POLYU SCHOOL OF DESIGN / SD2112 / WEEK 12

Can a machine

originate a design?

Course recap and discussion / ESFQ at 12:30

"The Analytical Engine has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform."

Ada Lovelace, Note G, 1843 — the first published program, for a machine that was never built

SD2112 · AI IN DESIGN · WEEK 12

02

WEEKS 1-3 / CONCEPTS, MEDIATION AND AI PARADIGMS

Write the rule, or show the examples.

MACHINE A · RULES if seat and back and legs >= 3: return "chair" definition → verdict exact · explainable · brittle MACHINE B · EXAMPLES 12 000 photos labelled "chair" → a feel for chair-ness fuzzy · fluent · cannot say why

Machine A: symbolic AI, 1956 onwards — definitions, logic, expert systems. Machine B: machine learning, 1958 / 1986 / 2012 — statistics over examples.

SD2112 · AI IN DESIGN · WEEK 12

03

GIO’S COURSE EXTENSION · MACHINE A / MACHINE B / A+B

Same request. Different machines.

THE REQUEST: “TWO CIRCLES”

MACHINE A · EXPLICIT INSTRUCTIONS

MACHINE B · LEARNED INTERPRETATION

circle(200, 300, 150);

circle(400, 200, 250);

Historical Stable Diffusion samples.

Appearance suggested; count not enforced.

A+B: code sets the boundary; generated pixels supply the surface.

SD2112 · AI IN DESIGN · WEEK 12

04

05 · HUMANS + CONCEPTS · MACHINES + CONCEPTS

Two theories. Two machines.

HUMANS + CONCEPTS MACHINES + CONCEPTS RULE-BASED CLASSICAL THEORY a concept is a definition: necessary and sufficient conditions Aristotle · Kant · the dictionary GOFAI · SYMBOLIC AI knowledge written down as rules, applied by a program Dartmouth 1956 · ELIZA · expert systems ADAPTIVE PROTOTYPE THEORY a concept is its best examples, membership a matter of degree Wittgenstein 1953 · Rosch 1975 CONNECTIONISM · MACHINE LEARNING the knowledge is in the weights, found from examples Rosenblatt 1958 · Hinton 1986 · 2012 · today the same two ideas, in a mind and in a machine: a rule you can read, or examples you cannot

Rule-based: classical theory + GOFAI, a definition a machine applies. Adaptive: prototype theory + connectionism, examples held in weights nobody can read. Your reflection is about this distinction.

SD2112 · AI IN DESIGN · WEEK 12

05

03 · SOL LEWITT · SPLAT · 2001 · CAIXAFORUM BARCELONA

A wall drawing: LeWitt wrote the instruction and signed a certificate; assistants drew it on this wall, and can draw it again on any other. Photo: Ardfern, 2014, CC BY-SA 3.0, Wikimedia Commons.

04 · GEORG NEES · SCHOTTER · c. 1968

One rule. One random number.

Twelve columns, twenty-two rows of squares. Each square is moved and turned by a random amount, and the amount grows with the row.

Row 0: a perfect grid. Row 21: gravel. Schotter means gravel.

The rule decides everything except two numbers per square. Chance is kept on a leash; the picture is order turning into disorder.

The sliders are the rule's two numbers: how much disorder, how many rows.

After Nees: our execution of his rule. Siemens 4004, ALGOL, Zuse Graphomat; the original print is 28 × 22 cm.

SD2112 · AI IN DESIGN · WEEK 12

07

Live · click = new dice↗

04 · FRIEDER NAKE · WALK-THROUGH-RASTER · SERIES 2, 1–4 · 1966

Look first. What repeats? What never happens? Four prints from one program: ALGOL 60 on a Zuse Graphomat Z64. Victoria and Albert Museum, E.955-2008.

04 · WALK-THROUGH-RASTER · THE RULE IN FOUR STEPS

Four states. The cell above decides.

A grid, drawn column by column, top to bottom. Every cell is one of four things: empty, a bar, a cap, or both (10 PRINT had two). Two rules decide: a bar is more likely near the diagonal, and a cap can only be drawn under an empty cell, so a cell remembers one thing about the one above it. The empty space flows up and to the right, and every field of caps has gaps.

1 · THE RASTER 2 · FOUR STATES 3 · THE RULE 4 · THE FLOW a grid of cells, 30 × 30 drawn column by column, top to bottom empty a bar a cap both every cell is one of these four a bar | on its left, a cap ¯ on its top two yes/no decisions per cell Nake's were his own; these are ours the cell above: empty so this one may get a cap under a bar or a cap: never a bar: likelier near the diagonal one fact carried from cell to cell empty space flows up and right so fields of caps have gaps

Nake, Walk-Through-Raster, series 2.1–4, 1966. Our reading of his rule: two yes/no decisions per cell, and one fact carried from the cell above.

SD2112 · AI IN DESIGN · WEEK 12

09

02 · THE CLASSICAL THEORY · ARISTOTLE TO THE DICTIONARY

Classical theory of concepts

A definition is a list of properties a thing must have to belong: necessary, and together sufficient. Bachelor = man + unmarried.

Categorising is then trivial: check each property, apply the concept. X is a bachelor if X is a man and X is unmarried. A machine can do that.

New ideas by assembly: put definitions together and you have a new one. Kant: analytic, all in the words.

Aristotle, pointing down at the things, is this theory: sort them by the properties they share and keep the ones that are necessary. Plato, pointing up, disagrees.

Plato and Aristotle, detail of Raphael's The School of Athens, 1509–1511. Public domain, Wikimedia Commons.

SD2112 · AI IN DESIGN · WEEK 12

10

03 · PROTOTYPE THEORY

Prototype theory

THE THEORY

A structured representation of what members tend to have.

Not a list of conditions but a picture of the typical case, and a distance from it. Membership is a degree: a robin is a very good bird, a penguin a poor one, and neither needs a definition.

