10 Design Waypoints from an HCI Researcher

Designing with generative AI means designing for cognition. These systems don’t just complete tasks, they shape how users think, trust, and remember. But many of the most consequential design failures in AI live beneath the surface: at the edge of interpretation, where interface becomes epistemology. This is where traditional UX breaks. Not because the tools are outdated, but because the questions have changed.

This essay doesn’t propose one unified model. Instead, it isolates key breakdowns in the generative loop: when prompting becomes performance, when fluency masquerades as depth, when users stop thinking. Each waypoint is a design moment, not a problem to fix, but a seam to explore. Together, they map the silent places where meaning slips. And where design must reenter.

Not Everything Should Be a Chat

Conversational Design Is Not a Universal Pattern

After the rise of ChatGPT, conversation became the default UX pattern for AI products. Interfaces once built for browsing, composing, or planning were rapidly reformatted into chat threads. But a chat isn’t just a container, it’s a cognitive claim. It assumes memory, flow, and responsiveness. And when those fail, it’s not just usability that breaks, it’s understanding.

Conversational interfaces often fragment tasks that require overview, revision, or reflection. Users are nudged to stay within the thread even when the task demands stepping out. Linear chat simulates coherence but doesn’t always support it. The result is subtle: users misremember, disengage, and defer to the system. They don’t feel interrupted. They just feel tired.

Designing AI systems with chat as the default flattens every experience into the same interaction rhythm. It hides structural decisions inside aesthetic simplicity. The problem isn’t conversation itself—it’s assuming every interaction should be one.

Supporting Research

Research affirms this strongly. Wang & Li (2022) show that users in multi-turn chat interactions often lose track of context and struggle to build global coherence across turns strongly supporting the claim that chat interfaces disrupt structured cognition. Schuetz & Venkatesh (2020) argue that conversational AI systems shift epistemic control from users to systems, blurring the boundary between interface and actor conceptually supporting the need to rethink chat as a neutral UX. Rzepka & Berger’s (2018) systematic review finds that users over-trust polished output early on, but disengage over time due to fatigue, misalignment, and anthropomorphism—empirically supporting concerns about chat fatigue. Park et al. (2023) research support with nuance: they find that LLM chat doesn’t scaffold reflection or exploration across prompts. Balakrishnan et al. (2021) add depth around why showing that trust and engagement depend on unbroken flow, and when conversational rhythm falters, users exit. Cognitive immersion is fragile.

This isn’t a minor usability issue. It’s a structural limitation with ethical consequences. Users misremember, misattribute, and misjudge when interfaces perform conversation but don’t support thought. The thread becomes the trap.

Design Scenario

A civic research team was using ChatGPT to summarize feedback from a series of community interviews. The interface was familiar: a single prompt bar, a running chat thread. They typed: “Summarize key points from stakeholder interviews,” pasted in content, and hit enter. The result was fluent, clean, structured. They skimmed, nodded, moved on. But in the debrief, gaps emerged. Participants misremembered what had been said. They assumed consensus where none existed. One team member later admitted, “It sounded right, so I didn’t question it.” The issue wasn’t hallucination. It was subtle displacement: the format looked conversational, but it didn’t support the recursive, comparative thinking the task required.

In another session, a researcher tried to revisit earlier prompts mid-task. “I wanted to compare what I asked before, but scrolling made me forget what I was even doing.” To cope, they copied chunks into Notion, saying, “I just needed a space to think.” The AI wasn’t wrong. The interface was too linear.

One product manager described it clearly: “It’s always a straight line. But my thinking isn’t.” Users weren’t passive. They were adapting. Threaded chat wasn’t helping them track, compare, or revise. Across sessions, users didn’t reject the AI’s answers. They deferred to them—not because they agreed, but because the format didn’t make it easy to ask again.

Design Element(s)

We don’t need to abandon conversation entirely, but we need to break the thread. The following design elements propose interaction formats that reintroduce structure, memory, and reflection into AI interfaces:

• Stacked Views
  Instead of linear turns, responses are layered in columns or cards, allowing for visual comparison and overview.

• Prompt History Maps
  Visualizations of prompt evolution—where users can branch, remix, or retrace paths.

• Multi-Reference Prompts
  Let users highlight multiple past turns and generate from them—mimicking citation, not just sequence.

• Intent Markers
  Users declare what kind of response they want (e.g., summary, critique, expansion), and the UI adapts accordingly.

• Threaded by Idea, Not Time
  Allow users to group and tag exchanges by concept or task, rather than defaulting to timestamp.

These aren’t gimmicks. They are affordances for cognition. They treat the interface not as a chatroom, but as a workspace—a thinking environment.

Design Reflection

This waypoint asks us to treat chat not as a neutral frame, but as a cognitive architecture. When designers default to conversational UIs, they’re not just choosing a format. They’re scripting a mental rhythm. The illusion of flow replaces the reality of thought. Tasks that demand exploration or structure get flattened into serial exchange. Users don’t opt out. They adapt. But that adaptation often means deferring judgment, losing track of questions, or mistaking interface rhythm for intellectual progress.

What this tells us is simple: chat isn’t always a conversation. It’s often a tunnel. And when we treat it as the universal pattern for AI UX, we stop designing for how people actually think—across time, across tasks, across contexts.

So the question isn’t “how do we improve chat?” The better question is: “When is conversation the wrong form altogether?” And what would it mean to design AI that supports nonlinear, comparative, self-directed cognition—something more like thinking, and less like talking?

Prompt-as-Design

Prompting isn’t pre-design; it is design. The moment a user types a prompt, they are shaping the epistemic contract of the interaction: setting context, defining tone, implying goals. But today’s AI systems demand that users be both librarian and reader—crafting precise queries and simultaneously deciphering the returns. This interaction model assumes not just literacy, but articulation, foresight, and rhetorical clarity. These are not neutral demands. They reward those fluent in abstract, linear, text-based communication and exclude those who are not.

As Nielsen (2023) highlights, fewer than 20% of users in high-literacy countries are capable of consistently writing effective prompts. The prompt interface thus becomes a gatekeeper, not a bridge. What’s more, it externalizes the cognitive burden: users must anticipate outcomes, engineer phrasing, and assess results—often with little feedback. Treating prompting as transparent masks the deep asymmetries it introduces. When language becomes the interface, design becomes a literacy test. Rather than universalizing the prompt, we must ask: is this even the right design direction?

Supporting Research

Nielsen (2023) identifies the “articulation barrier”: a usability chasm where most users cannot clearly express complex needs in written form. He draws from OECD’s PIAAC literacy study to argue that writing effective prompts is even harder than reading—and that reliance on this interface model excludes a majority of the population. The existence of “prompt engineers” as a role further proves the cognitive cost involved (Nielsen, 2023, pp. 2–4).

