I recently completed the Interaction Design Course titled “AI for Designers”. Here are some of my notes.
I have also included some questions at the end that helped with understanding the content.
.jpg)
- There is a difference between narrow AI and general AI
Narrow AI
The AI system is trained to do specific things within boundaries. When the AI system veers off the boundaries, it starts to break down and provide undesirable outputs.
General AI
The system aims to perform across many domains with human-like capabilities.
This means it should have “transfer”. It should learn something in one area and apply it reliably in a new one. It can plan, adapt, and handle unfamiliar situations without training wheels every time. We do not have general AI (yet).
[I think the instructor is referring to AGI here. The definitions get muddy depending on whom you’re speaking to]
- As designers, we must separate between designing for AI and designing with AI
Designing for AI
The designer is designing for the user experience of an AI powered product or feature in an existing product. These systems are probabilistic not deterministic.
Designing with AI
Using AI to speed up or improve your design process. Figma has a lot of these tools injected into the product and there are many tools out there that can help designers with their design process
- There is more to prompting than making simple requests.
Other things matter like order and structure of the prompt, one shot/few-shot techniques, chain of thought, iterative prompting is better than getting it in one go etc.
Simple requests are vague and lead to generic or non-useful outputs.
When designing with AI, increasing familiarity with the tools helps so prompting becomes second nature. The framework I like to use is the CARE from NNG.
- AI has come a long way in user research but it still does not understand visual cues, body language and certain nuances
AI is great for research in text (as well as speech to text) format. For research it can help with:
Transcription and clean-up of interviews, thematic clustering across lots of notes, tagging, summarising long sessions, finding patterns at scale in support tickets, and many other similar aspects
AI fails and needs help on the following:
Behavioural interpretation, sensitive moments, power dynamics, exploring contradictions, deciding what is real versus noise, and making product decisions.
- There are four types of bias to be aware of:
Systematic/systemic
Bias from the world we operate in. This includes historical inequality and uneven access
Statistical
Bias from the data itself. It can come from sampling bias, label bias, measurement bias and so on
Computational
Bias from the technical choices made when building the system. This includes the wrong objective functions, ranking decisions, type of architecture and feedback loops.
Human bias
Bias from the people building the system. Bias from designers, researchers, labelers, engineers and so on.
- Knowing when to summon AI. Users may not want AI “all the time”, so interruption avoidance matters.
If AI pops up mid-task, it can come across as very distracting to a user. It can also create unnecessary cognitive load as well as portray the wrong idea. E.g Asking if someone needs assistance can be obtrusive and may be offensive.
The material also mentions HAX Guidelines
HAX Guidelines: https://www.microsoft.com/en-us/haxtoolkit/library/
Stem: AI is defined as a technology with two traits. Which option correctly matches each trait to its meaning?
A. Autonomy: improving from feedback; Adaptivity: acting without any environment model
B. Autonomy: performing tasks without constant user guidance; Adaptivity: improving performance by learning from experience
C. Autonomy: predicting future events; Adaptivity: following explicit rules with no learning
D. Autonomy: generating creative outputs; Adaptivity: applying fixed heuristics consistently
Stem: Which prompt method is described as encouraging the model to explain its reasoning, rather than only producing an answer?
A. Few-shot prompting
B. Role prompting
C. One-shot prompting
D. Chain-of-thought prompting
Stem: In the Microsoft HAX guidelines, which guideline describes clarifying user intent before executing when the system is unsure?
A. Scope Services When in Doubt
B. Show Contextually Relevant Information
C. Support Efficient Correction
D. Match Relevant Social Norms
Stem: You have a highly polished screen design, but you want usability feedback focused on task flow rather than branding and aesthetics. What is the strongest rationale for converting the high-fidelity design back into a wireframe before testing?