THE GAIN

Only similarity is needed.

No definition to write: you judge a new thing by how much it resembles what you have seen. That is learning from examples, and it is why a machine can hold "chair" with no rule for it.

THE COST

Outliers, and combinations.

Exceptions are hard: there are fewer examples at the edge, so the edge is unsure. And combining concepts is a puzzle: which properties of "pet" and "fish" does "pet fish" keep?

SD2112 · AI IN DESIGN · WEEK 12

11

04 · ONE NEURON

A neuron is a weighted vote.

Two inputs, two weights, a bias, a threshold. Multiply, add, compare with zero: that is the whole unit. Three numbers hold everything it knows. Rosenblatt, 1958, called it a perceptron; the shape and the size are the week-1 cup.

ONE NEURON · ROSENBLATT, 1958 × +1.6 1.1 height ÷ width × -0.8 0.4 size INPUTS · THE OBJECT Σ + b b = -1.2 1.1×+1.6 + 0.4×-0.8 -1.2 = +0.24 WEIGHTS · THE KNOWLEDGE above 0? 1 GUESS +0.24 > 0 so: "cup" three numbers decide. learning = changing them when the guess is wrong

Rosenblatt, "The perceptron: a probabilistic model for information storage and organization in the brain", Psychological Review 65, 1958. In the html deck the neuron learns live, one example at a time.

SD2112 · AI IN DESIGN · WEEK 12

12

Live · click = pause · C = start again↗

05 · GEOFFREY HINTON

Fifty years betting on the brain.

A psychologist, like Rosenblatt. 1986: backpropagation, with Rumelhart and Williams, when almost nobody believed in networks.

Connectionism: a concept is a pattern of activity over many units, never a list. Rosch's graded categories as an engineering brief.

2012: with his students Krizhevsky and Sutskever, AlexNet wins ImageNet. 2018: the Turing Award. 2023: leaves Google to speak about the risks.

2024: the Nobel Prize in Physics, with John Hopfield, "for foundational discoveries and inventions that enable machine learning with artificial neural networks".

Geoffrey Hinton at the 2024 Nobel Lectures, Stockholm University. Photo: Jay Dixit, CC BY-SA 4.0, Wikimedia Commons.

SD2112 · AI IN DESIGN · WEEK 12

13

05 · BACKPROPAGATION · 1986 · THE PICTURE, LIVE

The guess goes forward. The error comes back.

THE EXAMPLE PIXELS IN HIDDEN HIDDEN GUESS OUT chair 0.35 cup 0.65 THE TRUTH: a chair chair should be 1.00 it said 0.35 error: 0.65 adjust the weights FORWARD · every unit sums its inputs and passes a number on · the guess comes out at the end BACKWARD · the error is shared out along the same connections · every weight moves a little, in proportion to its share

Nine pixels in, five and three hidden units, two out: nineteen units, 66 weights. Rumelhart, Hinton & Williams, "Learning representations by back-propagating errors", Nature 323, 1986.

SD2112 · AI IN DESIGN · WEEK 12

14

Live · click = pause · C = start again↗

06 · GPUS · NOT ONLY FOR GAMING

The chips made for games.

A graphics processing unit does the same small calculation on millions of pixels at once: textures, shading, vertices, physics. Games needed that, and paid for twenty years of it.

2007: CUDA. NVIDIA opens the chip to any calculation that has the same shape: many small identical jobs, no waiting.

A neural network is exactly that shape: multiply, add, everywhere, at once. What a CPU did in weeks, a graphics card did in days.

The other school of AI finally had its engine.

An NVIDIA GeForce GTX 580, late 2010: the card AlexNet was trained on, two of them. Photo: TheStriker, CC BY-SA 4.0, Wikimedia Commons.

SD2112 · AI IN DESIGN · WEEK 12

15

06 · 2012 · ALEXNET

Deep: the features are learned too.

30 September 2012: Krizhevsky, Sutskever and Hinton win the ImageNet challenge with an eight-layer network: 15.3% error against 26.2% for the runner-up, trained on 1.2 million labelled photos in about a week on two GTX 580s. The other entries ran on hand-crafted features; AlexNet grew its own from the pixels.

PIXELS EDGES PARTS OBJECTS A LABEL leg seat back chair 0.93 stool 0.05 table 0.02 input: the pixels layer 1 (convolutional) layers 2 – 5 (convolutional) layers 6 – 8 (fully connected) output: 1,000 scores AlexNet, 2012: 8 layers, 60 million weights, all found from 1.2 million labelled photos. nobody wrote a rule for "edge" or "leg"; the layers became those detectors because it lowered the error

ImageNet: Fei-Fei Li, from 2006; 14 million images labelled by 49,000 Mechanical Turk workers in 167 countries.

SD2112 · AI IN DESIGN · WEEK 12

16

IHDE · POSTPHENOMENOLOGY AND TECHNOLOGICAL MEDIATION

Ihde: technology mediates our world.

HUMAN TECHNOLOGY WORLD EMBODIMENT (I — TECHNOLOGY) → WORLD Look through the viewfinder at a scene. HERMENEUTIC I → (TECHNOLOGY — WORLD) Read a photograph as a representation. OTHER RELATIONS: ALTERITY / BACKGROUND

Same camera. Different relations—depending on how you use it.

SD2112 · AI IN DESIGN · WEEK 12

17

WEEKS 4-6 / LANGUAGE, VISUAL AND AUDIO

Using AI, or incorporating AI.

WEEKS 4-6 / THE CREATIVE PROCESS

Using AI

AI as a tool in how you design: a brief drafted with a chatbot, a moodboard from a diffusion model, generative fill, layouts from Figma Make.

You stay the author. You become the curator, the briefer, the editor.

WEEKS 8-11 / THE PRODUCT

Incorporating AI

AI as a material in what you design: a feed, a recommendation, an assistant, a filter — the product decides something for each person, on its own.

