Music off, lights up. People arrive skeptical, curious, anxious. Open warm, not salesy.
The deal
The deal in this room
You do not have to like AI. You do not have to let it make your art. Skepticism is welcome when it's specific.
Read it slowly. This slide buys trust with the hostile third of the room. "All three positions — excited, hostile, overwhelmed — are welcome and useful here."
Today
Today's promise
What these machines actually are
Where they reliably fail
Your own written rules for using them — or refusing them
Point at the third line: "That document is yours. You'll rewrite it in Week 15 and be surprised."
Deliverable
You will leave with
An AI Boundary Map — what you'll use AI for, what you won't
Your AI Use Manifesto v1 — 8–12 rules, in your words
Show a real example manifesto for 20 seconds (demos/week-01). Don't read it all — just prove it's small and real.
The one rule
AI for friction. Humans for meaning.
Friction = summaries, options, checklists, file names, transcripts Meaning = concept, taste, selection, ethics, final craft, your name
"Everything else this term is a footnote to this sentence." It's also on the board all day.
Today's map
Today's map
What these machines actually are — 50 min
The objections, taken seriously — 20 min
Demo: one task, three ways — 30 min
Lab: your map + your manifesto — 70 min
Share, critique, reflect — 40 min
Breaks every ~50 minutes. Protected.
Drop-in reassurance: "If this is your only week, you still leave with the full kit."
Part 1
What these machines actually are
Timer check: should be ~0:10.
LLMs
Inside ChatGPT and Claude
A large language model: trained on enormous amounts of text to do exactly one job —
predict what plausibly comes next.
"No fact database. No looking things up (unless search is on). No understanding the way you understand. Prediction — one token at a time, billions of times."
LLMs
Predict → sample → repeat
It writes one word-chunk (token) at a time. Each one: "what's likely next?"
The scale is the miracle. The method is the catch.
The miracle: prediction at this scale captures logic, style, code. The catch: it optimizes for plausible, never for true.
Video
Watch
The best 3 minutes on this
Play 0:00–3:30 — the predict/sample/repeat loop. Rest is optional homework (it's excellent).
Say first: "Watch what it's actually doing when it 'writes'." Play only the opening segment; full video is linked on the week page.
The catch
Plausible ≠ true
It optimizes for sounds right. It has no concept of is right.
And it cannot tell you which one you're getting.
"This isn't a bug they'll patch next version. It's what the machine is."
The trap
How sure does it sound? vs. How right is it?
Sounds
Confident. Confident. Confident.
Is
Sometimes right. Sometimes wrong. Sometimes invented entirely.
The confidence is constant. The truth isn't.
"The fluency is the trap. The better it reads, the less you check. Train the opposite reflex."
Try this now
90 seconds
Ask your chat tool:
"What do you know about [your name / your townland / your tiny local landmark]?"
Hands up when you catch an invention.
Phones fine. Collect 2–3 examples aloud. Someone always gets confident nonsense about a small place. If everything's right: "Today it was right. You still had to check — that's the habit."
Two limits
Two limits worth knowing
Token — the word-chunk it reads/writes. Everything is counted in these.
Context window — its working memory. Outgrow it and it forgets your brief mid-conversation.
Fix: keep briefs in a doc. Re-paste. Fresh chats.
This explains the #1 confusing experience: "it was great, then it got stupid." It forgot. Not moody — full.
Multimodal
"Multimodal" = it has eyes now
Upload a screenshot of your layout → get critique. Upload your storyboard → get continuity questions.
The critique loop works on images. This matters more for you than generation does.
Seed for Week 10. For artists who never want to generate: this is the AI skill that still serves them.
Image models
How image generators actually work
Diffusion models learn to turn noise into images, step by step, steered by your words.
Nothing is stored. Nothing is pasted. Everything is synthesized.
The Vox diffusion animation is two slides ahead if the room wants the visual.
Image models
What it doesn't do
It does not keep a folder of stolen images. It does not cut and paste anyone's work.
There is nothing "in there" to copy out — only learned patterns.
Be precise here; the room will quote you. The mechanical claim is narrow and true.
Ethics
Image models
What it did do
It was trained on millions of artists' images — mostly scraped, mostly without consent.
It reproduces the patterns of what it saw.
Equally true, equally important. Don't rush past this slide.
Ethics
Both at once
Hold both facts
"It doesn't copy" — true, and doesn't settle the question. "The data was taken without asking" — true, and doesn't make every use theft.
Tool choice becomes an ethical act.
Mention Firefly's licensed-data training vs. scraped-data tools. "Choosing your tool is one of your levers."
