab 5

Lab 5

Copyright Frontier

Develop a policy framework for AI copyright and training data rights.

Lab 5 — Copyright Frontier

Analyze the unresolved legal questions and develop a policy framework.

The AI opens with the verbatim memorization evidence from NYT v. OpenAI and asks whether it changes your view of training data fair use.
Develop your framework for training data rights and AI output copyrightability.
Address the Balganesh proposal: should there be a new authorship category for human-AI work?

Consider: transformativeness, market substitution, the minimal human authorship threshold, and opt-in vs opt-out frameworks.

🎯 AI GuideLab 5

Back to Quiz Next: Lesson 6

Lesson 6

AI and Human Creativity Together

Augmentation theory, creative labor markets, and the political economy of AI creativity.

McKinsey's 2023 economic analysis estimated that generative AI could automate 60-70% of the tasks currently performed by knowledge workers. For creative workers specifically, the pattern was asymmetric: senior creative directors and art directors saw their productivity multiply; junior illustrators, stock photo contributors, entry-level copywriters, and session musicians saw their markets shrink rapidly. The same technology augmented at the top and replaced at the bottom of the creative labor market simultaneously.

The Asymmetry of Creative Augmentation

AI creative tools do not affect all creative workers equally. The pattern that has emerged:

Concept-level workers benefit: Those whose value lies in direction, taste, and creative judgment can now execute faster. AI handles execution; humans handle meaning.
Execution-level workers are displaced: Those whose value was in the ability to execute at a certain quality level (illustrate, photograph, write copy) now compete with AI that executes at comparable quality at near-zero cost.
The learning pipeline breaks: Junior creative roles — historically where people built execution skills before moving to concept roles — are the ones being eliminated first.

The Learning Pipeline Problem

The junior roles being eliminated are not just jobs — they are training grounds. Illustrators learn by doing thousands of illustrations; copywriters learn by writing thousands of pieces of copy. If AI eliminates the market for junior-level execution work, it potentially breaks the pipeline that produces future senior creatives.

Structural Question

If AI eliminates the entry-level roles that train future concept-level creative directors, who does the concept-level work in ten years? The augmentation argument assumes a pipeline that AI may be disrupting.

Back Take Quiz

Quiz 6

AI and Human Creativity Together

5 questions — free, untracked, retake anytime.

was the asymmetric pattern of AI creative tools on the labor market in 2023?

✓ Correct — ✅ The asymmetry: AI augments concept-level work (direction, taste, judgment) and replaces execution-level work (illustration, copywriting, photography). Benefits flowed up the hierarchy; displacement flowed down.

❌ The asymmetry: AI augmented concept-level workers (who direct and curate) while displacing execution-level workers (who implement). Same technology, opposite effects at different career levels.

does the elimination of junior creative roles create a structural risk beyond just job loss?

✓ Correct — ✅ Junior roles are not just jobs — they're how people develop the execution skills that eventually become concept skills. Eliminating them may break the pipeline that produces future creative leaders.

❌ Junior roles = training ground for future senior creatives. Illustration, copywriting, photography skills develop through practice. Eliminating those roles may break the pipeline from execution to concept work.

'concept-level workers benefit' argument assumes:

✓ Correct — ✅ The augmentation argument assumes a pipeline of skilled concept-level workers directing AI. But if AI eliminates the junior roles that develop those skills, the pipeline dries up.

❌ The augmentation argument assumes there will be skilled concept-level workers to direct AI. But if the junior training ground disappears, who develops those skills?

does McKinsey's 60-70% task automation estimate mean for creative workers specifically?

✓ Correct — ✅ The estimate is about tasks, not workers — and impact is uneven: concept-level tasks are less automatable than execution tasks, so senior creatives are less exposed than junior ones.

❌ The estimate covers tasks, not workers, and the impact is asymmetric: execution tasks are more automatable than concept tasks, so junior execution workers are more exposed than senior concept workers.

does it mean to say AI 'breaks the learning pipeline' in creative work?

✓ Correct — ✅ The learning pipeline: junior execution work → skill development → senior concept work. AI eliminating junior execution roles removes the on-ramp to creative mastery.