You design a behaviour, not a picture. Data, bias, trust and accountability become design problems.

SD2112 · AI IN DESIGN · WEEK 12

18

01 · ELIZA · 1966

Pattern matching can sound like understanding.

SD2112 · AI IN DESIGN · WEEK 12

19

02 · TRANSFORMER · CAUSAL SELF-ATTENTION · VASWANI ET AL. 2017 ↗

One sequence. Many positions working together.

1 Tokens 2 Context 3 Training 4 Generation A robot robot writes writes a a brief brief . Parallel next-token predictions → compare with known targets Context: each position uses itself and earlier tokens only. Training works across positions together. Generation adds one token at a time.

SD2112 · AI IN DESIGN · WEEK 12

20

Live · choose a step · click a token · replay training or generation↗

03 · QUESTION ANSWERING

Retrieve evidence, then write one answer.

01 · QUESTION

What do I need to know?

A visitor asks: “What time does the event start?”

02 · SEARCH

Find a source.

Search the official event page for the schedule.

03 · EVIDENCE

Select a passage.

Bring the relevant text and its URL into context.

04 · ANSWER

Write from evidence.

Answer once; mark details the source does not give.

SD2112 · AI IN DESIGN · WEEK 12

21

04 · ONE TOOL CALL · SEARCH

The model requests. The harness executes.

MACHINE B · REQUEST

“Search the event page.”

Tool: search

Query: exhibition start time

→

MACHINE A · EXECUTE

Check, then run.

The harness checks access, runs the tool, and records the result.

→

OBSERVATION

Return the evidence.

The page text and URL enter the model’s context.

A tool call is a structured request to software outside the model.

SD2112 · AI IN DESIGN · WEEK 12

22

CLIP · CONTRASTIVE LANGUAGE–IMAGE PRETRAINING

CLIP brings words and images into a shared space.

CAPTION a chair with curved back TEXT ENCODER caption to vector IMAGE ENCODER picture to vector SHARED EMBEDDING SPACE matching pair different caption mismatched pairs move apart Similarity is not synthesis. A matching vector is not a picture.

Matching pairs move closer; mismatched pairs move apart. CLIP does not generate an image.

SD2112 · AI IN DESIGN · WEEK 12

23

CLASSIC LATENT DIFFUSION · GENERATION

Generation turns new noise into an image.

TEXT FEATURES NEW NOISY LATENT DENOISER learned prediction FINAL LATENT VAE DECODER OUTPUT IMAGE REPEAT LATENT UPDATE Decode after the last update.

The loop changes the latent, not the trained weights. Decode the final latent after the last update.

SD2112 · AI IN DESIGN · WEEK 12

24

CASE REFLECTION · SD2112 CO-CREATION

The brief emerged through choices.

01 · CLEAN MARK

02 · FOREST VARIANT

03 · BRIGHTER EDIT

→

→

A recognizable form to keep.

A setting and material enter.

A chosen refinement—not a finish.

The mark did not arrive fully formed. What entered the brief after seeing a candidate?

SD2112 · AI IN DESIGN · WEEK 12

25

CASE REFLECTION · SD2112 CO-CREATION

A better prompt was not always the right repair.

When the failure is in the transition, change the transition.

EARLIER ITERATIONS

Rejected montage.

Coherent origami repair.

Later, a burn to refine.

→

V8 · OBSERVED MISS

Flames inside the counter

and along the front edge.

→

V9 · TARGETED CHANGE

Replace only that transition:

front-facing, top → bottom.

Prompt repair? Reference repair? Workflow repair? Choose from the failure you can point to.

V9 delivered · review pending · not a claim of final acceptance

What did you delegate—and what did you insist on keeping?

SD2112 · AI IN DESIGN · WEEK 12

26

WEEKS 4-6 / MUSIC / MACHINE A

Musical choices in a sequencer

FIX

Tempo and bass

Set 100 BPM and repeat the same four-bar bass pattern.

VARY

Drum fills

At the end of each phrase, choose one of three fills or leave a rest.

LISTEN

Groove and space

Listen back. Does the fill lift the next phrase or crowd it?

SD2112 · AI IN DESIGN · WEEK 12

27

MACHINE B · TWO DOCUMENTED APPROACHES

A picture of sound. Or tokens of sound.

Historic Riffusion v1 (2022): a documented magnitude-spectrogram route DIFFUSION text + noise MAGNITUDE PICTURE generated spectrogram PHASE RECONSTRUCTION Griffin-Lim + inversion AUDIO waveform Magnitude alone is not enough: phase must be estimated before reconstructing audio. MusicGen (2023): a documented audio-codec-token route CODEC TOKENS EnCodec audio codes PREDICT TOKENS learned token predictor CODEC DECODE tokens to waveform AUDIO waveform Training audio is codec-encoded; generation predicts tokens. No claim about Suno internals.

Historical Riffusion v1 and MusicGen illustrate mechanisms—not Suno’s unpublished internals.

SD2112 · AI IN DESIGN · WEEK 12

28

MACHINE B · SUNO

Your demo can guide a new musical proposal.

Supply your own exported demo as audio reference; describe the transformation.

Listen to the actual result: arrangement, pacing, timbre, voices and surprises.

Keep a candidate because it serves the brief—not because it arrived first.

Suno · generate through the configured Easel integration in the workshop.

Public Suno landing page · 9 Oct 2026 · not the Easel interface.

SD2112 · AI IN DESIGN · WEEK 12

29

WEEKS 8-11 / AI INSIDE PRODUCTS / R8

People chose humans but rated AI higher.

WHAT THEY FOUND

About 57-62% of choices went to humans, but AI replies generally rated higher for empathy and feeling heard.

WHAT IT DOES NOT SHOW

Sources were chosen, not randomly assigned. Brief tasks, not established relationships or clinical benefit.

FOR YOUR DESIGN

Should users choose their source?