Optional video
If the room wants the visual
AI art, explained — Vox
Play 5:30–8:00 — the diffusion animation. Made 2022: tools dated, questions current.
Use if the diffusion idea didn't land visually. Otherwise skip — it's on the week page for homework.
Why output is generic
The average machine
Trained on everything → defaults to the average of everything.
"Make it cyberpunk" returns the mean of every cyberpunk image online.
Specific direction is the only escape from average.
This reframes prompting before Week 2 teaches it: not magic words — escaping the mean.
Part 1, continued
The four failure modes
(memorize these — they're structural)
Timer check: ~0:35. "Every tool, every week, fails these same four ways."
Failure 1 of 4
Hallucination
Invented facts. Fake sources. Artists who don't exist. Delivered in perfect confidence.
Counter-move: verify every name, date, quote, citation. Every one.
"Not lying — it has no concept of truth, only plausibility. You are the fact-checker now."
Failure 2 of 4
Genericness
Competent. Soulless. Looks like everything else. The mean of the internet, served warm.
Counter-move: constraints, references, an "avoid" list — your taste.
"Generic isn't a moral failure of the machine. It's the machine working as designed. Direction is your job."
Failure 3 of 4
Bias
Default faces. Default beauty. Default art history. The data's blind spots become the output's defaults.
Counter-move: notice the defaults. Direct against them deliberately.
Concrete example: ask for "a designer" and look at who it pictures. Worth doing live if time allows.
Failure 4 of 4
Amnesia
Long chat → brief forgotten → quality collapses. (That's the context window, full.)
Counter-move: restate constraints. Fresh chats. Briefs live in docs, not chats.
The practical habit that saves students the most frustration this term.
Try this now
2 minutes
"Name three real Irish illustrators and their best-known work."
Check one claim. Hands up when you find the invention.
Usually at least one hallucinated or mismatched detail in the room. Tie-off: "Hallucination isn't rare. It's ambient."
The point
Notice what all four counter-moves are
Verify. Direct. Notice. Restate.
All human habits. None automatable. This is why judgment is the skill — and why this course exists.
End of Part 1. Break next.
Break
Break — 10 minutes
Back at :00
Protect it. Hostile students often approach 1:1 now — that conversation matters more than restarting to the second.
Part 2
The objections, taken seriously
Tone shift: slower, no jokes for 15 minutes. This section is why skeptics stay.
Ethics
Objection 1
Consent
Training data was largely scraped without permission. True. Lawsuits ongoing. Laws moving.
Tools differ: some train on licensed content (Firefly); most don't. → Your tool choice is an ethical act.
Don't soften, don't catastrophize. "If this makes you angry, your anger is well-founded. Now make it precise."
Ethics
Objection 2
Copyright
In most jurisdictions: purely AI-generated output can't be copyrighted. Human authorship is the legal test.
The law agrees with your art teachers: the human decisions are the work.
US Copyright Office is explicit; linked on the week page. This quietly reframes the course: judgment isn't just ethics — it's ownership.
Ethics
Objection 3
Labor
Some creative tasks are getting automated. True.
What's defensible: taste, judgment, accountability, relationships — the things clients actually pay a person for.
Anxious students need this straight: "Some doors narrow. The ones that stay open belong to people whose judgment sits on top of the tools."
Ethics
Objection 4
Slop
Feeds and stock sites are flooding with low-effort AI content.
Consequence: generic is now free. Distinctive human work just became more valuable, not less.
The market logic skeptics rarely hear: when average is free, only above-average sells.
Ethics
Objection 5
Disclosure
Audiences feel deceived when AI use is hidden. The EU AI Act is making synthetic-media transparency law.
We practise specific, honest disclosure every single week.
"Not because you're suspects — because it's about to be a professional skill, like accessibility."
So
Where does that leave us?
Not worship. Not refusal-from-fear.
Literacy → then your own rules.
Bridge slide. Energy back up.
The framework
AI for friction. Humans for meaning.
GREEN — AI freely: no authorship at stake
YELLOW — AI with care: control + maybe disclosure
RED — human only: where your voice lives
You draw these lines yourself, today, in the lab.
Preview of lab Part A. "Nobody else can draw them — nobody else knows which parts of your practice are load-bearing."
Demo
One task, three ways
Timer check: ~1:20. Switch to the chat tool. Prompts pre-staged in DEMOS_AND_VISUAL_AIDS.md — paste, don't improvise.
Demo
Stage 1
The weak prompt
Make a poster concept for an art school degree show.
Watch how confident it is.
Paste it. Read the output flatly. Then THE question to the room: "What did it get wrong?" Collect three answers before you offer yours.
