August 2026 · Session 03 of 03
Opening · 5 minutes
Same product. Same empty state. One of these was written by a designer. One was generated. Which is which?
A
No items yet
Add your first item to start tracking your inventory.
Add item
B
Shelves are empty — for now
Add what you've got and we'll keep count from here. Takes about a minute.
Add first item
Hands up for A. Hands up for B. Now — one word for why.
The Reveal
01
Same model. Same session. Eleven seconds apart.
02
The only difference was the prompt. B carried a specification: who the user is, when they will read this, what the product must never sound like.
03
What you just detected was not human versus machine. It was specified versus unspecified.
Kosch et al., ACM TOCHI 2022 (N=369, N=100): participants told an adaptive AI was assisting them expected to perform better, and the expectation held after the task. No AI ever ran. Being told a machine was involved moved their judgment before they had evaluated anything.
So this session is not about learning to spot AI. It is about the judgment that survives not knowing.
Session 03 of 03
01
The Thinking Behind Design
Design is not decoration. Bad design creates debt you pay for later.
02
Design Thinking in Real-World Projects
A great-looking design still fails if it solves the wrong problem.
03
The Designer in the Age of AI
The machine makes good-looking free. Both of the above get harder, not easier.
Part I · 12 min
— John Heskett
The Strategy
The Labor
The Plan
The Output
Historically, 80% of our time was trapped in the labor and the output.
Human Mandate
Strategy and judgment become your primary currency.
Automated
Generative layouts, instant visual exploration.
Human Mandate
Directing the system architecture and curation.
Automated
Instant high-fidelity mockups and code generation.
The shift: We are moving from Makers to Editors & Directors.
“Roles are blurring as designers take on PM and engineering work, and vice versa.” — AI in Design Report 2026 (Designer Fund with Foundation Capital, 900+ responses)
“With the ability to build new functionality more quickly, teams may flood products with low-value features.” — NN/g
The New Mental Model
Why "Torque" matters more than "Speed" in the AI era.
The rate of motion. It is a scalar quantity, meaning it has magnitude but no inherent direction.
AI provides infinite speed. Generating 100 screens a minute is easy. But going 100mph in the wrong direction is just a faster way to fail.
The rotational force. It is a vector quantity. It is the leverage applied in a specific direction to move a heavy object.
The human provides the torque. It’s the strategic leverage, empathy, and accountability required to point the machine's speed at the correct problem.
Part II · 33 min
The Academic Consensus
Verganti, Vendraminelli & Iansiti · Harvard Business School Working Paper 20-091
Design in the Age of Artificial Intelligence
When execution is free, the human bottleneck shifts to four pillars.
“Preserving judgment, taste, and skill development.” — AI in Design Report 2026, on craft
Making sense of the human context to choose the exact right problem to solve, and carrying accountability for the outcome.
Writing intent as an executable gear mesh. Being precise and nuanced enough that a machine cannot build the wrong thing.
When high-fidelity output is produced faster than anyone can review, you must know what to sample, trust, and let pass.
Designing the system loops that run and evaluate themselves, moving away from driving the AI turn by turn.
01
Outer loop: Understand
Building the Skillset: Torque (1/2)
AI is trained on millions of standard UI patterns. It knows exactly what a checkout flow should look like.
However, AI doesn't know why humans abandon a flow. You must learn behavioral economics (trust, cognitive load, anxiety) to make sense of user behavior.
The Prompt: "Generate a high-converting checkout page."
AI Output: A sleek, minimalist 1-click Shopify-style layout. Conversion tanks.
Human Torque: You recognize this is a $10,000 B2B software purchase. Users are anxious, not impatient. You dictate adding robust trust signals, security badges, and an ROI calculator. Conversion jumps.
Building the Skillset: Torque (2/2)
When stakeholders ask, "What should we build?", designers must use sensemaking to shift the question to, "What behavior are we trying to create?"
AI makes it dangerously easy to just build exactly what was asked for, even if it's the wrong thing.
The Request: "Users aren't engaging. Redesign the dashboard."
Speed (AI Only): Generates 10 beautiful dashboard UI variations in 60 seconds. You ship it. Engagement still drops.
Torque (Human): You run 3 user calls and realize users aren't logging in because the notification emails are broken. You direct the AI to rewrite the email logic, not the UI.
02
Outer loop: Frame intent
Building the Skillset: Specification (1/2)
Historically, we judged an artifact by the human labor it required. Today, a polished UI takes three seconds to produce.
You must shift from passive consumer to active interrogator. What datasets informed this? What constraints or system prompts shaped its boundaries? Seeing invisible mechanisms prevents us from taking AI at face value.
Passive Artifact: "Generate a user profile card."
Active Architecture:
Building the Skillset: Specification (2/2)
AI is a maximalist by default. If you give it freedom, it will over-design and hallucinate unnecessary features.
Often, the best way to control an AI agent is by defining the absolute guardrails first. Tell it what it cannot do.
Instead of asking for a design, write parameters a machine cannot violate.
03
Outer loop: Verify & decide
The Verification Paradigm
1972
What it destroyed
The unique "aura" of art. A painting's meaning was no longer fixed to the one wall it hung on.