Wenger et al. (2026), Fig. 1; published PDF p.5.

Cropped plot; CC BY 4.0. Dashed line = equal choice.

SD2112 · AI IN DESIGN · WEEK 12

30

AGENCY / CALIBRATED TRUST

Can the person disagree and act?

BEFORE

Know the boundary.

What does the system do? What evidence and data does it lack?

DURING

Question the answer.

Compare, correct, override or ask for a source before the decision matters.

AFTER

Leave or recover.

Keep an exit, a human route and a way to repair the consequences.

SD2112 · AI IN DESIGN · WEEK 12

31

FEEDBACK / CONCEPTUAL LOOP

The next screen changes the next data.

Possible feedback, not a measured universal law.

SD2112 · AI IN DESIGN · WEEK 12

32

READ THE SOURCE / RESULT + LIMIT / R4

Read the error row, not just the average.

DF/DM: darker female/male.

LF/LM: lighter female/male.

IBM error: All 12.1%.

DF 34.7% / LM 0.3%.

Within-group rates.

2017 binary classification.

Not identification.

Causes/current rates unknown.

Table 4 / PDF p.9 / PMLR 2018

SD2112 · AI IN DESIGN · WEEK 12

33

WEEK 10 / DESIGN PROCESS

The Double Diamond

Discover and Define: understand the people and decide which problem the AI would address.

Develop and Deliver: compare interactions, test with people, and revise when the evidence changes.

Week 8 / AI in the design process

SD2112 · AI IN DESIGN · WEEK 12

34

THE RESULT / D

Better ratings. More similar stories.

Doshi & Hauser (2024), Fig. 2A crop, CC BY 4.0. Dots: differences from human-only; bars: uncertainty.

MAIN FINDING

AI access raised average novelty and usefulness ratings. The two AI conditions also produced more similar stories.

LIMIT

A short writing task, not studio learning; similarity does not count cultural voices.

FOR YOUR DESIGN

Inspect the portfolio, not only its best item.

SD2112 · AI IN DESIGN · WEEK 12

35

COURSE RECAP / RECOMMENDERS, BIAS AND MEDIATION

Week 9 papers

SD2112 · AI IN DESIGN · WEEK 12

36

COURSE RECAP / CURATION, AUTHORSHIP AND RESPONSIBILITY

Week 11 papers

SD2112 · AI IN DESIGN · WEEK 12

37

AGENCY + AUTHORSHIP / ORIGINAL CONTRIBUTION MAP

Trace four different decisions.

01 / GENERATE

Tool proposes.

AI writes a draft from the source text.

→

02 / SELECT

Designer chooses.

Choose a draft that names its source.

→

03 / REVISE

Designer edits.

Check the facts and rewrite the text.

→

04 / AUTHORISE

Team releases.

A person approves it for visitors.

AI drafts. People choose, edit and publish. Keep a record of that work.

SD2112 · AI IN DESIGN · WEEK 12

38

FOUR DISTINCT QUESTIONS

Feeling, work, permission, law.

01 / FEEL

Does it feel mine?

A reported feeling is not a copyright certificate.

02 / CONTRIBUTE

Who did what?

Credit sources, selection, edits and review.

03 / PERMISSION

May we reuse it?

Check rights and consent for the material.

04 / PROTECTION

What is protected?

Human expression is a separate legal question.

SD2112 · AI IN DESIGN · WEEK 12

39

12:07-12:13 / THE DESIGNER'S ROLE

Designers as curators

OUTPUTS

What makes the cut?

Choose what fits the brief. Be able to explain why you kept, edited or rejected a result.

TRAINING DATA

Whose work shapes it?

Look at whose work and whose perspective the examples represent. Check permissions and gaps.

HUMAN OVERSIGHT

Who has the last word?

Decide where an agent can act and when a person must review, intervene or stop it.

SD2112 · AI IN DESIGN · WEEK 12

40

12:13-12:19 / PROMPTS, RULES AND ACTIONS

Language as an interface

A phrase can specify a rule, as in a geometry instruction.

A description can steer a model toward a range of possible stories or images.

A music brief can do both. Fix the tempo and leave the arrangement open.

When the prompt can trigger an action, what is left implicit?

What should the interface show before a person lets the agent act?

Week 4 / The model requests, the harness executes

SD2112 · AI IN DESIGN · WEEK 12

41

12:19-12:25 / CONTRIBUTION AND CREDIT

Fractional authorship

A team makes a poster with AI. Who shaped the idea, made the source material, selected the image, or edited the final composition?

Credit each contribution you can identify and describe what the model did. What should you disclose if the origins of training material are unclear?

Ask separately who gave permission to use the material and who holds rights in the finished work.

Week 11 / Trace four different decisions

SD2112 · AI IN DESIGN · WEEK 12

42

12:25-12:27 / SYNTHESIS

What answer would you give Lovelace now?

One responsibility you will carry into your work.

One skill you will keep practising.

One question you still want to answer.

SD2112 · AI IN DESIGN · WEEK 12

43

COURSE ARCHIVE

Earlier slides

SD2112 · AI IN DESIGN · WEEK 12

44

12:27-12:30 / ASK ME ANYTHING

Course AMA

Submit a short question about the course in ClassPoint.

A confusing concept, a case we discussed, or an ethical question you are still thinking about.

SD2112 · AI IN DESIGN · WEEK 12

45

Short answer

12:30 / STUDENT FEEDBACK

ESFQ

SD2112 · AI IN DESIGN · WEEK 12

46

AFTER ESFQ / POSTER FAIR

Claim your poster-fair space

Open the Excalidraw room link shared in class.

One editor per group finds the exact group name and adds the project title.

Add a draft poster or placeholder. Landscape and portrait both work.

Add viewing links for the video and mediation brief when available.

Ask another group to find your space and open the links. Submit the final work in Canvas.

SD2112 · AI IN DESIGN · WEEK 12

47

GROUP PROJECT / ASK ME ANYTHING

Group project AMA

Submit a short question about your group project in ClassPoint.