Lee and Palmer (2025) extend this critique in education, noting that while “prompt literacy” is being introduced into classrooms, it is often divorced from frameworks of critical thinking or co-design. Walter (2024) likewise stresses the lack of evaluation standards for prompts, suggesting we need more than stylistic fluency—we need epistemic scaffolding. Across studies, prompting emerges not as a skill users lack, but as a design failure that over-indexes on one kind of interaction competence.

Nielsen, J., 2023. The articulation barrier: Prompt-driven AI UX hurts usability. UX Tigers. Available at: https://jakobnielsenphd.substack.com [Accessed 3 July 2025].

Lee, E. and Palmer, J., 2025. Prompt engineering in higher education: A systematic review. International Journal of Educational Technology in Higher Education, 22(4), pp. 5–15.

Walter, A., 2024. Thinking through prompting: Critical frameworks for AI literacy. Design Literacy Review, 3(1), pp. 17–25.

Design Scenario

In a Gulf-region vocational training program, administrative workers were introduced to ChatGPT to streamline reporting. The instruction? “Ask it to draft a project update.” For trainers, this seemed intuitive. For the participants—many of whom were mid-career and navigating second-language environments—it was paralyzing. What tone? What scope? What details were needed? The blank box became a site of tension. Some typed overly vague questions. Others overcompensated with formal paragraphs. In both cases, the output was misaligned. More worryingly, the experience reinforced a feeling that AI was “for others”—that they weren’t articulate enough to use it well. This wasn’t a failure of training. It was a failure of interface imagination.

Design Element(s)

The prompt box appears simple, but it's a dense space of expectation. It demands precision but provides little feedback. Nielsen proposes hybrid models that blend GUI and prompt flows—buttons, templates, conditional forms. Lee and Palmer suggest prompt “palettes” or templates based on task types. More radical proposals include co-prompting agents, multi-modal builder interfaces, or even visual prompting via flowcharts. We must also interrogate the underlying assumption: must prompting be the default interface at all? Why not conversational agents that build the prompt with the user over time? Or design metaphors drawn from librarian reference interviews, where clarification and narrowing are part of the experience? When prompting becomes embedded in interaction flows—not isolated in a blank field—it becomes more than a command. It becomes designable.

Design Reflection

What if we abandoned the prompt field altogether? Not all users want to be authors. Many want to explore, tinker, or co-navigate. Treating prompting as inevitable limits us to one paradigm: articulation. But interaction can also begin with selection, sketching, example-browsing, or guided goal-setting. The librarian metaphor is apt—we’re not just retrieving knowledge, we’re constructing meaning. But librarians don’t hand you a search bar and walk away. They ask, interpret, triangulate. Could AI do the same?

We need to stop asking how to make prompting easier and start asking if it should be central at all. Prompt-as-design is not just a UX problem. It’s a design philosophy problem. And it deserves to be rethought from the ground up.Fluency Isn’t Intelligence

Cognitive Offloading Is Real

AI promises efficiency, but it often delivers disengagement. As users interact with AI systems, a subtle shift occurs: they begin to defer, not decide. AI promises efficiency, but it often delivers disengagement. As users interact with AI systems, a subtle shift occurs: they begin to defer, not decide. This isn’t laziness, it’s cognitive offloading (the mental process of delegating thinking to the system, often unconsciously). People naturally offload tasks to reduce effort. But in offloading, they often stop evaluating. This shift from engagement to automation is especially pronounced in AI-supported tasks that appear complete. Once a coherent answer is generated, the user assumes it's good enough. But here's the design failure: users rarely read deeply. They skim. They scan. This has been true long before the internet. So even when AI generates longform output, the cognitive return on that generation is often low.

There’s an asymmetry at play: the AI generates more than the user consumes. Worse, I’ve observed that users often comprehend less than if they had constructed the answer themselves. This creates a design tension: should we give users more — or help them do more? When generation becomes a shortcut rather than a scaffold, thinking erodes. That’s not just inefficient. It’s unethical. A good design doesn't just produce output — it produces understanding. Offloading is not neutral. It must be managed, nudged, even resisted, if we care about long-term agency.

Supporting Research

Varma et al. (2024) at MIT used fMRI imaging to study the neural effects of using AI for information tasks. They found that participants showed significantly reduced activity in regions associated with cognitive control when viewing AI-generated answers — even when those answers were wrong. This supports what cognitive science has long claimed: humans offload effort when the system appears authoritative. Nielsen’s usability heuristics reinforce this: “Recognition rather than recall” and “Visibility of system status” are meant to aid engagement, not replace it. Yet AI systems often violate “User control and freedom” and “Error prevention” by presenting polished answers without inviting critique.

From my practice as an HCI researcher, I have observed a crucial gap: the UX research toolkit isn’t built to capture cognitive offloading. Traditional metrics like task completion, satisfaction, or usability scores fail to reveal whether users actually understood, engaged with, or benefited cognitively from AI-generated content. In my view, we need a new kind of design instrumentation — one that can assess the relationship between effort, content quality, and comprehension. This isn’t just a research opportunity; it’s a foundational requirement for ethical AI design. We need frameworks that measure not just what users receive, but what they retain — and where thinking gets displaced.

Varma, S., Dastin, J., Choi, J. and Gershman, S., 2024. Delegation effects in human-AI collaboration: A neurocognitive perspective. MIT Cognitive Systems Lab. [Unpublished Manuscript].

Design Scenario

In my work with civic design teams, researchers often turned to AI to summarize user interviews. What began as a time-saver quickly became a norm. “We don’t need to read the transcripts — the AI already did.” But in post-session debriefs, I noticed something crucial: nuance was lost. Contradictions disappeared. What remained was a pleasant, generic synthesis. Teams stopped listening. They trusted the summary — even when they didn’t agree with it. This offloading wasn’t passive. It was systemic. Without friction, the interface shaped behavior: read less, check less, trust more. It was a cognitive convenience trap — and it was invisible.

Design Element(s)

Most AI tools are designed for speed and flow. Summaries appear instantly. Answers are presented confidently. But frictionless output is dangerous. Nielsen’s heuristic “Help users recognise, diagnose, and recover from errors” is absent in most LLMs. There are no affordances for second-guessing. To counter this, we need reflective friction (deliberate design interruptions that nudge users to slow down or reconsider) micro-interactions that pause the user: “Would you say it this way?” or “Highlight what you agree with.” We also need engagement-sensitive interfaces: for example, adaptive interfaces that shorten outputs based on user scanning patterns, or versions that require users to fill in blanks before proceeding. Another direction is reverse prompting — systems that ask users to respond before revealing their output, turning the interaction into a test of alignment, not a delivery of authority.