Demo
Stage 1 debrief
What did it get wrong?
No audience. No constraints. No venue. No voice. Could be any college, any year, anywhere.
The model didn't fail. The brief did.
This sentence is the bridge to Week 2. Let it sit.
Demo
Stage 2
The creative brief
Same task + audience + tone + constraints (riso, 2-colour, typography-led) + references +
"Don't pick a winner — that's my job."
Paste the full Field Notes brief from the demos file. While output streams, narrate WHY each constraint exists. Point hard at the last line: the human keeps the decision.
Demo
Compare
Same model. Different human.
Weak prompt
Generic poster #4,000,000 Model decided everything 10 seconds
Creative brief
5 usable directions + risks Human kept every decision 90 seconds
The quality lives in the brief.
"Eighty seconds of thinking bought all of that."
Demo
Stage 3
The critique loop
Direction 3 is the weakest. Critique it against my brief: where is it generic, where does it fight the riso constraint? Then give me two sharper variants that fix those problems.
Let the ROOM pick the weakest direction first — instant engagement. The model critiquing itself, steered by human taste, is the move students remember.
Demo
The planned failure
When it confidently breaks the rules
(One direction usually violates a constraint — full-colour photography in a 2-colour riso brief.)
It was confident about an impossible direction. This is why you stay the director.
If it happens live, celebrate it. If not, show the saved screenshot from prep. Never pretend failures don't happen — they're the syllabus.
Demo
Division of labor
What the human supplied
Audience. Constraints. References. Taste. The decision.
What the model supplied
Speed. Volume. Options. A tireless first draft.
"That division of labor is the entire course. Everything from here is discipline-specific versions of this slide."
Circulate. You're listening for the "oh" sound. Pull one before/after to the projector if a volunteer offers.
Break
Break — 10 minutes
Lab next: you write your rules.
~1:50. Stage the lab slide before you walk away.
Lab · 70 minutes
Your map. Your rules.
A. Boundary Map — 20 min · B. Same brief, two tools — 30 min · C. Manifesto v1 — 20 min
Instructions stay on screen during each part. Write the three parts + times on the board too.
Lab A · 20 min
Part A
AI Boundary Map
Three columns. Your real tasks, sorted:
GREEN AI freely · YELLOW AI with care · RED human only
Minimum 5 real tasks per column. "Ideation" is not a task. "Finding type references" is.
Paper first. Circulate: "What did you actually make this week? Which column does each step go in?" Empty red column = romanticizing; empty green = underestimating. Say both aloud.
Lab B · 30 min
Part B
Same brief, two tools
5-line brief for a real project (recipe in handout)
Run in ChatGPT and Claude — pairs welcome
Compare: differences? shared genericness?
Critique-loop the better one
Write down 3 places the human had to intervene
The three interventions are the deliverable from B — they feed the manifesto. Keep saying it.
Always available
Part B alternative
Zero-generation track
Same lab, no generation: paste a past project description into both tools → pure critique mode → compare the critiques → same 3 interventions.
Equal credit. Equal lesson.
Offer it neutrally, to everyone. Some of your strongest students take this track on principle. The course means it — so should you.
Lab C · 20 min
Part C
Manifesto v1 — 8–12 bullets
Weak
"I'll use AI responsibly."
Strong
"AI can generate layout options; final composition is always drawn by me, and I keep the rejected options as process evidence."
The weak/strong pair does the teaching. Demand a real task or real rule in every bullet. This document returns in Week 15 as the pre/post instrument.
Share-out · 30 min
Volunteers: one map row + one manifesto bullet
Room responds with:
Where's the human point of view in that rule?
Protecting craft, or avoiding learning — how do you tell?
Would that bullet survive a deadline?
4–6 volunteers. If anyone wrote a refusal statement, have it read aloud and give it equal weight — the room must see refusal-with-reasons earns full respect.
Every week
Integrity check
The weekly four
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 goes public?
Same four questions every week, until they're reflexes.
Reflection
Write 4 lines
One useful thing AI did today
One thing it got wrong (and how you caught it)
One human decision that improved the result
One boundary you're keeping
Two quiet minutes. These notes accumulate into their process portfolio — say so.
Before you leave
Keep your manifesto somewhere findable
Week 15: you'll write v2 next to it. The distance between them is what this course measures.
Practical: photograph it, notes app, email to self — now, before leaving.
Next Wednesday
Prompting Like an Art Director
Briefs, references, and critique loops — turning today's "the brief is everything" into a repeatable skill.
Drop-ins always welcome. Bring a real project.
End on time. Stay 10 minutes for 1:1 questions — Week 1 corridor conversations decide who comes back.