Ask about your design or evidence, the poster, video, mediation brief, Excalidraw, or Canvas.

SD2112 · AI IN DESIGN · WEEK 12

48

Short answer
SD2112 / AI in Design / Week 12
1

POLYU SCHOOL OF DESIGN / SD2112 / WEEK 12

Can a machine

originate a design?

Course recap and discussion / ESFQ at 12:30

2

"The Analytical Engine has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform."

Ada Lovelace, Note G, 1843 — the first published program, for a machine that was never built

Week 1 / Original slide

3

WEEKS 1-3 / CONCEPTS, MEDIATION AND AI PARADIGMS

Write the rule, or show the examples.

MACHINE A · RULES if seat and back and legs >= 3: return "chair" definition → verdict exact · explainable · brittle MACHINE B · EXAMPLES 12 000 photos labelled "chair" → a feel for chair-ness fuzzy · fluent · cannot say why

Machine A: symbolic AI, 1956 onwards — definitions, logic, expert systems. Machine B: machine learning, 1958 / 1986 / 2012 — statistics over examples.

Week 1 / Original slide

4

GIO’S COURSE EXTENSION · MACHINE A / MACHINE B / A+B

Same request. Different machines.

THE REQUEST: “TWO CIRCLES”

MACHINE A · EXPLICIT INSTRUCTIONS

MACHINE B · LEARNED INTERPRETATION

circle(200, 300, 150);

circle(400, 200, 250);

Historical Stable Diffusion samples.

Appearance suggested; count not enforced.

A+B: code sets the boundary; generated pixels supply the surface.

Week 5 / Original slide

5

05 · HUMANS + CONCEPTS · MACHINES + CONCEPTS

Two theories. Two machines.

HUMANS + CONCEPTS MACHINES + CONCEPTS RULE-BASED CLASSICAL THEORY a concept is a definition: necessary and sufficient conditions Aristotle · Kant · the dictionary GOFAI · SYMBOLIC AI knowledge written down as rules, applied by a program Dartmouth 1956 · ELIZA · expert systems ADAPTIVE PROTOTYPE THEORY a concept is its best examples, membership a matter of degree Wittgenstein 1953 · Rosch 1975 CONNECTIONISM · MACHINE LEARNING the knowledge is in the weights, found from examples Rosenblatt 1958 · Hinton 1986 · 2012 · today the same two ideas, in a mind and in a machine: a rule you can read, or examples you cannot

Rule-based: classical theory + GOFAI, a definition a machine applies. Adaptive: prototype theory + connectionism, examples held in weights nobody can read. Your reflection is about this distinction.

Week 3 / Original slide

6

03 · SOL LEWITT · SPLAT · 2001 · CAIXAFORUM BARCELONA

A wall drawing: LeWitt wrote the instruction and signed a certificate; assistants drew it on this wall, and can draw it again on any other. Photo: Ardfern, 2014, CC BY-SA 3.0, Wikimedia Commons.

Week 2 / Original slide

7

04 · GEORG NEES · SCHOTTER · C. 1968

One rule. One random number.

Twelve columns, twenty-two rows of squares. Each square is moved and turned by a random amount, and the amount grows with the row.

The sliders are the rule's two numbers: how much disorder, how many rows.

After Nees: our execution of his rule. Siemens 4004, ALGOL, Zuse Graphomat; the original print is 28 × 22 cm.

Week 2 / Original slide

Open the live sketch › — click = new dice

8

04 · FRIEDER NAKE · WALK-THROUGH-RASTER · SERIES 2, 1–4 · 1966

Look first. What repeats? What never happens? Four prints from one program: ALGOL 60 on a Zuse Graphomat Z64. Victoria and Albert Museum, E.955-2008.

Week 2 / Original slide

9

04 · WALK-THROUGH-RASTER · THE RULE IN FOUR STEPS

Four states. The cell above decides.

A grid, drawn column by column, top to bottom. Every cell is one of four things: empty, a bar, a cap, or both (10 PRINT had two). Two rules decide: a bar is more likely near the diagonal, and a cap can only be drawn under an empty cell, so a cell remembers one thing about the one above it. The empty space flows up and to the right, and every field of caps has gaps.

1 · THE RASTER 2 · FOUR STATES 3 · THE RULE 4 · THE FLOW a grid of cells, 30 × 30 drawn column by column, top to bottom empty a bar a cap both every cell is one of these four a bar | on its left, a cap ¯ on its top two yes/no decisions per cell Nake's were his own; these are ours the cell above: empty so this one may get a cap under a bar or a cap: never a bar: likelier near the diagonal one fact carried from cell to cell empty space flows up and right so fields of caps have gaps

Nake, Walk-Through-Raster, series 2.1–4, 1966. Our reading of his rule: two yes/no decisions per cell, and one fact carried from the cell above.

Week 2 / Original slide

10

02 · THE CLASSICAL THEORY · ARISTOTLE TO THE DICTIONARY

Classical theory of concepts

A definition is a list of properties a thing must have to belong: necessary, and together sufficient. Bachelor = man + unmarried.

Aristotle, pointing down at the things, is this theory: sort them by the properties they share and keep the ones that are necessary. Plato, pointing up, disagrees.

Plato and Aristotle, detail of Raphael's The School of Athens, 1509–1511. Public domain, Wikimedia Commons.

Week 3 / Original slide

11

03 · PROTOTYPE THEORY

Prototype theory

THE THEORY

A structured representation of what members tend to have.

Not a list of conditions but a picture of the typical case, and a distance from it. Membership is a degree: a robin is a very good bird, a penguin a poor one, and neither needs a definition.

THE GAIN

Only similarity is needed.

No definition to write: you judge a new thing by how much it resembles what you have seen. That is learning from examples, and it is why a machine can hold "chair" with no rule for it.