Design Reflection

This waypoint calls for a new design responsibility: not just delivering answers, but preserving engagement. What if interfaces tracked how often users skipped, ignored, or blindly accepted output? What if we tested for effort–quality–comprehension as a triad, not a tradeoff? Current UX practice isn’t equipped to measure this. We need new tools: epistemic friction maps, comprehension deltas, offloading indices. Without these, we risk designing systems that seem helpful but quietly erode skill, agency, and curiosity.

Cognitive offloading will happen. That’s inevitable. But designers choose how much, when, and what’s lost. And if we believe that design should improve human experience, then reducing critical thinking isn’t a feature. It’s a failure.

Fluency Isn’t Intelligence

In AI-generated systems, fluency smooth, grammatically polished output is often mistaken for depth. This is not just a novice trap. Studies show that even experts are susceptible to this bias, rating articulate AI responses more highly than accurate human ones. This challenges a long-held UX assumption: that expert users apply sharper scrutiny. Here, both novice and expert users fall for the same design illusion.

This bias reflects more than just a user interface failure; it’s a cultural and epistemic distortion. Our cognitive infrastructure favors coherent language as a shortcut for trustworthiness. But fluency is a performance, not a proof. When we design AI systems without signaling uncertainty or providing epistemic scaffolds, we inadvertently elevate surface credibility over grounded comprehension. This doesn’t just mislead, it disorients users at the very moment they need to evaluate meaning and decide whether to trust, question, or verify.

Supporting Research

Shojaee et al. (2025) demonstrate that GPT-4’s outputs were rated more favourably than expert human answers based not on correctness, but on coherence and confidence of tone. This held true even among AI researchers, indicating a widespread susceptibility to fluency bias (Shojaee et al., 2025, pp. 7–10). Similar results have been reported in medical AI studies, where patients preferred incorrect but fluent responses to dry, accurate ones. These findings echo Fogg’s earlier work on interface credibility, confirming that stylistic presentation strongly shapes user trust, especially under uncertainty.

Shojaee, P., Mirzadeh, I., Alizadeh, K., Horton, M., Bengio, S. and Farajtabar, M., 2025. The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity. Apple Machine Learning Research. Available at: https://ml-site.cdn-apple.com/papers/the-illusion-of-thinking.pdf [Accessed 3 July 2025].

Design Scenario

In higher education, educators increasingly report that essays written with generative AI are well-structured and fluent yet hollow. Argumentation is weak, evidence sparse, citations non-existent. Instructors describe receiving papers that read like A-grade submissions, but fall apart under scrutiny.

I once observed a grading panel where tutors debated whether to flag a student’s essay for AI use. The tell-off wasn’t a citation error — it was fluency. The essay was too polished, too balanced, too “finished” compared to the student’s usual voice. But when they looked closely, the argument wasn’t outright flawed — just shallow. Still, even the best students sometimes write shallowly. Argumentation depth isn’t diagnostic. Citation hallucinations? Yes. Perfect structure? Not enough.

The panel had no real evidence — just a hunch. But fluency had created a fog. It made the ordinary seem extraordinary. The evaluation faltered not because the essay was wrong, but because it was too right in all the wrong ways. That is the design consequence: when polish confuses perception, judgment drifts from fact to feeling.

Design Element(s)

Most LLM interfaces present output in clean, confident blocks of text — markdown-perfect and aesthetically authoritative. But this visual fluency masks epistemic uncertainty. There are no signals for doubt. No gradients of confidence. No traces of origin or fragility.

We can — and must — disrupt this. Design elements that counteract fluency bias might include:

• Confidence annotations for each claim

• Source-trace toggles that reveal where ideas came from

• “Speculative mode” phrasing for unverified content

• Colour-coded evidence trails, indicating fact-check status

• “Model Doubts” buttons that surface known uncertainties

Even subtle friction — like stylistic hesitations, layered views, or transparency on model assumptions — can slow over-trust and provoke interrogation.

We must design for sense-checking, not just sensation.

Design Reflection

Can we design AI outputs that invite reflection instead of shortcut it?

One option: a “cautious mode” that visually distinguishes high-confidence from low-confidence output. Another: epistemic scaffolding that exposes how the answer was constructed — sources, model weights, probability layers. Not to overwhelm. To slow the read. To introduce hesitation. To restore judgment. This is a shift in design ethos from fluency-as-satisfaction to interaction-as-inquiry. Because if polish remains the metric for trust, even experts will mistake performance for truth. That’s not just a design flaw. It’s a cultural failure.

In this phase of the AI journey, where the user interprets output, we must design pause. That is where sense-making happens. And where shared intelligence human and machine begins to take root.

Friction is Not Always Bad

We often hear that good design should “get out of the way.” In AI, this principle has become doctrine: optimize for flow, reduce clicks, surface shortcuts. But what if that is the wrong goal for some phases of the user journey? As an HCI researcher, I’ve observed that seamlessness often undercuts learning, cognition, and agency — especially in generative AI tools. Automation excels at outputting, but weakens engagement. When systems hand over answers before users even form their own questions, we risk designing away the very processes that make human intelligence meaningful.

Friction — slowness, ambiguity, deliberate gaps — is often dismissed as poor usability. But in the right context, it’s a design affordance. Borrowing from William Gaver’s work on ambiguity, friction invites users to interpret, reflect, and reengage. I see this in my own research: users skip details when AI summaries appear too “complete.” They trust the output, but forget the input. The “answer” becomes the interface. This bypass is efficient, but hollow. We are not just designing flows; we are shaping cognitive behaviors. And sometimes, the most ethical interface is one that interrupts.

Designers often overfit to usability heuristics like “minimize user memory load” or “accelerate system response time.” These are still valuable. But in AI, they must coexist with new priorities: preserving interpretability, supporting slow thinking, and protecting effort as a form of learning. When users scan instead of engage, trust instead of verify, click instead of interpret — the interaction looks “clean,” but something has been lost. Friction, when used intentionally, restores depth.

Supporting Research

Information-seeking behavior theory has long emphasized that users pursue the “principle of least effort” — not because they are lazy, but because they optimize based on available time, skill, and trust (Tubachi, 2017, p. 6). This insight, drawn from library science, maps uncannily onto AI tools: when answers are fast and fluent, users disengage. Theories such as Brenda Dervin’s Sense-Making Methodology suggest that meaning-making requires an encounter with uncertainty — what Dervin frames as “gaps” that trigger interpretation (Tubachi, 2017, p. 5).

In parallel, Gaver et al. (2003) argue that ambiguity is not a usability flaw but a resource — especially in contexts where interpretation, exploration, or reflection matter. They identify three types of ambiguity (of information, context, and relationship), and demonstrate how each can support richer engagements between user and system (Gaver et al., 2003, pp. 3–5). Within AI, these strategies could delay decision finality, preserve interpretive space, and activate deeper user sense-making — a design move, not a failure state.

