What LLMs and image models actually do; the four reliable failure modes.
All tracks · 4 hours · drop-in friendly
AI Without the Ick: Literacy, Integrity, and Creative Control
You leave today knowing what AI actually is, where it reliably fails, and with your own first-draft rules for using it without losing your voice.
01This Week at a Glance
You will make: an AI Boundary Map + your Personal AI Use Manifesto v1 (8–12 bullets).
Consent, copyright, labor, slop, disclosure — taken seriously. Then: friction vs. meaning.
One task, three ways: weak prompt → creative brief → critique loop.
Boundary Map → two-tool comparison → Manifesto v1 → critique.
Why this matters in real work
AI is already inside studios, agencies, and client expectations. You don't have to like it — but there's a difference between refusing from knowledge and refusing from fear. This course builds the knowledge; what you do with it is your artistic decision.
02Key Concepts
1 · An LLM is a prediction machine, not a knowledge machine
The model inside ChatGPT/Claude does one job: predict what plausibly comes next, one token at a time. That produces useful writing, critique, and code — but it optimizes for plausible, not true, and sounds equally confident either way.
In real work
Treat it like a brilliant, fast, overconfident intern: great at drafts, options, structure. Never the final authority on facts, sources, or specs.
2 · Image models don't collage — they synthesize (and that cuts both ways)
Diffusion models turn noise into images, steered by your words. They don't store or paste anyone's image — and they were trained on millions of artists' works, mostly scraped without consent. Both facts are true. The mechanism being non-copying doesn't settle the consent question.
Watch out
The output is statistically average by default. "Make it cyberpunk" returns the mean of every cyberpunk image online. Specific direction is the only escape from average.
3 · The four reliable failure modes
| Failure | Looks like | Your counter-move |
|---|---|---|
| Hallucination | Invented facts, fake sources, artists who don't exist | Verify every name, date, quote, citation |
| Genericness | Competent, soulless, like everything else | Constraints, references, your taste |
| Bias | Default faces, default beauty, default art history | Notice the defaults; direct against them |
| Amnesia | Forgets your brief mid-chat (context window full) | Restate constraints; fresh chats |
4 · The artist's objections — taken seriously
- Consent: training data was largely scraped without permission. True. Tools differ (Firefly trains on licensed content) — tool choice is an ethical act.
- Copyright: purely AI-generated output generally can't be copyrighted — human authorship is the legal test. Your judgment isn't just ethics; it's ownership.
- Labor: some tasks are getting automated. The defensible ground: taste, judgment, accountability, relationships — this course's syllabus.
- Slop: generic is now free, so distinctive human work gains value.
- Disclosure: the EU is making synthetic-media transparency law. Honesty is becoming a hard skill.
5 · The one rule
The course in one line
Use AI for friction. Keep meaning human. Friction = the parts that aren't your art (summaries, options, checklists, file names). Meaning = concept, taste, selection, ethics, final craft, your name on it. Where the line sits is yours to draw — that's today's lab.
03This Week's Words
LLMText-prediction model behind ChatGPT/Claude. Plausibility machine, not truth machine.
TokenThe word-chunk unit models read and write. Limits and pricing count in these.
Context windowThe model's working memory. Outgrow it and it forgets your brief.
HallucinationConfident invention of facts, sources, or artists. Verify before trusting.
Multimodal modelHandles images/audio too — why you can screenshot work and ask for critique.
Diffusion modelImage/video generator: noise → image, steered by your prompt.
Training dataWhat the model learned from. The ethical centre of the artist debate.
PromptEverything you give the model. Output ceiling = brief quality.
Model biasDefaults inherited from data. Counter with specific direction.
SlopLow-effort AI content at volume. What your work must visibly not be.
Synthetic mediaAny AI-generated image/video/audio. EU labelling law is phasing in.
DisclosureSaying specifically what AI did. Specific beats vague, always.
04Demo: One Task, Three Ways
I made this, so can you
The instructor runs the same task three ways. Watch what changes — it isn't the model.
Prepared demo artifact
Instructor backup / finished example: comparison-board.html. Use it if the live tool run fails, or to reveal the finished structure after critique.
- Weak prompt. "Make a poster concept for an art school degree show." → fast, confident, generic. Could be any college, anywhere.
- Creative brief. Same ask + audience, tone, constraints (2-colour riso, typography-led), references, and the line that matters: "Don't pick a winner — that's my job." → five usable directions with risks.
- Critique loop. "Direction 3 is weakest. Critique it against my brief, then give two sharper variants." → the model attacks its own genericness, steered by human taste.
What to notice
The human supplied audience, constraints, taste, and the final call. The model supplied speed and volume. That division of labor is the whole course. Adapt it: swap in any project, keep the structure.
05Step-by-Step Workflows
Your first real AI conversation (the non-sucky version)
- Give it a role and a job. Not "hi." Who to be, what you're working on.
- Paste context. Project, audience, constraints. Nothing private, nothing that isn't yours to share.
- Ask for options, not answers. "Five directions with risks" beats "what should I do."
- Make it ask first. "Ask clarifying questions before answering" — instantly better output.
- React with specifics. "No. 2 is closest but the tone is too corporate" beats "make it better."
- Keep your decision. Choose, reject, refine yourself — and note why. That's process documentation.
The 60-second sanity check (before any AI use)
- Is this task friction or meaning? (Meaning → don't outsource.)
- Would I be embarrassed if the raw output were shown as mine?