THE COST

Outliers, and combinations.

Exceptions are hard: there are fewer examples at the edge, so the edge is unsure. And combining concepts is a puzzle: which properties of "pet" and "fish" does "pet fish" keep?

Week 3 / Original slide

12

04 · ONE NEURON

A neuron is a weighted vote.

Two inputs, two weights, a bias, a threshold. Multiply, add, compare with zero: that is the whole unit. Three numbers hold everything it knows. Rosenblatt, 1958, called it a perceptron; the shape and the size are the week-1 cup.

ONE NEURON · ROSENBLATT, 1958 × +1.6 1.1 height ÷ width × -0.8 0.4 size INPUTS · THE OBJECT Σ + b b = -1.2 1.1×+1.6 + 0.4×-0.8 -1.2 = +0.24 WEIGHTS · THE KNOWLEDGE above 0? 1 GUESS +0.24 > 0 so: "cup" three numbers decide. learning = changing them when the guess is wrong

Rosenblatt, "The perceptron: a probabilistic model for information storage and organization in the brain", Psychological Review 65, 1958. In the html deck the neuron learns live, one example at a time.

Week 3 / Original slide

Open the live sketch › — click = pause · C = start again

13

05 · GEOFFREY HINTON

Fifty years betting on the brain.

A psychologist, like Rosenblatt. 1986: backpropagation, with Rumelhart and Williams, when almost nobody believed in networks.

2024: the Nobel Prize in Physics, with John Hopfield, "for foundational discoveries and inventions that enable machine learning with artificial neural networks".

Geoffrey Hinton at the 2024 Nobel Lectures, Stockholm University. Photo: Jay Dixit, CC BY-SA 4.0, Wikimedia Commons.

Week 3 / Original slide

14

05 · BACKPROPAGATION · 1986 · THE PICTURE, LIVE

The guess goes forward. The error comes back.

THE EXAMPLE PIXELS IN HIDDEN HIDDEN GUESS OUT chair 0.35 cup 0.65 THE TRUTH: a chair chair should be 1.00 it said 0.35 error: 0.65 adjust the weights FORWARD · every unit sums its inputs and passes a number on · the guess comes out at the end BACKWARD · the error is shared out along the same connections · every weight moves a little, in proportion to its share

Nine pixels in, five and three hidden units, two out: nineteen units, 66 weights. Rumelhart, Hinton & Williams, "Learning representations by back-propagating errors", Nature 323, 1986.

Week 3 / Original slide

Open the live sketch › — click = pause · C = start again

15

06 · GPUS · NOT ONLY FOR GAMING

The chips made for games.

A graphics processing unit does the same small calculation on millions of pixels at once: textures, shading, vertices, physics. Games needed that, and paid for twenty years of it.

The other school of AI finally had its engine.

An NVIDIA GeForce GTX 580, late 2010: the card AlexNet was trained on, two of them. Photo: TheStriker, CC BY-SA 4.0, Wikimedia Commons.

Week 3 / Original slide

16

06 · 2012 · ALEXNET

Deep: the features are learned too.

30 September 2012: Krizhevsky, Sutskever and Hinton win the ImageNet challenge with an eight-layer network: 15.3% error against 26.2% for the runner-up, trained on 1.2 million labelled photos in about a week on two GTX 580s. The other entries ran on hand-crafted features; AlexNet grew its own from the pixels.

PIXELS EDGES PARTS OBJECTS A LABEL leg seat back chair 0.93 stool 0.05 table 0.02 input: the pixels layer 1 (convolutional) layers 2 – 5 (convolutional) layers 6 – 8 (fully connected) output: 1,000 scores AlexNet, 2012: 8 layers, 60 million weights, all found from 1.2 million labelled photos. nobody wrote a rule for "edge" or "leg"; the layers became those detectors because it lowered the error

ImageNet: Fei-Fei Li, from 2006; 14 million images labelled by 49,000 Mechanical Turk workers in 167 countries.

Week 3 / Original slide

17

IHDE · POSTPHENOMENOLOGY AND TECHNOLOGICAL MEDIATION

Ihde: technology mediates our world.

HUMAN TECHNOLOGY WORLD EMBODIMENT (I — TECHNOLOGY) → WORLD Look through the viewfinder at a scene. HERMENEUTIC I → (TECHNOLOGY — WORLD) Read a photograph as a representation. OTHER RELATIONS: ALTERITY / BACKGROUND

Same camera. Different relations—depending on how you use it.

Week 5 / Original slide

18

WEEKS 4-6 / LANGUAGE, VISUAL AND AUDIO

Using AI, or incorporating AI.

WEEKS 4-6 / THE CREATIVE PROCESS

Using AI

AI as a tool in how you design: a brief drafted with a chatbot, a moodboard from a diffusion model, generative fill, layouts from Figma Make.

You stay the author. You become the curator, the briefer, the editor.

WEEKS 8-11 / THE PRODUCT

Incorporating AI

AI as a material in what you design: a feed, a recommendation, an assistant, a filter — the product decides something for each person, on its own.

You design a behaviour, not a picture. Data, bias, trust and accountability become design problems.

Week 1 / Original slide

19

01 · ELIZA · 1966

Pattern matching can sound like understanding.

Week 4 / Original slide

20

02 · TRANSFORMER · CAUSAL SELF-ATTENTION · VASWANI ET AL. 2017 ↗

One sequence. Many positions working together.

1 Tokens 2 Context 3 Training 4 Generation A robot robot writes writes a a brief brief . Parallel next-token predictions → compare with known targets Context: each position uses itself and earlier tokens only. Training works across positions together. Generation adds one token at a time.

Week 4 / Original slide

Open the live sketch › — choose a step · click a token · replay training or generation

21

03 · QUESTION ANSWERING

Retrieve evidence, then write one answer.

01 · QUESTION

What do I need to know?

A visitor asks: “What time does the event start?”

02 · SEARCH

Find a source.

