Teaching "AI in Creativity" to Architect mode (ages 15–18)

Hook & Warm-Up

Open with a direct framing question, aimed at students who may be weighing a creative or technical career path: "If generative AI can already produce publishable-quality images, listenable songs, and passable prose, what's actually left that's distinctly human about creative work — and is that changing, or is it permanent?"

"AI can now create art, music, video, and code. Every creative industry is being reshaped. Let's examine the technology and ethics of AI creativity."

That's the lesson's own hook — treat it as an agenda, not just an opener. Follow with the concrete legal anchor for the whole discussion:

"In 2025, a US federal appeals court ruled on Thaler v. Perlmutter: a work generated entirely by an AI system, with no human author, cannot receive copyright protection under US law. That's a real, decided case, not a hypothetical. But it only answers the easy version of the question — pure AI output, zero human involvement. It says almost nothing about the much more common real-world case: a human using AI as one tool among several in a creative process. That gap is where almost all of the live legal and economic conflict actually sits, and it's where we're going to spend most of today."

Add the local frame immediately, since this is a Philippine classroom, not a US one:

"That's US law. Copyright here is governed by Republic Act No. 8293, the Intellectual Property Code of the Philippines, administered by IPOPHL. It was written in 1997 — years before anything like Midjourney or Suno existed. RA 8293 defines a protectable work as an 'original intellectual creation,' language that, like the US law behind Thaler, points toward requiring a human author. But no Philippine court or IPOPHL ruling has yet tested that language against an actual AI-generated work the way Thaler v. Perlmutter did in the US. So unlike the American question, which now has at least a partial answer, the Philippine question is still completely open."

Ask directly: "If you were the first person to bring this question to a Philippine court or to IPOPHL, what argument would you make, and what do you think they should decide?" This positions the lesson as an analysis of an active, unresolved legal and economic question with direct career relevance — one where the Philippines specifically hasn't caught up yet — not a tour of impressive demos.

Main Activity

The in-app lesson covers three scenes — Visual Art, Music & Audio, and Writing & Film. For this age band, use each scene as a case study in a specific generation technique and the specific legal or economic conflict around it.

Scene 1 — Visual Art: diffusion models and the training-data dispute

• AI Paintings 🖼️ (Midjourney, DALL-E) — diffusion-based text-to-image generation.
• Photo Enhancement 📸 — AI upscaling, background removal, colorization of archival photos.
• Style Transfer 🎭 — separating and recombining content and style representations.
• Oil Painting 🖌️ (not AI) — requires embodied, years-long skill development.

Give the mechanism at a genuine technical level: a diffusion model is trained with a denoising objective — real images are progressively corrupted with Gaussian noise across many timesteps, and the model learns to predict and subtract that noise at each step, conditioned on an accompanying text caption via a shared embedding space (typically produced by a model like CLIP, which learns to associate images and text descriptions). Generation runs this process in reverse: starting from pure noise, the model iteratively denoises, at each step conditioned on the target text prompt, until a coherent image emerges. This is genuine generative modeling — the model is sampling from a learned probability distribution over images, not retrieving or splicing training examples — but that same fact is exactly what's at the center of the legal dispute: training that distribution required feeding the model enormous quantities of existing copyrighted images, almost never with the original artists' consent or compensation, and several major lawsuits (against companies including Stability AI and Midjourney, brought by visual artists and stock-image companies) are actively litigating whether that training process itself constitutes infringement, separate from any question about the copyrightability of the output.

Push further on what "training data" actually contains, since it's usually treated as an abstract legal abstraction rather than a concrete one with a geography. Large-scale image datasets like LAION, which underlie tools such as Stable Diffusion, are scraped overwhelmingly from image-heavy, already-digitized corners of the internet — which is not remotely representative of art or visual culture worldwide. Ask students to reason through a specific, real case rather than a hypothetical: a diffusion model prompted for "T'nalak weaving pattern," "Ifugao woodcarving," or "traditional Filipino komiks art" has almost certainly seen far fewer training examples of these than of, say, Japanese ukiyo-e or European Impressionism, both extensively digitized by museums, archives, and stock-image sites for decades. The practical result is a model that either can't render a distinctly Filipino visual tradition with any real accuracy, or worse, defaults to a generic, flattened "tribal" or "Asian" aesthetic that erases the specific, meaningful differences between, say, a T'boli textile motif and an Ifugao bulul carving. Ask: "If a tool's output quality depends entirely on how much of a given culture's art has already been digitized and put online, what does that predict about which artistic traditions AI will represent well, and which it will misrepresent or flatten together — and is that a technical limitation, or a cultural one that technical fixes alone won't solve?"

Scene 2 — Music & Audio: neural audio codecs and identity

• AI Music 🎵 (Suno, Udio) — full song composition, including vocals, from text.
• AI Mastering 🎧 — automated track enhancement, replacing work that cost roughly $1,000 per track.
• Voice Cloning 🎤 — voice replication from roughly 3 seconds of audio.
• Guitar Playing 🎸 (not AI) — embodied skill AI can simulate the acoustic result of, but not perform.

Explain neural audio codecs, since they're the specific technology that made systems like Suno and Udio possible: a codec like Meta's EnCodec compresses raw audio into a sequence of discrete tokens, in a way conceptually similar to how text is broken into tokens for a language model. Once audio can be represented as a token sequence, the same autoregressive, next-token-prediction approach that works for text generation can be applied to audio — this is what lets modern music-generation systems be trained and operated with techniques borrowed directly from large language models, rather than needing an entirely separate generative architecture for sound.