A. A wireframe lets you measure visual delight more reliably because branding is removed
B. A wireframe prevents users from forming any first impressions, improving validity
C. A low-fidelity wireframe reduces distraction and encourages candid, task-focused feedback
D. A low-fidelity wireframe removes the need for any further iterations after the test
Stem: A team uses a predictive heatmap tool and gets a clarity score of 58/100 (“moderate difficulty”). They propose skipping real user testing to save time. Which response best matches what they should do?
A. Accept the score as empirical proof and proceed to build, since the tool quantified clarity
B. Use the output as a pre-test signal, interpret it critically, then validate with real users
C. Discard the output entirely, because AI evaluation tools provide no value in any scenario
D. Replace the test with more prompts until the clarity score becomes high enough
Stem: You simulate an interview with an AI-generated persona to explore a problem space. What use of the output best aligns with what you learned?
A. Treat the interview as representative qualitative data and proceed directly to solutions
B. Use the interview as definitive themes, then validate only with analytics after launch
C. Use the interview to replace early discovery research, then do usability testing later
D. Use the interview as provisional pointers, then challenge and validate with real users
Stem: Imagine a next-generation research tool that (1) can process full-session video (visuals + behaviour) and (2) accepts study goals, participant context, tasks, and past research. Which limitation would still remain largely unresolved unless separately designed for?
A. The difficulty of verifying accuracy if outputs lack citations and blur notes vs participant data
B. The inability to observe user actions because the tool cannot process visual input
C. The absence of research context, leading to generic summaries unrelated to the study goals
D. The risk that transcript-only analysis misses actions because participants do not verbalise everything
Stem: A team shifts from a rigid, command-by-command automation flow to a goal-based AI assistant where users state outcomes and the system works out steps. According to the materials, which statement best captures what changes and what must be designed differently?
A. Nothing fundamental changes; only the UI copy needs to be re-written to sound friendlier
B. The system becomes general AI, so trust patterns and safeguards become less necessary
C. The interaction becomes less controllable upfront, so expectation-setting and error handling become central
D. The user no longer needs to provide context, because the system can infer intent from goals alone
Stem: To address a hyper-personalisation risk, a music app occasionally recommends a song that does not match the user’s profile “to introduce them to something different.” Which concern is this tactic primarily meant to mitigate?
A. Privacy risk from data collection
B. Fatigue from being repeatedly shown similar content
C. Bias risk from training data and content exposure
D. Accountability risk from unclear ownership of outcomes
Stem: Consider this argument about AI research tools:
Stem: A designer is prompting Midjourney for a logo. After a mediocre result, they keep appending more descriptors (styles, lighting, multiple metaphors, and extra brand adjectives) and conclude: “More detail always improves results.” What is the best correction for this?
A. Keep prompts specific, but avoid over-describing; refine through focused iterations
B. Add even more detail until the prompt becomes fully exhaustive and unambiguous
C. Stop iterating, because the first output is the most reliable output for branding
D. Replace prompting with transcript analysis, since visuals are too subjective to direct
Stem: A productivity app adds proactive AI push notifications suggesting rewrites. It sends them even when it detects the user is driving, arguing: “We followed the HAX guideline about timing services based on context by notifying them immediately.” What change most directly fixes the first failure?
A. Add a disclaimer that the suggestions may be wrong
B. Show an explanation of the model’s reasoning after each push notification
C. Postpone or suppress notifications when the user context suggests they are busy or driving
D. Increase the number of proactive notifications so the user sees more options
Stem: Which statement most accurately distinguishes AI “insight generators” from AI “collaborators” in research tools?
A. Insight generators accept contextual inputs, while collaborators are transcript-only systems
B. Collaborators accept some researcher-provided context and can support tagging/themes beyond transcripts, while insight generators summarise transcripts alone
C. Insight generators are general AI systems, while collaborators are narrow AI systems
D. Collaborators eliminate bias by design, while insight generators mainly suffer from usability issues
Stem: Two predictive attention tools appear similar in features, but the materials highlight a key difference in their underlying training basis. Which option matches that distinction?