- Does anything need verification before reuse?
- Would an honest viewer want to know AI was involved?
- Whose material am I uploading — and is it mine to upload?
06Prompt Recipes
Copy, adapt, reuse. The copy button grabs the whole recipe.
The practical creative directorI am an NCAD student working in [your track]. I want help with this project: [describe project in 2–3 sentences, with audience and constraints]. Act as a practical creative director, not a replacement artist. Ask clarifying questions first if needed. Help me create options, critique them against my goals, and identify what I should make or revise myself. Don't pick winners — that's my job.
The critique loopHere is my [brief/draft/image]. Critique it against these criteria: [paste 3–5 criteria from your brief]. Be specific and unsentimental. For each weakness: name it, explain why it matters for my audience, and suggest one concrete revision. Do not rewrite the whole thing.
The verification passList every factual claim you just made (names, dates, sources, statistics, artists, tools). For each one, tell me how confident you are and how I could verify it independently. Flag anything you may have invented.
The generic detectorRead this back and tell me: which parts could appear in any art student's work, and which parts could only have come from my specific project? Be brutal about the generic parts.
07Lab: Boundary Map + Manifesto (70 min)
Cold start friendly — this lab needs nothing from before today.
Part A — AI Boundary Map (20 min, paper first)
Three columns; sort your actual creative tasks. Minimum 5 per column. "Ideation" is not a task — "finding type references" is.
| Green — AI freely | Yellow — AI with care | Red — human only |
|---|---|---|
| Friction; no authorship at stake | Touches creative territory; needs control + maybe disclosure | Where your voice lives |
Part B — Same brief, two tools (30 min, pairs welcome)
- Write a 5-line brief for a real project (Creative Director recipe above).
- Run it in ChatGPT and Claude — one account each per pair is plenty.
- Compare: where do they differ? What did both get generic?
- Run the Critique Loop on the better one.
- Write down three places the human had to intervene — they feed your manifesto.
Zero-generation track
Do Part B entirely as critique: paste a past project description into both tools and compare their critiques. Same lesson, no generation, equal credit.
Part C — Personal AI Use Manifesto v1 (20 min)
8–12 bullets defining ethical, useful AI use for your practice. Specific enough to act on:
- ✗ Weak: "I'll use AI responsibly."
- ✓ Strong: "AI can generate layout options, but the final composition is always drawn by me — and I keep the rejected options as process evidence."
Keep it findable
This is v1. You'll rewrite it in Week 15, and the distance between versions is what the course measures. Photograph it now.
08Integrity Check
The weekly four — same every week
- What part of this work requires your taste?
- What can AI safely accelerate?
- What must be verified before you trust it?
- What should be disclosed if this becomes portfolio or client work?
This week's special case: your manifesto itself. If AI helped you draft it — did the bullets stay yours? If you can't tell, that's worth a bullet of its own. Full agreement: Student AI Integrity Pact.
09Troubleshooting
| Problem | Fix |
|---|---|
| Free tier ran out mid-lab | Pair with a neighbour, or switch to the zero-generation track — it's half the lesson anyway |
| Output feels generic | It is. Add constraints, audience, references, and an "avoid" list. Generic in = generic out |
| A "fact" smells wrong | It probably is. Run the Verification Pass recipe; check anything you'd repeat aloud |
| It forgot my brief mid-chat | Context window full. Fresh chat, re-paste the brief (keep briefs in a doc) |
| I feel gross using this | Legitimate. Stay in the room: take the critique/planning track and put the discomfort into your manifesto. Informed refusal is a course outcome, not a failure |
| It refused a harmless request | Rephrase with context about your legitimate creative purpose — tools over-refuse art tasks sometimes |
10Resources
- VideoBut what is a GPT? — 3Blue1BrownThe best visual explanation of what's inside an LLM — embedded above in Concepts. First 10 minutes minimum.
- VideoAI art, explained — VoxImage generation's history and the artist-ethics framing; diffusion animation at 5:30. From 2022 — tools dated, questions current.
- ArticleLinkedTransformers, visual lesson — 3blue1brown.comThe same explanation as the video, readable at your own pace.
- VideoVerify before teachingIntro to Large Language Models — Andrej KarpathyA calm one-hour deep dive from a leading researcher, for students who want the full picture.
- OfficialLinkedCopyright and AI — US Copyright OfficeClassroom takeaway: purely AI-generated work generally can't be copyrighted — human authorship is the test, which is why your judgment legally matters.
- OfficialLinkedEU AI Act explorerPlain-language guide to the law phasing in around you: synthetic-media transparency is becoming a legal requirement, not a courtesy.
- ToolLinkedHave I Been Trained — SpawningSearch whether artwork appears in common training datasets and register opt-outs. Makes the training-data debate concrete in 90 seconds.
- ArticleLinkedOne Useful Thing — Ethan MollickHype-free ongoing commentary; a good "stay current without doomscrolling" subscription.
Goes stale
Tool names and free-tier limits change fast. Anything tool-specific here was accurate June 2026 — check TOOLS_AND_ACCESS_JUNE_2026.md for the current state.
Optional extension (fast finishers)
- Read the full Integrity Pact; mark one line you'd change and bring the edit next week.
- Run your Part B brief through a third tool (Gemini free tier); add one comparison note.
- Hard mode: write a 100-word refusal statement — one AI use you reject, with specific reasoning. Informed refusal stated well is rarer than enthusiasm.
Next week: Prompting Like an Art Director →