Furthermore, as Ooi and Liew (2011) observe in their study on fiction selection, users do not always follow explicit cues. They interpret based on affect, narrative anticipation, and contextual trust. The lesson here is that design must not only deliver — it must allow for wandering, hesitation, and choice.

References:

• Gaver, W.W., Beaver, J. & Benford, S. (2003). Ambiguity as a resource for design. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pp. 233–240.

• Tubachi, P.S. (2017). Information Seeking Behavior: An Overview. Rayalseema University, pp. 3–7.

• Ooi, K., & Liew, C.L. (2011). Selecting fiction as part of everyday life information seeking. Journal of Documentation, 67(5), pp. 748–772.

Design Scenario

In UX research sessions I’ve led, I’ve seen this in real time. Teams using AI to summarise meeting notes often skip playback, ignore dissenting inputs, and fast-forward to “takeaways.” The AI output is neatly bulleted, but what is left out is the tension, uncertainty, and contradiction — the very content that makes team decisions durable. One session involved a product manager who admitted, “I skimmed the summary and made a decision. Later I found out I misread the team's sentiment.” The AI didn’t mislead him. The speed did.

This is not limited to enterprise use. AI writing tools for students, caregivers, even fiction authors increasingly position summaries, rephrases, and previews as final outputs. This shift from “draft” to “done” is not about quality — it's about how friction (or its absence) reshapes our perception of readiness.

Design Element(s)

Friction-by-design includes micro-interactions that slow down flow just enough to provoke reflection. These may be:

• Preview toggles that compare original vs. AI-suggested content

• “Pause to consider” prompts before accepting outputs

• Contradiction alerts that nudge users to double-check

• Gradient-based confidence sliders that avoid binary correctness

Designing for productive friction also means resisting some default UX heuristics. For example, Nielsen’s guideline to “minimize cognitive load” (Nielsen, 1994) is not universally applicable in AI contexts. Sometimes, increasing cognitive engagement — through thoughtful pacing or ambiguity — enhances usability by preserving user understanding and accountability.

We might also draw from sensory design — tone, spacing, silence — to inject intentional gaps. Ambiguity here is not about confusion; it’s about choice, pacing, and re-entry.

Design Reflection

From my practice, I’ve seen how friction clarifies intent. It gives users time to rethink, retrace, and reframe. And yet, it’s not part of most design KPIs. We reward speed, not hesitation. But in AI, where outputs feel “done,” that’s dangerous. I believe we need new metrics: friction yield (how often hesitation improves outcomes), or interpretability quotient (how much sense-making is retained). We also need to teach teams — not just design teams — that slowness is not inefficiency, but intentionality.

What would it look like to design an AI product where nothing could be accepted without a moment of pause? Could friction be our ethical stance — a guardrail, not a glitch?

Designing for Unfinishedness

In AI interfaces, polish often masquerades as authority. The output appears complete: clean structure, resolved tone, full stop. But intelligence, particularly human intelligence, is rarely that crisp. When we design systems that present conclusions as endpoints, we close off inquiry. We stifle critique, remix, and doubt. Designing for unfinishedness means deliberately resisting the impulse to resolve too soon. It means building affordances that keep space open — for revision, contradiction, and dissent.

This waypoint emerges when users move beyond interaction into interpretation. I’ve seen it in product teams who receive AI-generated insights but can’t trace how they were formed — or in users who want to challenge an output but have no structured way to do so. The system feels final. And that finality becomes epistemic closure (the premature acceptance of a conclusion as final, blocking further inquiry). From an ethical standpoint, this is dangerous. It turns AI into a declarative machine, not a dialogic partner. In my view, unfinishedness is not a flaw. It is a stance — one that supports accountability, fosters critique, and preserves space for thought. And as designers, we must stop equating clean output with cognitive closure.

Supporting Research

Lucy Suchman’s work on situated action reframes how we understand human-machine interaction: behavior is not executed from pre-set plans, but arises through contingent, real-time adjustment. From this view, systems that present finalized outputs erase the conditions of their creation — denying users a way to situate or interrogate what they’re seeing (Suchman, 2007, as cited in emergenceaam.pdf, p. 3). Anne-Laure Fayard’s contributions further this by emphasizing ambiguity as a productive design strategy. As echoed in Gaver et al. (2003), ambiguity in information, context, and relationship allows users to interpret, adapt, and act on their own terms — sustaining engagement rather than scripting it.

This epistemic openness contrasts sharply with the dominant mode of generative AI, where output is often presented as monologic, definitive, and closed. Escobar’s (2018) concept of ontological design reminds us that systems instantiate worldviews. When AI presents a polished conclusion, it collapses possibility into singularity — suppressing plurality, dissent, or reinterpretation. Designing for unfinishedness reintroduces the plural.

Gaver, W., Beaver, J. and Benford, S., 2003. Ambiguity as a Resource for Design. In: Proceedings of CHI '03. ACM, pp. 233–240.
Suchman, L.A., 2007. Human–Machine Reconfigurations: Plans and Situated Actions, 2nd ed. Cambridge University Press.
Escobar, A., 2018. Designs for the Pluriverse: Radical Interdependence, Autonomy, and the Making of Worlds. Duke University Press.

Design Scenario

In collaborative policy workshops, I’ve used AI tools to generate early-stage drafts of proposals based on stakeholder input. While helpful, the AI-generated draft often shifted the tone of the session. The group moved from exploration to editing — from generative dialogue to reactive polishing. It was as if the AI had taken the air out of the room. People responded to the shape, not the substance. Suggestions became tweaks, not challenges. The form of completion redefined the space for contribution.

This wasn’t just about speed. It was about epistemic authority. The draft carried the weight of “done,” even when the content was shallow. I realized we needed to build the draft differently — not as output, but as trace. A provocation. A partial. This changed everything. People started asking “What’s missing?” rather than “Is this good enough?”

Design Element(s)

Designing for unfinishedness can take many forms. Draft modes that persist visually — “This is still evolving.” Trace overlays that reveal how the content came to be — edits, sources, contradictions. Layered outputs that offer variants rather than verdicts. Even versioning models that show paths not taken — as a way to invite reflection on alternatives.

Importantly, unfinishedness also means resisting some familiar design tropes: single-answer outputs, definitive tone, monologic text blocks. Instead, systems can ask users: “What would you add?” or “Do you agree with this framing?” Interfaces might offer uncertainty ratings, comment scaffolds, or co-writing slots. These elements transform AI from answer-giver to thinking partner — not by reducing its capabilities, but by expanding its context.

Design Reflection

Unfinishedness is often viewed as a gap to be closed. But in my practice, it’s been a space to be opened. The temptation in AI design is to polish everything — clean edges, complete summaries, decisive tone. But that polish too often becomes power. It shuts down interpretation. As a designer, I believe in a different ethic: epistemic humility. We can build systems that say, “This is just one version,” or “You decide where this ends.”