❌ The learning pipeline: junior execution roles → practice → skill development → senior concept work. AI eliminating those junior roles removes the path that develops creative mastery.

Back to Lesson Start Lab 6

Lab 6

Creative Labor Economics

Analyze AI's asymmetric impact on creative labor markets.

Lab 6 — Creative Labor Economics

Analyze the political economy of AI creative tools and develop a policy response.

The AI opens with the learning pipeline problem: if AI eliminates junior roles, who trains the next generation of concept-level creatives?
Develop your analysis of the asymmetric augmentation pattern.
Address: what policy responses (if any) would you advocate for — and who should bear the costs?

Consider: training levies, opt-in licensing, universal basic income, creative apprenticeship programs, disclosure requirements.

🎯 AI GuideLab 6

Back to Quiz Next: Lesson 7

Lesson 7

The Ethics of AI Art

Consent, training data, artists' rights, and the moral philosophy of AI creativity.

Sarah Andersen, Kelly McKernan, and Karla Ortiz — three artists whose styles had been heavily used as prompts in AI image generation — filed a class action lawsuit against Stability AI, Midjourney, and DeviantArt in 2023. Their attorney argued that the models had been trained on their work without consent, and that the ability to generate images "in their style" constituted a form of theft. The defendants argued that training is transformative and that style isn't copyrightable. The case raised questions that copyright law wasn't designed to answer.

Consent and Training Data

AI image models were trained on billions of images scraped from the internet — including work from living artists who had made their work publicly accessible for viewing, not for AI training. The ethical question precedes the legal one: even if scraping is legal, is it ethical to use an artist's creative output to build a commercial tool that competes with them?

Opt-in vs opt-out frameworks: Some researchers argue for opt-in consent (artists must affirmatively agree to have their work included). Current practice is effectively opt-out, with imperfect tools for removal.
Do Not Train registries: HaveIBeenTrained.com and ArtStation's opt-out tools offer partial, unverified protection.
Compensation models: Adobe Firefly paid contributors; Getty Images licensed its library. These are the minority.

The Moral Philosophy Question

Beyond copyright: does training on an artist's work without consent violate something that copyright law doesn't protect? Philosopher Nils-Hennes Stear argues that there is a moral right to creative attribution and creative integrity that is distinct from copyright — and that using someone's creative work to train a competitive tool violates both, regardless of legal status.

The Consent Principle

Making creative work publicly accessible does not constitute consent to have it used for every downstream purpose. The same principle applies to data: sharing data publicly is not the same as consenting to that data being used to train a commercial AI competitor.

Back Take Quiz

Quiz 7

The Ethics of AI Art

5 questions — free, untracked, retake anytime.

is the central ethical argument of artists suing AI image companies, beyond copyright?

✓ Correct — ✅ The ethical claim beyond copyright: public accessibility ≠ consent to training. Using publicly available work to build a tool that competes commercially with the creator violates something beyond what copyright addresses.

❌ The ethical claim: public accessibility is not the same as consent to training. Using creative work to build a commercial competitor violates something beyond what copyright law addresses.

does an 'opt-in' framework for AI training data mean?

✓ Correct — ✅ Opt-in: the default is exclusion from training. Artists affirmatively agree to be included. Current practice is effectively the opposite: scraped by default, opt-out with imperfect tools.

❌ Opt-in: default = excluded. Artists must affirmatively consent to inclusion. Current practice is effectively opt-out (scraped by default, removal tools available but imperfect).

distinguishes the Adobe Firefly and Getty Images approach from standard practice?

✓ Correct — ✅ Adobe Firefly paid contributors; Getty licensed its library. Both obtained consent and provided compensation — in contrast to the scrape-by-default approach of most models.

❌ Adobe Firefly and Getty both obtained consent from and compensated contributors — a sharp contrast to scrape-by-default standard practice.

Nils-Hennes Stear argues that training on artists' work without consent violates:

✓ Correct — ✅ Stear argues there's a moral right to creative attribution and integrity — something copyright doesn't fully protect — that's violated when work is used to train a competitive tool without consent.