Search the official event page for the schedule.

03 · EVIDENCE

Select a passage.

Bring the relevant text and its URL into context.

04 · ANSWER

Write from evidence.

Answer once; mark details the source does not give.

Week 4 / Original slide

22

04 · ONE TOOL CALL · SEARCH

The model requests. The harness executes.

MACHINE B · REQUEST

“Search the event page.”

Tool: search

Query: exhibition start time

→

MACHINE A · EXECUTE

Check, then run.

The harness checks access, runs the tool, and records the result.

→

OBSERVATION

Return the evidence.

The page text and URL enter the model’s context.

A tool call is a structured request to software outside the model.

Week 4 / Original slide

23

CLIP · CONTRASTIVE LANGUAGE–IMAGE PRETRAINING

CLIP brings words and images into a shared space.

CAPTION a chair with curved back TEXT ENCODER caption to vector IMAGE ENCODER picture to vector SHARED EMBEDDING SPACE matching pair different caption mismatched pairs move apart Similarity is not synthesis. A matching vector is not a picture.

Matching pairs move closer; mismatched pairs move apart. CLIP does not generate an image.

Week 5 / Original slide

24

CLASSIC LATENT DIFFUSION · GENERATION

Generation turns new noise into an image.

TEXT FEATURES NEW NOISY LATENT DENOISER learned prediction FINAL LATENT VAE DECODER OUTPUT IMAGE REPEAT LATENT UPDATE Decode after the last update.

The loop changes the latent, not the trained weights. Decode the final latent after the last update.

Week 5 / Original slide

25

CASE REFLECTION · SD2112 CO-CREATION

The brief emerged through choices.

01 · CLEAN MARK

02 · FOREST VARIANT

03 · BRIGHTER EDIT

→

→

A recognizable form to keep.

A setting and material enter.

A chosen refinement—not a finish.

The mark did not arrive fully formed. What entered the brief after seeing a candidate?

Week 5 / Original slide

26

CASE REFLECTION · SD2112 CO-CREATION

A better prompt was not always the right repair.

When the failure is in the transition, change the transition.

EARLIER ITERATIONS

Rejected montage.

Coherent origami repair.

Later, a burn to refine.

→

V8 · OBSERVED MISS

Flames inside the counter

and along the front edge.

→

V9 · TARGETED CHANGE

Replace only that transition:

front-facing, top → bottom.

Prompt repair? Reference repair? Workflow repair? Choose from the failure you can point to.

V9 delivered · review pending · not a claim of final acceptance

What did you delegate—and what did you insist on keeping?

Week 5 / Original slide

27

WEEKS 4-6 / MUSIC / MACHINE A

Musical choices in a sequencer

FIX

Tempo and bass

Set 100 BPM and repeat the same four-bar bass pattern.

VARY

Drum fills

At the end of each phrase, choose one of three fills or leave a rest.

LISTEN

Groove and space

Listen back. Does the fill lift the next phrase or crowd it?

Week 6 / Original slide

28

MACHINE B · TWO DOCUMENTED APPROACHES

A picture of sound. Or tokens of sound.

Historic Riffusion v1 (2022): a documented magnitude-spectrogram route DIFFUSION text + noise MAGNITUDE PICTURE generated spectrogram PHASE RECONSTRUCTION Griffin-Lim + inversion AUDIO waveform Magnitude alone is not enough: phase must be estimated before reconstructing audio. MusicGen (2023): a documented audio-codec-token route CODEC TOKENS EnCodec audio codes PREDICT TOKENS learned token predictor CODEC DECODE tokens to waveform AUDIO waveform Training audio is codec-encoded; generation predicts tokens. No claim about Suno internals.

Historical Riffusion v1 and MusicGen illustrate mechanisms—not Suno’s unpublished internals.

Week 6 / Original slide

29

MACHINE B · SUNO

Your demo can guide a new musical proposal.

Supply your own exported demo as audio reference; describe the transformation.

Listen to the actual result: arrangement, pacing, timbre, voices and surprises.

Keep a candidate because it serves the brief—not because it arrived first.

Suno · generate through the configured Easel integration in the workshop.

Public Suno landing page · 9 Oct 2026 · not the Easel interface.

Week 6 / Original slide

30

WEEKS 8-11 / AI INSIDE PRODUCTS / R8

People chose humans but rated AI higher.

WHAT THEY FOUND

About 57-62% of choices went to humans, but AI replies generally rated higher for empathy and feeling heard.

WHAT IT DOES NOT SHOW

Sources were chosen, not randomly assigned. Brief tasks, not established relationships or clinical benefit.

FOR YOUR DESIGN

Should users choose their source?

Wenger et al. (2026), Fig. 1; published PDF p.5.

Cropped plot; CC BY 4.0. Dashed line = equal choice.

R8 / Wenger, Cameron & Inzlicht, 2026

Week 8 / Original slide

31

AGENCY / CALIBRATED TRUST

Can the person disagree and act?

BEFORE

Know the boundary.

What does the system do? What evidence and data does it lack?

DURING

Question the answer.

Compare, correct, override or ask for a source before the decision matters.

AFTER

Leave or recover.

Keep an exit, a human route and a way to repair the consequences.

Week 8 / Original slide

32

FEEDBACK / CONCEPTUAL LOOP

The next screen changes the next data.

Possible feedback, not a measured universal law.

Week 9 / Original slide

33

READ THE SOURCE / RESULT + LIMIT / R4

Read the error row, not just the average.

DF/DM: darker female/male.

LF/LM: lighter female/male.

IBM error: All 12.1%.

DF 34.7% / LM 0.3%.

Within-group rates.

2017 binary classification.

Not identification.

Causes/current rates unknown.

Table 4 / PDF p.9 / PMLR 2018

R4 / Buolamwini & Gebru / Gender Shades, 2018 · CC BY 4.0

Week 9 / Original slide

34

WEEK 10 / DESIGN PROCESS

The Double Diamond

Discover and Define: understand the people and decide which problem the AI would address.