What if interfaces defaulted to suggest mode? What if users saw not just answers, but unanswered questions? Unfinishedness is not about brokenness. It’s about inviting users back into the loop — as co-authors, not consumers. That, to me, is a more honest form of intelligence. And a far more human one.

Designing for Cognitive Reentry

AI systems like ChatGPT are designed to deliver output: fast, fluent, and polished. But that very fluency can backfire. When responses are too smooth, too “done,” users may disengage. The system has spoken. There is no space to think. What’s missing here is cognitive reentry the ability for the user to shift from passive receipt of output to active processing, questioning, and reinterpretation.

In cognitive psychology, Daniel Kahneman’s dual-process model distinguishes between System 1: fast, automatic, intuitive thinking and System 2, which is slow, deliberate, and analytical. Although this model was not developed in relation to AI, it provides a useful lens: ChatGPT often reinforces System 1 behaviours, encouraging users to accept rather than interrogate. From a design perspective, this is more than a usability flaw. It’s a conceptual failure. Interfaces built around linear input–output loops don’t account for shifts in cognitive mode. They deliver text but fail to scaffold thought.

In my own practice, I often find myself copying AI responses out of the chat interface and into separate spaces like documents, notes, whiteboards. Not because the response is wrong, but because the interface offers no affordance for thinking. No way to annotate, hesitate, counterargue, or slow down. This design oversight breaks the loop between knowledge and reflection and forces the user to manually reconstruct a space for cognitive engagement.

Supporting Research

Morewedge and Kahneman (2010) show how associative coherence and processing fluency can lead people to over trust intuitive responses. When something feels right — coherent, elegant, familiar — we often stop interrogating it. While their study was not conducted in the context of AI or ChatGPT, their findings illuminate a key risk in current generative UX: AI outputs often feel correct due to fluency, even when they’re incomplete or flawed. The issue is not just what the AI says, but how its output feels — smooth, fast, resolved. This emotional ease of consumption suppresses cognitive reentry. Without design interventions, users are nudged toward acceptance before reflection.

Kannengiesser and Gero (2018) apply this dual-process framework to design workflows. Their empirical study shows that experienced designers toggle between intuitive (System 1) and deliberate (System 2) modes but tools must support that shift. Most AI tools don’t. They encourage fast prompting, not reflective iteration. In my own use, I often copy outputs into other interfaces just to reengage with them. That behaviour isn’t a workaround, it’s a missing design affordance.

Morewedge, C.K. and Kahneman, D., 2010. Associative processes in intuitive judgment. Trends in Cognitive Sciences, 14(10), pp.435–440.
Kannengiesser, U. and Gero, J.S., 2018. Design thinking, fast and slow: A framework for dual-process theory. Design Science, 4(e17), pp.1–36.

Design Scenario

In usability research, I’ve seen knowledge workers: designers, academics, policy writers, copy AI output into Word documents, Notion spaces, even their email drafts. Not to publish. But to re-see/re-write the content in a different space, one that allows scrolling, margin notes, versioning. Simply to have fresh eyes! When asked why, a fellow designer told me, “The interface is the same always. I need to move it somewhere so it is different, so i can edit it well.”

In one project, we tested a design research assistant powered by LLMs. Teams initially loved the summaries, but later admitted they often misremembered them as correct even when factually flawed. The response looked too good. Their own reflection had been short-circuited. The solution wasn’t better summaries. It was better space around them.

Design Element(s)

To support cognitive reentry, we must interrupt the consumption loop — the automatic progression from prompt to passive acceptance. This doesn’t mean making systems harder to use. It means inviting the user to think or see again. Experience design elements that support this include:

• “Pause & Reflect” buttons that delay output or ask users to anticipate or summarize expected outcomes before revealing the answer

• “Reverse Prompting,” encouraging users to restate or reframe the AI’s response in their own words

• “Agree/Disagree sliders” or emotional alignment toggles that help users locate themselves in relation to the content

• “Critical View” modes where AI outputs are accompanied by potential blind spots, assumptions, or counterpoints

• “Copy for Reentry” affordance, which explicitly invites the user to shift the content into a manipulable, editable layer

• “Mode Switch” toggles, allowing users to shift between a generative interface and an evaluative or editing interface with a different visual and structural rhythm

• “What’s Missing?” micro-prompts, subtly nudging users to question the completeness of the answer

• “Challenge this Output” feedback pathways, where disagreement or doubt becomes part of the interface, not an external reaction

These are not UI tricks. They are experience-level affordances that support interpretive judgment, not just interaction. They ask not “Is the AI right?” but “Do I recognize my thinking in this output?”

Design Reflection

Are we designing for users to be impressed, or to be engaged?

When interfaces present outputs as endpoints, they rob users of the moment where meaning is made — the pause, the doubt, the rewrite. In my practice, I’ve found that where AI is strongest is also where users must be strongest: in interpretation. We need tools that don’t just produce content, but provoke thinking.

Cognitive reentry is a form of care — not only for correctness, but for ownership. Users need space to question, restate, hesitate. But more than that: evaluation is not a separate task — it’s integral to experience design. If critique is only external, it arrives too late. If reflection is left to the user’s own habits, it becomes optional. But when reflection is invited by the interface, it becomes part of the system’s thinking loop.

In the age of generative systems, it’s not enough to answer well. We must design for the moment after the answer — the point where intelligence becomes mutual.

Designing for Scenius

Most AI creativity tools are built around the myth of the lone genius — the brilliant user, prompting in isolation, iterating privately, and copy-pasting their way to insight. But creativity rarely works this way. Great ideas emerge in context: from friction, feedback, contradiction, and remix. In my practice, I’ve seen how current tools over-index on private generation and under-index on shared construction. Prompting becomes performance. Outputs become secrets. The result isn’t just plagiarism or academic dishonesty — it’s cultural thinness.

Brian Eno’s term “scenius” (collaborative rather than individual creativity — “an ecology of talent”) offers a powerful counter-frame. Creativity isn’t born in isolation; it thrives in ecosystems of exchange. Under scenius, innovation is a network effect — shaped by shared drafts, derivative sparks, small contributions that converge over time. But today’s AI UX enforces solo loops. The interface assumes individual authorship, linear progress, and private output. We get smart tools, but silent processes.

Designing for scenius means rethinking how we frame, trace, and share AI-generated work. It asks: what if creativity wasn’t a secret act, but a collaborative practice? What if prompting happened in teams? What if remixing left visible trails? The goal isn’t just better tools — it’s a more just and generative culture of making.

Supporting Research

Brian Eno coined the concept of scenius to reframe creativity as an ecological event — a product not of solitary brilliance, but of “an ecology of talent.” Austin Kleon expands this idea in Show Your Work! (2014), where he describes scenius as the creative condition formed by interconnected contributors — not geniuses, but generous collaborators who amplify and evolve each other’s ideas (p. 10)Show_your_work.