❌ Stear: there's a moral right to creative attribution and integrity that's distinct from and goes beyond copyright. Using work to train a competitor violates both, regardless of legal status.

is the key principle regarding public accessibility and consent to AI training?

✓ Correct — ✅ Public accessibility ≠ consent. Sharing work for public viewing is not the same as consenting to that work being used to train a commercial AI that competes with the creator.

❌ Public accessibility and consent are separate. Making work viewable publicly does not constitute consent to use it for AI training — especially for a commercial tool that competes with the creator.

Back to Lesson Start Lab 7

Lab 7

Consent and Creative Rights

Develop an ethical framework for AI art and training data.

Lab 7 — Consent and Creative Rights

Develop your ethical and policy framework for AI training data and artists' rights.

The AI opens with the consent principle: public accessibility ≠ consent to training. Do you agree?
Develop your position on opt-in vs opt-out frameworks and compensation models.
Address: is there a moral right to creative attribution and integrity beyond copyright that AI violates?

Consider: what would a genuinely ethical AI art ecosystem look like? What would it cost, and who would bear it?

🎯 AI GuideLab 7

Back to Quiz Next: Lesson 8

Lesson 8

Building Your Own AI Story

Practical prompting, iterative creation, and building a human-AI creative practice.

Author Robin Sloan has written publicly about using language models in his creative process since 2021. His practice: generate large amounts of raw material (thousands of words of rough prose, alternative scenes, character dialogue in different registers), then curate aggressively, extracting the phrases and images that feel genuinely surprising or useful. "I'm not looking for the AI to write my novel," he said in an interview. "I'm looking for it to show me things I wouldn't have thought of, so I can decide if they're interesting."

Building a Prompting Practice

The most productive AI-assisted creative practices share common features:

Volume generation: Generate significantly more than you need. AI is cheap to run; aggressive curation is the skill.
Constraint injection: Specify form constraints, voice constraints, structural constraints. These force the model off its statistical defaults and toward the unexpected.
Context management: For long-form work, maintain an external world-state document — characters, timeline, established facts — and inject it into prompts to compensate for recency bias.
Voice protection: Generate, then rewrite. Let AI produce the raw material, then bring your voice to it in editing. Never publish unedited AI output as your creative work.

Iterative Creation Protocol

A practical framework for AI-assisted narrative creation:

Phase 1 — Direction: Write your core premise, character sketches, and thematic intention without AI.
Phase 2 — Generation: Use AI to generate rough prose, alternative scenes, unexpected details. Generate 5-10x what you need.
Phase 3 — Curation: Extract what's genuinely surprising or useful. Discard the statistical average.
Phase 4 — Voice: Rewrite everything in your voice. Use AI output as raw material, not finished product.

The Core Principle

AI is a raw material generator. You are the author. The work is yours when you've provided the direction, made the curatorial decisions, and written the voice. That's what authorship has always meant.

Back Take Quiz

Quiz 8

Building Your Own AI Story

5 questions — free, untracked, retake anytime.

does Robin Sloan use AI for in his creative process?

✓ Correct — ✅ Sloan uses AI as a raw material generator, not a drafter. He looks for what's surprising and unexpected, then decides if it's interesting — curation is the skill.

❌ Sloan: generate large amounts of raw material, curate aggressively for what's genuinely surprising. AI shows him things he wouldn't have thought of; he decides if they're interesting.

does generating 5-10x more than you need improve AI-assisted creative work?

✓ Correct — ✅ Volume + aggressive curation: by generating far more than you need, you can afford to discard everything that feels like the statistical average and keep only what's genuinely distinctive.

❌ Volume generation enables aggressive curation: generate far more than you need, then discard the statistical average and keep only what's genuinely surprising or useful.

is the purpose of an external world-state document in AI-assisted long-form narrative?

✓ Correct — ✅ World-state documentation compensates for recency bias: inject established characters, timeline, and facts into each prompt to prevent the model from forgetting or contradicting earlier content.

❌ External world-state compensates for recency bias: inject established characters, facts, and timeline into each prompt so the model doesn't forget or contradict earlier content.

'voice protection' principle in AI-assisted creative work means:

✓ Correct — ✅ Voice protection: AI generates raw material; you bring your voice in editing. Never publish unedited AI output. The rewriting is where your authorship lives.