A. Predict by Neurons uses only survey sentiment, while Attention Insight uses only behavioural friction cues
B. Attention Insight extends to physical environments, while 3M VAS is limited to screens
C. Predict by Neurons is trained on institutional policies, while Attention Insight is trained on demographic personas
D. Attention Insight bases predictions on eye-tracking studies, while Predict by Neurons uses eye-tracking plus neuroscientific experiments
Stem: You are designing an AI feature for workplace decisions and want the minimum set of UI commitments that simultaneously supports transparency, accountability, fairness, and privacy as framed in the materials. Which option best satisfies all four at once?
A. Consent + data deletion controls; explanation of why outputs were produced; clear correction/undo paths; in-product reporting of biased outcomes
B. Consent + data deletion controls; friendly tone and social norms; prompt templates; proactive suggestions that cannot be dismissed
C. Explanation of why outputs were produced; confidence score; faster generation; removal of all user controls to reduce misuse
D. Clear correction/undo paths; reporting of biased outcomes; faster onboarding; no disclosure of limitations to avoid distrust
Stem: You are designing an AI writing assistant. You want it to (1) teach users how to get better outputs over time, (2) learn from users, and (3) stay non-intrusive with strong user control. Which option best fits?
A. Always-on proactive suggestions; no dismissal controls; a single onboarding tooltip; no feedback capture
B. Full-screen AI wizard on every document open; hidden prompt templates; correction via support tickets only
C. Prompt templates and persistent custom instructions; explicit feedback capture; manual invocation and easy dismissal/correction
D. Automatic rewrites of user text; delayed explanations; no undo; novelty suggestions to reduce fatigue
Stem: A PM argues: “Disclaimers and limitations messaging reduce adoption. Since AI can hallucinate, we should hide uncertainty and present outputs confidently.” Which critique is strongest under the trust guidance?
A. Hiding limitations is fine as long as the system is narrow AI, not general AI
B. Confidence is sufficient if the model is updated frequently and users never see changes
C. Removing disclaimers is acceptable if users can report issues through an external form
D. Trust is built by setting realistic expectations about what the system can do and how well, alongside transparency and control
Stem: A team wants to use an LLM to generate a long-form help article about a stressful account-recovery process. They plan to ship it with minimal human editing because “AI writing tools are powerful.” Which critique best matches?
A. This is fine because long-form content is more predictable than microcopy in most LLMs
B. Long-form output is less predictable and AI lacks empathy, so humans must contextualise and humanise the content
C. This is fine if the LLM is role-prompted as “an empathetic support agent”
D. This is fine if the team adds more examples, because examples eliminate inaccuracies
Stem: You want to run a fast, AI-augmented design loop for a new mobile productivity app while staying faithful to the cautions you learnt. Which plan is most aligned?
A. Generate personas and interviews with AI, accept the themes as ground truth, then ship the top Uizard concept without user testing
B. Use a transcript-only insight generator for discovery, then rely on predictive heatmaps as the main validation signal before launch
C. Start broad and iteratively refine prompts for problem-space exploration; define a clear feature list; generate starter wireframes; use AI to explore multiple visual concepts; pre-test with predictive tools; then validate with real users and iterate
D. Skip wireframes to avoid biasing layouts; go straight to Midjourney marketing visuals; infer usability from engagement with ads
Stem: You are designing an adaptive UI for complex CAD software that learns from behaviour to surface likely commands. You must address hyper-personalisation concerns and HAX-style interaction safeguards. Which bundle best matches?
A. Seek explicit consent; explain data use; offer data deletion; continually audit/refine for bias; add occasional novelty to reduce fatigue; provide manual invocation/dismissal/correction; notify users when the model changes
B. Collect behavioural data silently; optimise only for speed; hide uncertainty; rely on a single “report issue” link; keep the model update process invisible
C. Ask for consent once; avoid novelty; remove dismissal controls to reduce interface clutter; rely on confidence scores as the main safeguard
D. Use demographic personas as the main training input; disable correction to prevent “user tampering”; only show context when the system is fully certain