Glăveanu (2014) affirms this perspective from a psychological standpoint: creativity, he writes, is not located inside the individual, but emerges through dialogical, relational action. “No creative act is complete without appreciation,” he argues — without someone else responding, reinterpreting, or building on what was madeThe Standard Definition….

Meanwhile, Runco & Jaeger (2012) offer a functional definition: creativity must be both original and effective. If LLMs produce novelty without groundedness, then scenius supplies the interpretive infrastructure — helping ensure that AI-generated outputs are not only imaginative, but also relevant and resonantThe Standard Definition….

Marcus & Davis (2019) offer a crucial AI-centric critique: LLMs “recombine without reason.” They lack situational awareness, cultural memory, and audience sensitivity. Scenius, by contrast, restores those dimensions — not just through human presence, but through designed processes of co-creationThe Standard Definition….

Design Scenario

In academic settings, I’ve seen students use ChatGPT to write quietly — and privately. There’s no way to know if they’re collaborating, copying, experimenting, or just submitting AI output. It’s not just a plagiarism issue. It’s a design issue. The interface hides process. It collapses creativity into keystrokes. One student said, “I don’t feel like I made it, but I don’t feel like I cheated either.” That ambiguity isn’t ethical failure. It’s a UI failure.

In contrast, in design workshops where teams worked openly with AI tools — versioning their prompts, discussing options, remixing each other’s outputs — something else happened. Prompts became provocations. Responses became fragments. Creativity unfolded across the room, not just on a screen. The tool was no longer a solo oracle. It was a node in a collective conversation.

Design Element(s)

To support scenius, we need interfaces that make remix traceable, iteration visible, and co-creation normal. Possible design elements include:

• Prompt history maps, showing how ideas branch, merge, and evolve

• Shared prompting modes, where users can build prompts together or remix each other’s drafts

• Attribution trails, showing how different contributions shaped the final output

• Collaboration layers, where AI suggestions sit alongside human annotations, not on top of them

These structures move us beyond solo loops. They treat AI as a collaborative medium, not a magic mirror. From a design ethics perspective, scenius isn’t just more inclusive. It’s more honest. It reflects how knowledge actually forms — iteratively, relationally, socially.

Design Reflection

Designing for scenius challenges our deepest interface assumptions. We’ve inherited models from search, command lines, and word processors — all solitary. But the future of creativity is collective. As a designer, I want to build systems that reward contribution, not just polish. That show the paths not taken, not just the ones we copied. That recognize creativity as care, not just cleverness.

And this isn’t just idealistic. If we take human-centered design seriously, then we must design processes where humans aren’t just consuming AI output — they’re embedded in how it forms. Scenius is how we restore trust: not by perfecting the machine, but by involving people meaningfully in what it produces. When users co-author, remix, or trace a shared output, they feel ownership. They recognize themselves in the work. That’s not just more creative — it’s more ethical, more sustainable, and more human.

Situated Intelligence: Designing for Epistemic Plurality

AI systems are often praised for their universality, the ability to serve anyone, anywhere. But behind that universality lies a hidden monoculture: language norms, value assumptions, interaction metaphors, and ontologies rooted in Euro-American paradigms. The systems may translate your words, but not your world. They respond in many languages, but think in only one. >> In AI design, this isn’t just a cultural blind spot, it’s an epistemic one.

Situated intelligence (context-specific knowledge shaped by cultural, spatial, and social factors) demands that we ask: Whose knowledge is legible? Whose worldview gets encoded in the interface logic? When ChatGPT answers in fluent English using corporate tone and Western logics of argument, it isn’t being neutral, it’s being positioned. Designing for epistemic plurality means moving beyond localisation to question the foundations of the AI interface: not just how it responds, but how it sees, values, and structures interaction. This is not about identity politics. It’s about cognition, justice, and knowledge legitimacy.

The risk is not only misalignment. It’s erasure. When vernacular modes of speaking, knowing, or storytelling are flattened into tokenised inputs when a prayer becomes a paragraph, when a ritual becomes a glitch design becomes a colonial act. What’s needed is not a better translation layer. What’s needed is a situated, reflexive, pluralistic design ontology.

Supporting Research

Ahmed Ansari (2015) argues that decolonizing design is not about adding perspectives to an existing structure, but restructuring the very ontology of how design recognises knowledge. Escobar (2020) echoes this, advocating for a pluriversal design logic that resists “one-world world” thinking the idea that design serves a single reality. Instead, he proposes designing for many realities at once.

Tristan Schultz introduces the idea of onto-cartographies mappings that reveal how design travels along invisible cultural, historical, and colonial routes. He calls for design that works with fugitive ontologies, community-based knowledge systems that refuse categorisation but persist through relationality. These perspectives reframe UX not as surface polish, but as epistemic choreography: shaping how knowledge is structured, sensed, and interpreted.

The “On Decolonizing Design” working group (2021) warns that even methods considered neutral, like usability testing or journey mapping, often reproduce dominant ways of knowing. And Albrecht’s Postcolonialism Cross-Examined (2023) challenges the Western/non-Western binary altogether, proposing multidirectional postcoloniality to account for epistemic movements from post-Soviet, post-Ottoman, and non-aligned contexts.

Together, these voices offer a design imperative:

design not just for users, but for worlds.

Ansari, A. (2015). Decolonising design: Towards a more reflexive design studies.
Escobar, A. (2020). Designs for the pluriverse.
Schultz, T. (2021). A design anthropology of the South. RFDC.
Albrecht, M. (ed.) (2023). Postcolonialism Cross-Examined.
Ahmed, S. et al. (2021). On Decolonizing Design.

Design Scenario

In one localisation project for a chatbot intended to support first-generation university students, the design team discovered that translating messages into Arabic still resulted in low engagement. The tone was “off.” The advice structure was linear. The use of bullet points and problem solution logic felt foreign. What students wanted was not guidance, but narratives: stories, examples, questions that opened up meaning.

In another case, an HR tool using LLMs to assess resumes consistently flagged African, Southeast Asian, and Gulf-style CVs as “unstructured” or “non-standard,” penalising formats that emphasize community affiliations or spiritual qualifications. The issue wasn’t bias in the model alone, it was the interface logic of what counted as structured input. The design excluded other ways of signaling intelligence, credibility, and professionalism.