Develop and Deliver: compare interactions, test with people, and revise when the evidence changes.

Week 8 / AI in the design process

Design Council, The Double Diamond. CC BY 4.0.

35

THE RESULT / D

Better ratings. More similar stories.

Doshi & Hauser (2024), Fig. 2A crop, CC BY 4.0. Dots: differences from human-only; bars: uncertainty.

MAIN FINDING

AI access raised average novelty and usefulness ratings. The two AI conditions also produced more similar stories.

LIMIT

A short writing task, not studio learning; similarity does not count cultural voices.

FOR YOUR DESIGN

Inspect the portfolio, not only its best item.

Doshi & Hauser / Science Advances 2024

Week 11 / Original slide

36

COURSE RECAP / RECOMMENDERS, BIAS AND MEDIATION

Week 9 papers

Covington, Adams & Sargin (2016) / Deep Neural Networks for YouTube Recommendations

Gauthier et al. (2026) / The Political Effects of X's Feed Algorithm

Glickman & Sharot (2025) / How human-AI feedback loops alter human perceptual, emotional and social judgements

Buolamwini & Gebru (2018) / Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification

Gebru et al. (2021) / Datasheets for Datasets

Mitchell (1980) / The Need for Biases in Learning Generalizations

Verbeek (2015) / Beyond Interaction: A Short Introduction to Mediation Theory

Van Den Eede (2011) / In Between Us: On the Transparency and Opacity of Technological Mediation

37

COURSE RECAP / CURATION, AUTHORSHIP AND RESPONSIBILITY

Week 11 papers

Verbeek (2015) / Beyond Interaction: A Short Introduction to Mediation Theory

Doshi & Hauser (2024) / Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content

Gebru et al. (2021) / Datasheets for Datasets

Naqvi, He & Kaur (2025) / Catalyst for Creativity or a Hollow Trend?: A Cross-Level Perspective on The Role of Generative AI in Design

Joshi & Vogel (2026) / Interaction Techniques that Encourage Longer Prompts Can Improve Psychological Ownership when Writing with AI

U.S. Copyright Office (2025 report) / Copyright and Artificial Intelligence, Part 2: Copyrightability

38

AGENCY + AUTHORSHIP / ORIGINAL CONTRIBUTION MAP

Trace four different decisions.

01 / GENERATE

Tool proposes.

AI writes a draft from the source text.

→

02 / SELECT

Designer chooses.

Choose a draft that names its source.

→

03 / REVISE

Designer edits.

Check the facts and rewrite the text.

→

04 / AUTHORISE

Team releases.

A person approves it for visitors.

AI drafts. People choose, edit and publish. Keep a record of that work.

Week 11 / Original slide

39

FOUR DISTINCT QUESTIONS

Feeling, work, permission, law.

01 / FEEL

Does it feel mine?

A reported feeling is not a copyright certificate.

02 / CONTRIBUTE

Who did what?

Credit sources, selection, edits and review.

03 / PERMISSION

May we reuse it?

Check rights and consent for the material.

04 / PROTECTION

What is protected?

Human expression is a separate legal question.

Week 11 / Original slide

40

12:07-12:13 / THE DESIGNER'S ROLE

Designers as curators

OUTPUTS

What makes the cut?

Choose what fits the brief. Be able to explain why you kept, edited or rejected a result.

TRAINING DATA

Whose work shapes it?

Look at whose work and whose perspective the examples represent. Check permissions and gaps.

HUMAN OVERSIGHT

Who has the last word?

Decide where an agent can act and when a person must review, intervene or stop it.

41

12:13-12:19 / PROMPTS, RULES AND ACTIONS

Language as an interface

A phrase can specify a rule, as in a geometry instruction.

A description can steer a model toward a range of possible stories or images.

A music brief can do both. Fix the tempo and leave the arrangement open.

When the prompt can trigger an action, what is left implicit?

What should the interface show before a person lets the agent act?

Week 4 / The model requests, the harness executes

42

12:19-12:25 / CONTRIBUTION AND CREDIT

Fractional authorship

A team makes a poster with AI. Who shaped the idea, made the source material, selected the image, or edited the final composition?

Credit each contribution you can identify and describe what the model did. What should you disclose if the origins of training material are unclear?

Ask separately who gave permission to use the material and who holds rights in the finished work.

Week 11 / Trace four different decisions

43

12:25-12:27 / SYNTHESIS

What answer would you give Lovelace now?

One responsibility you will carry into your work.

One skill you will keep practising.

One question you still want to answer.

44

COURSE ARCHIVE

Earlier slides

Week 1 / Lovelace and the opening question

Week 2 / Rules and specifications

Week 3 / Learning from examples

Week 4 / Language models

Week 5 / Technological mediation

Week 6 / Sound and representations

Week 7 / Midterm recap

Week 8 / AI in the design process

Week 9 / Recommendation and feedback

Week 11 / Credit and responsibility

45

12:27-12:30 / ASK ME ANYTHING

Course AMA

Submit a short question about the course in ClassPoint.

A confusing concept, a case we discussed, or an ethical question you are still thinking about.

ClassPoint · short answer — answer on the projector

46

12:30 / STUDENT FEEDBACK

ESFQ

47

AFTER ESFQ / POSTER FAIR

Claim your poster-fair space

Open the Excalidraw room link shared in class.

One editor per group finds the exact group name and adds the project title.

Add a draft poster or placeholder. Landscape and portrait both work.

Add viewing links for the video and mediation brief when available.

Ask another group to find your space and open the links. Submit the final work in Canvas.

48

GROUP PROJECT / ASK ME ANYTHING

Group project AMA

Submit a short question about your group project in ClassPoint.

Ask about your design or evidence, the poster, video, mediation brief, Excalidraw, or Canvas.

ClassPoint · short answer — answer on the projector