❌ Voice protection: let AI generate rough material, then rewrite everything in your voice. The rewriting is the authorship. Never publish unedited AI output as your creative work.

the iterative creation protocol, what is the function of Phase 1 (Direction)?

✓ Correct — ✅ Phase 1 happens without AI: you establish your premise, characters, and thematic intention. Human intent drives the work before AI generates anything.

❌ Phase 1 (without AI): establish your core premise, character sketches, and thematic intention. Human direction precedes AI generation — so the work's meaning is yours from the start.

Back to Lesson Start Lab 8

Lab 8

Building Your Creative Practice

Develop your personal AI-assisted creative framework.

Lab 8 — Building Your Creative Practice

Develop your personal framework for AI-assisted creative work.

The AI opens with Sloan's framework and asks what you'd adapt or change for your own creative practice.
Develop your four-phase protocol for a specific type of creative work you do or want to do.
Address: how do you protect your creative voice while using AI extensively as a raw material generator?

Be specific about the type of creative work. Vague frameworks produce vague practices.

🎯 AI GuideLab 8

Back to Quiz Module Test

Module 4 Test

8 questions covering all lessons. Free, untracked, retake anytime.

structural property causes AI narrative to lose coherence over long distances?

✓ Correct — ✅ Recency bias: the model weights recent context more than distant context, causing earlier established characters, facts, and plot logic to fade and producing continuity errors.

❌ Recency bias: the model attends more strongly to recent context, causing earlier established content to fade — producing continuity errors and long-range incoherence.

Ross Goodwin's '1 the Road', authorship was:

✓ Correct — ✅ Authorship was genuinely distributed: Goodwin designed, drove, and curated; the AI generated; the training data supplied the aesthetic vocabulary. No traditional category fully fits.

❌ Distributed authorship: Goodwin (system design, driving, curation) + AI (text generation) + training data (aesthetic vocabulary). None of the traditional categories fully account for the work.

fine-tuning enables:

✓ Correct — ✅ LoRA: a small, efficient model adaptation that can inject a specific style or subject (including a living artist's style) with minimal training data — making style replication cheap and targeted.

❌ LoRA: efficient injection of a specific style or subject into a model with minimal training data — enabling targeted, cheap replication of a living artist's distinctive style.

C2PA provenance standard addresses synthetic media by:

✓ Correct — ✅ C2PA: cryptographic signatures at creation time. Verifies genuine content rather than detecting forgery — more durable because it doesn't rely on artifact detection that improves away.

❌ C2PA embeds cryptographic provenance at creation — a chain of custody that verifies authentic content rather than trying to detect forgery after the fact.

learning pipeline problem in AI creative labor means:

✓ Correct — ✅ Junior execution roles = the practice ground for developing creative mastery. Eliminating them may break the pipeline from entry-level execution to senior concept work.

❌ Junior execution roles are where creative mastery develops. AI eliminating those roles may break the pipeline that produces future senior creative directors.

ethical principle 'public accessibility is not consent to training' means:

✓ Correct — ✅ Public accessibility ≠ consent to training. Sharing work for public viewing is not the same as consenting to that work being used to train a commercial AI competitor.

❌ Public accessibility and consent are separate. Viewing permission is not training permission — especially for a commercial system that competes with the creator.

the iterative creation protocol, Phase 4 (Voice) is:

✓ Correct — ✅ Phase 4: rewrite AI raw material in your voice. This is where your authorship lives — direction (P1) and voice (P4) are the human phases; generation and curation are where AI contributes.

❌ Phase 4 = rewrite in your voice. AI generates raw material; you bring your voice in the rewriting. That's where your authorship is most fully expressed.

NYT verbatim memorization evidence in its lawsuit against OpenAI was significant because:

✓ Correct — ✅ Verbatim reproduction of protected expression on demand weakens the transformativeness argument: if training produced a model that can reproduce the original works, it's harder to argue training transformed them.

❌ Verbatim memorization evidence undermines transformativeness: if the model can reproduce articles word-for-word, training arguably didn't transform the work — it preserved it for reproduction.

Back Academy Home