What it opened
Images travelled to people instead of people travelling to images. Anyone could set them side by side and argue with them. Reproduction made imagery a language ordinary people could speak.
2026
What it destroys
The "aura" of human technical labor. A polished artifact no longer proves skill, time, or intent.
What it opens
Anyone who can describe an intent can produce the artifact. The design vocabulary leaves the profession, and the number of people who can argue visually goes up by orders of magnitude.
Berger's second move: that same language was captured by publicity. His question was never "is this authentic?" — it was who is using the language, and for what.
To develop better judgment, we must shift from being passive consumers of outputs to active interrogators of systems.
Building the Skillset: Verification (1/2)
When technical execution drops to zero, human curation is the differentiator.
Machines produce the mathematical average of their training data—resulting in technically perfect but sterile outputs.
Your job is to look at ten machine-generated variations and identify the single one that genuinely connects with human emotion. You must actively reject the average.
The Output: AI instantly generates a gorgeous, translucent "glassmorphism" UI with heavy blur filters.
Passive Acceptance: "Wow, this looks beautiful. Let's ship it."
Vibe Engineering: You know heavy CSS backdrop-filters feel sterile and will tank performance on low-end Androids. You reject the AI's aesthetic and force a highly-intentional, performant flat design.
Building the Skillset: Verification (2/2)
As synthetic media becomes indistinguishable from reality, asking "Is this real?" is a losing battle. We must shift from passive consumers to active interrogators.
Instead of evaluating how an artifact was made, we must interrogate why it was generated and what behavior it aims to elicit.
| Traditional (Creation) | Modern (Generation) |
|---|---|
| How was this physically made? | Why was this output generated? |
| Is this an accurate representation? | What behavior is it designed to elicit? |
| Who is the author/artist? | Who controls the model's worldview? |
04
Outer loop: Learn
Building the Skillset: Orchestration (1/3)
When code is no longer the constraint
Because execution is free, designers are absorbing engineering and PM roles. You must have a working knowledge of APIs, databases, and latency.
If you don't understand how data flows, you will orchestrate impossible systems.
The Request: Design a live-updating dashboard.
AI Output: Generates a beautiful UI assuming data loads in 0.01 seconds.
Human Orchestration: You know this specific legacy database takes 4 seconds to query. You reject the AI's instant-load UI and mandate the inclusion of "optimistic updates" and skeleton loaders to manage user perception during the wait.
Building the Skillset: Orchestration (2/3)
AI outputs are adequate, but they cannot explain tradeoffs, justify decisions, or connect solutions to the bottom line.
Orchestrating a system means using emotional intelligence and business vocabulary to align human stakeholders around the why.
Designer Talk: "We need to fix this onboarding flow because the cognitive load is too high and the UI is messy." (Founders tune out).
Business Storytelling: "We are seeing a 30% drop-off at this step. By deploying this new AI-generated flow, we can recapture 10% of those users, which translates to $50k in retained ARR this quarter."
Building the Skillset: Orchestration (3/3)
Instead of viewing isolated screens, human judgment requires a holistic view. You are taking disparate, machine-produced components and weaving them into a cohesive product strategy.
The machine produces the raw material, but human orchestration dictates whether that material is useful, truthful, or meaningful.
Orchestrators actively look for the gaps that machines cannot see:
Scaling Orchestration
The industry is currently obsessed with the bottom layer: writing the perfect prompt.
But to orchestrate a true AI system, a designer must move up the stack. Your value multiplies as you move from designing a single artifact to designing the entire ecosystem.
*Adapted from AI Engineering Architecture
Service design. Wiring multiple touchpoints and actors to hit business outcomes.
Designing the continuous UX loop (User Acts → AI Reasons → UI Adapts).
Establishing UX guardrails, accessibility constraints, and fail states.
Injecting the Design System, user state, and brand guidelines.
Generating a single specific UI component, screen, or copy asset.
Part III · 8 min
The AI Trap
If you give an AI a generic business prompt, it will give you a perfectly acceptable, completely boring output.
AI predicts the most mathematically probable next word or pixel. By definition, it generates the average of all human thought.
Result: A dark UI, a picture of a guy running, a headline that says "Unleash Your Potential." It looks exactly like every other app on the market.
The Antidote
How to beat synthetic sameness using the pillars we just learned.
Don't ask for a UI. Ask for a behavior.
Provide the AI with real-world context, emotional tension, and the user's psychological state. Give the prompt a specific direction.
Be a ruthless editor up front.
Define the exact constraints. Tell the AI what it cannot do. Force it to avoid clichés by establishing strict rules for the output.
Part IV · 20 min
Live Exercise (20 Minutes)
You are designing the onboarding flow for a B2B SaaS tool that helps restaurants manage their inventory. The users are exhausted chefs and managers doing this at 2:00 AM after a 14-hour shift.
Do not use the generic prompt: "Design a restaurant inventory onboarding."
Write a new prompt that beats synthetic sameness. Inject the Torque (the emotion of an exhausted chef) and set the Specifications (the exact constraints the AI must follow).
Pair up. You have 15 minutes to write. 5 minutes to share.
Part V · 10 min
Closing & Q&A
AI just handed everyone the tools. The tools became everyone's. The judgment did not.