Then connect this to voice cloning as a distinct application of similar underlying representations: a small amount of reference audio (the lesson's figure of about 3 seconds is realistic for current commercial tools) is enough to condition a voice-generation model on a specific speaker's vocal characteristics, producing new speech in that voice. Frame the ethical stakes precisely: this isn't a generic "AI is impressive" concern, it's an identity and authentication concern — voice has historically functioned as an informal proof of identity (a phone call from a family member "sounding like them" used to be reasonably trustworthy evidence it was them), and that assumption no longer safely holds.

Scene 3 — Writing & Film: language models and the economics of generation

• AI Writing 📝 — articles, poetry, code, screenplays; struggles with truly original ideas.
• AI Video 🎬 (Sora) — text-to-video generation of scenes that never existed.
• Game Design AI 🎮 — procedural generation of levels, character designs, and storylines.
• Novel Writing 📖 (not AI) — grounded in lived human experience.

Frame "struggles with truly original ideas" precisely: a language model is trained to predict statistically likely continuations of text given its training distribution, which structurally biases outputs toward the statistically central tendencies of that distribution rather than toward its tails — where genuine novelty tends to live. This is a property of the training objective itself, not simply a current limitation expected to disappear with more scale, though how much it matters in practice for a given creative task is a legitimate open question.

Close the activity by having students map each of the twelve items across all three scenes onto one of: fully generative, human-and-AI-assisted, or purely human — and identify, for each "assisted" item, exactly which economic role is being displaced or changed (e.g., AI mastering doesn't replace the musician, it replaces the professional audio engineer's post-production labor).

Push one level further with Game Design AI as a case study in a more constrained generative problem: procedural level and character generation typically operates under hard functional constraints (a level must be completable, difficulty must scale, assets must fit a technical budget), which is a meaningfully different engineering problem from open-ended text or image generation with no correctness criterion beyond "does this look or read well." Ask students to articulate why a studio might trust AI generation for this narrower, more verifiable task well before trusting it with a game's core narrative — a useful frame for thinking about where in a creative pipeline AI assistance is currently most defensible.

Discussion

These questions are live policy and legal debates, not settled textbook material — treat disagreement among students as the expected, correct outcome, and press each position for its practical consequences.

• "Thaler v. Perlmutter denies copyright to fully AI-authored work but doesn't define how much human creative input is enough for a mixed work to qualify. If you were writing that legal standard, what test would you propose, and what problems would your test create?"
• "Training a diffusion model on copyrighted art without consent, versus a human artist studying other artists' work without paying them — are these meaningfully different in a way the law should treat differently? Make the strongest case for each side."
• "If a diffusion model has seen a million paintings in a European style and a few hundred examples of a specific Filipino artistic tradition, is its 'Filipino art' output actually representing that tradition, or producing a flattened guess shaped mostly by what it saw more of? Whose responsibility is that — the dataset's, the model builder's, nobody's?"
• "RA 8293 predates generative AI by almost three decades and has never been tested against an AI-authorship question in the Philippines. Should the law be amended to explicitly address AI-generated work, or is applying its existing 'original intellectual creation' language by analogy enough? If you were advising a Philippine legislator, what would you propose?"
• "Neural audio codecs let audio generation borrow language-model techniques. What does it tell you about the underlying nature of these generative systems that a technique built for text turned out to transfer to music?"
• "The economic impact of generative AI on creative professionals — musicians, illustrators, writers, editors — is real and already underway. Who do you think should bear responsibility for managing that transition: the AI companies, government policy, the creative industries themselves, or some combination? Defend your allocation."
• "Opt-out frameworks — like the EU's text-and-data-mining exception, which lets rights holders formally reserve their work from AI training — put the burden on creators to actively opt out, rather than on AI companies to get opt-in consent first. Is that the right default? What would change if it were reversed?"
• "If you were advising a musician, illustrator, or writer just starting their career today, what would you tell them about how to position their work and their skills given how fast generative AI is improving in their field?"

Quiz Walkthrough

Wrap-Up & Extension

Close with: "Every technique we looked at today — diffusion models, neural audio codecs, language models applied to text and video — is real, working technology that's already reshaping creative industries, not a future hypothetical. But the legal framework around ownership, consent, and compensation is still being actively litigated and written, case by case, right now. If any of you go into a creative field, a technical field building these tools, or law, you are very likely to be working inside rules that are still being decided during your career, not rules that were already settled before you got there. That's worth taking seriously either way you plan to sit relative to this technology — building it, regulated by it, or competing with it."

Extension activity: Assign each student one specific, real, ongoing legal case or policy development related to generative AI and creative work (examples to research: an artists'-rights or authors' lawsuit against an AI company, a specific country or the EU's AI-and-copyright policy approach, a documented voice-cloning fraud case and its legal outcome, or how RA 8293 and IPOPHL might handle an AI-authorship question in the Philippines, since no case has tested this yet). Each student writes a one-page brief covering: the core legal or ethical question at stake, the strongest argument on each side, and the current status of the case or policy as of their research. Compile briefs into a shared class reference document — this both reinforces that these are live, evolving issues and gives the class a genuinely useful resource to revisit.