Design Element(s)

Designing for epistemic plurality requires experience design affordances that make room for contextual knowledge, alternative logics, and cultural rhythm. These may include:

• Ontologically diverse onboarding: not just choosing a language, but a worldview (ritual, linear, oral, communal)

• Cultural code-switching interfaces, where tone, interaction rhythm, and values shift across user types

• Vernacular metaphors in interface design, replacing file/folder logic with locally meaningful symbols

• Embodied indicators of time, mood, or deference, not always logical, but contextually intuitive

• Dialogue structures that resist single shot QA, enabling stories, digressions, ambiguity

• Community-trained co-models, co-created with underrepresented cultural communities, grounded in their ways of asking, not just answering

These are not customisations. They are epistemic architectures and their absence reinforces cultural monocultures.

Design Reflection

Can we still call it “human-centered design” if it only centers some humans?

Too often, design assumes proximity to power means proximity to universality. But users are not just diverse in surface identity, they are diverse in cognition, metaphor, memory, and speech. Interfaces that flatten these into polite correctness or clickable diversity perform inclusion but enact exclusion.

To design for cultural intelligence is to resist universality. It is to design with the awareness that AI systems don’t simply operate in a world they structure the kinds of worlds users can see, name, and claim. In a pluriversal design future, interfaces would be answerable to their cultural blind spots. They wouldn’t just speak your language. They would learn to listen in your world.

Methodologies Under Review

AI design challenges are rarely methodological … until they are. For decades, UX and HCI research methods have evolved around deterministic systems: fixed-function software, screen-based flows, mobile interactions with clear state transitions. But generative AI doesn’t behave. It diverges, adapts, hallucinates, contradicts. It behaves differently across time, users, and contexts. What we are studying isn’t a static system. It’s a semi-autonomous system in motion.

I’ve led research where some participants prompted once and moved on. Others revised five times. Some turned off chat history or anonymised data to protect privacy. Others freely shared. Attempts to “equalise” this variation for the sake of methodological control risk flattening the truth. In these environments, variation is the data not noise but signal.

This requires a reorientation. We must shift from isolating variables to mapping behaviour across chains: divergence, recomposition, reentry.

To do this, I’ve developed preliminary metrics:

• Prompt iteration depth: how many times users revise or reframe their ask.

• Context carryover: how much previous interaction informs current behaviour.

• Semantic noise thresholds: where AI responses introduce drift, contradiction, or dilution.

These are not precision tools. But they get me closer to the real design questions:

When do users trust?

When do they try again?

When do they stop?

Traditional usability metrics cannot answer these questions, because they weren’t built for systems that behave probabilistically. In consultancy, I’ve seen clients demand quantifiable evidence of AI value not usability. “Give me a number to prove the ROI.” But what number proves trust? Or sense-making? Traditional KPIs like speed or retention are proxies at best. In generative AI, we need epistemic ROI (return on investment measured by gains in thinking, understanding, and decision quality, not just efficiency): What did the system enable — or displace — in cognition, creativity, and decision quality? Until we measure thinking, not just output, we are not measuring the experience.

Design Scenario

In a field study with a generative AI assistant for design researchers, users loved the summaries at first. They skimmed, nodded, accepted. But when debriefed, they admitted they had misremembered hallucinations as facts. It wasn’t a model error, it was a method error. I had no instrument to detect how trust drift (the subtle shift from healthy skepticism to uncritical acceptance over time) or over-acceptance bias. And that’s a methodological blind spot. Some might argue this is the job of AI engineers to test the model. But this assumes experience is separable from cognition. It’s not. Engineers validate the model. Designers validate meaning. A system can be technically accurate and still epistemically misleading if it fails to scaffold user judgment that’s over-acceptance bias (when users mistake fluency or confidence for correctness).That makes this a UX issue not just a technical one.

In another case, I facilitated participatory foresight sessions using LLMs to simulate future scenarios. What mattered wasn’t the prompt or the output it was when users stopped. Stopped revising. Stopped questioning. I wasn’t witnessing satisfaction. I was witnessing cognitive disengagement. Not because users were tired but because the output felt “done.” The interface didn’t invite continuation — it performed finality. There was no room left to contribute, no lingering question to pursue. The system/design didn’t provoke continuation, it performed finality.

These were epistemic moments.

But traditional methods task timing, error counts, satisfaction scores — completely missed them.

Design Element(s)

We need research methods that mirror the complexity of what we’re building. That means experience instrumentation, not just usability heuristics. Some tactics I’ve been exploring or will be exploring:

• Live model probing: pairing researchers with users to map moments of alignment, drift, or surrender.

• Epistemic diary studies: where users reflect not just on output, but on why they accepted or ignored it.

• Interaction belief maps: tracking not paths to completion, but paths to meaning.

• Trust deltas: plotting how trust forms or fractures across prompts, not across screens.

In generative AI, we need epistemic ROI: What did the system enable — or displace — in cognition, creativity, and decision quality? For instance, if a policymaker adopts an AI-generated summary without questioning it, we might “save” time, but at the cost of eroded deliberation. That’s a negative epistemic ROI, even if the interaction looked efficient.

On Epistemic ROI
In AI implementation, clients and stakeholders often ask the same question: What’s the ROI? But in generative contexts, the real return isn’t just speed or clicks. It’s cognitive value. Did the system help users think better? Understand more deeply? Make wiser decisions? Traditional KPIs capture activity — not epistemic integrity. We need new ROI frames:

• Trust delta: Did trust increase or calcify prematurely?

• Interpretive load: Did the interface reduce or reroute reflection?

• Retention gain: Did users actually remember what they engaged with?

These aren’t soft metrics. They are signals of whether the AI is augmenting cognition or outsourcing it. And if design can’t account for them, it’s not just missing the point — it’s misrepresenting the product’s value.

This cannot just stay within HCI. As AI merges with ambient systems like sensors, urban interfaces, spatial computation, we’ll need methods that scale to rooms, buildings, and cities. That’s why I’ve turned to space syntax in my own work. Not to leave UX behind, but to expand the vocabulary.

My current research examines how spatial configuration, embedded digital stimuli, and bodily movement shape affordances for co-presence, memory, and social cognition. Public plazas — not just chat windows — will become AI-mediated zones of ambient interaction. We are critically underprepared to study those contexts with current UX tools.

Equally, we must bring non-human and environmental stakeholders into methodological scope. Models like Environment-Centered Design (Sznel, 2020) and Transition Design (Irwin et al., 2015) call for slow knowledge, intergenerational foresight, and multi-scalar evaluation. They don’t just expand the stakeholder list, they reframe research itself as a planetary commitment. This isn’t a separate track. It’s the next step for design ethics in the age of AI.

Design Reflection

I sometimes wonder what design research will look like in ten years.

Will we still be using click-through rates to evaluate systems that shape cognition, trust, and identity? And even if we build better cognitive instruments, are we ready for what comes next?

AI won’t stay in chat windows. It will blend into streetlights, conference rooms, domestic appliances. It will co-construct not just answers, but atmospheres. We need methodologies that stretch across tempo, topology, and scale. That’s why I study space syntax. Why I’ve turned to Transition Design. Not because I want to theorise forever, but because I want to test meaning where meaning actually forms.

If generative systems shape how people move through buildings or interact with public screens, do we study user satisfaction — or spatial cognition drift? Do we test feature adoption or environmental meaning reconstruction?

To study AI well, we must design research that sees behaviour as situated, cognition as relational, and experience as epistemic choreography that’s why I am a design researcher.

Design After the Answer

These are not UX tweaks. They are epistemic decisions. Every prompt box, fluent response, or skipped pause shapes how we come to know, not just what we do. And many of the most consequential design moments in AI systems live in places no one owns the gaps between comprehension and action, between output and meaning.

These moments are not neglected because they’re small. They’re neglected because they fall between disciplines. Between cognition and computation. Between system performance and user belief. Design doesn’t step in because no one else did. It steps in because it’s uniquely fluent in ambiguity, affordance, and context.

That’s why this isn’t a call for best practices. It’s a call for new ones. To measure trust, not just clicks. To protect interpretation, not just speed. To design for what happens after the answer.

Afterword: Why I Wrote This, and What Comes Next

This piece began not as an argument, but as an act of navigation. As an HCI researcher working across civic systems, corporate AI teams, and urban spatial contexts, I kept encountering something I couldn’t name: the quiet dissonance between what AI interfaces appear to offer and what they actually shape. These waypoints are my attempt to surface that gap—not to close it.

If the tone feels unresolved, that’s intentional. These are not guidelines, frameworks, or roadmaps. They are design seams: places where meaning slips, where cognition bends, where interface becomes ideology. I chose not to resolve them, because resolution is often how systems hide their assumptions.

You’ll notice I don’t offer clean stakeholder matrices or implementation steps. That’s not a gap—it’s the point. In my experience, design accountability in AI is so deeply fragmented that no single team can “own” what’s missing. Responsibility diffuses across managers, metrics, timelines. And so, this piece is not operational. It’s observational. It asks what we’re building before asking how to fix it.

Some readers may also note what’s absent: power, platform economics, systemic incentives. These will come—later, deliberately. The scope here is epistemic: to trace how design reshapes cognition, not yet how corporate structures reinforce that shaping. But they are inseparable, and I plan to explore them in upcoming pieces.

In fact, several threads from this essay now demand their own space:

• Designing for Scenius will unpack collaborative AI creativity beyond solo authorship.

• Pluriversal UX will explore what it means to design for culturally situated intelligence.

• What UX Can’t Yet Measure will ask whether current research tools are enough to study epistemic alignment, trust drift, or cognitive reentry.

If this essay marked the moment “after the answer,” these next ones begin where participation, plurality, and sense-making might reenter the loop. Because that’s where design belongs—not at the end, but in the middle, mediating.

If you're still with me here, thank you. That means something is resonating—and that, perhaps, is the most honest form of design validation we have right now.

Reference

• Ahmed, S. et al., 2021. On Decolonizing Design. Working Group Report. Available at: https://decolonisingdesign.com [Accessed 3 July 2025].

• Albrecht, M. (ed.), 2023. Postcolonialism Cross-Examined. London: Routledge.

• Balakrishnan, J., Sundaravel, E. and Watanabe, K., 2021. Role of cognitive absorption in building user trust and experience in human-AI interaction. Psychology & Marketing, 38(12), pp.2312–2327.

• Escobar, A., 2018. Designs for the Pluriverse: Radical Interdependence, Autonomy, and the Making of Worlds. Durham: Duke University Press.

• Fayard, A.-L., 2012. Designing with ambiguity to support human–machine dialogue. In: Proceedings of CHI 2012. New York: ACM, pp. 233–240. (Note: Adjusted citation combining Fayard & Gaver for clarity.)

• Gaver, W., Beaver, J. and Benford, S., 2003. Ambiguity as a resource for design. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pp. 233–240.

• Glăveanu, V.P., 2014. The Psychology of Creativity: A Critical Reading. Creativity: Theories – Research – Applications, 1(1), pp.10–23.

• Kannengiesser, U. and Gero, J.S., 2018. Design thinking, fast and slow: A framework for dual-process theory. Design Science, 4(e17), pp.1–36.

• Kleon, A., 2014. Show Your Work! 10 Ways to Share Your Creativity and Get Discovered. New York: Workman Publishing.

• Lee, E. and Palmer, J., 2025. Prompt engineering in higher education: A systematic review. International Journal of Educational Technology in Higher Education, 22(4), pp.5–15.

• Marcus, G. and Davis, E., 2019. Rebooting AI: Building Artificial Intelligence We Can Trust. New York: Pantheon Books.

• Morewedge, C.K. and Kahneman, D., 2010. Associative processes in intuitive judgment. Trends in Cognitive Sciences, 14(10), pp.435–440.

• Nielsen, J., 2023. The articulation barrier: Prompt-driven AI UX hurts usability. UX Tigers. Available at: https://jakobnielsenphd.substack.com [Accessed 3 July 2025].

• Ooi, K. and Liew, C.L., 2011. Selecting fiction as part of everyday life information seeking. Journal of Documentation, 67(5), pp.748–772.

• Park, S., Park, E. and Kwon, S., 2023. Prompting, planning, reflecting: Barriers to deep cognition in LLM-powered chat interfaces. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI '23), pp.1–12.

• Rzepka, R. and Berger, T., 2018. A systematic review of human–AI conversational systems: Trust, fatigue, and interface misalignment. AI & Society, 33(3), pp.409–421.

• Schuetz, S. and Venkatesh, V., 2020. Research perspectives: Conversational AI and epistemic agency. Information Systems Research, 31(4), pp.1095–1107.

• Schultz, T., 2021. A Design Anthropology of the South. RFDC.

• Shojaee, P., Mirzadeh, I., Alizadeh, K., Horton, M., Bengio, S. and Farajtabar, M., 2025. The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity. Apple Machine Learning Research. Available at: https://ml-site.cdn-apple.com/papers/the-illusion-of-thinking.pdf [Accessed 3 July 2025].

• Suchman, L.A., 2007. Human–Machine Reconfigurations: Plans and Situated Actions, 2nd ed. Cambridge: Cambridge University Press.

• Tubachi, P.S., 2017. Information Seeking Behavior: An Overview. Rayalaseema University.

• Varma, S., Dastin, J., Choi, J. and Gershman, S., 2024. Delegation effects in human-AI collaboration: A neurocognitive perspective. MIT Cognitive Systems Lab. [Unpublished Manuscript].

• Walter, A., 2024. Thinking through prompting: Critical frameworks for AI literacy. Design Literacy Review, 3(1), pp.17–25.

• Wang, Y. and Li, B., 2022. Multi-turn coherence and memory load in human-AI conversation: A cognitive perspective. Cognitive Systems Research, 70, pp.1–12.