Design thinking practitioners and UX researchers are working with the wrong map. They are modeling people as fixed archetypes — and building AI that pays the price. Here is what they are missing, and why it matters more than ever.
"A mental model is not a portrait of a person. It is a map of how a mind moves — and you cannot build AI that serves humanity if you do not understand how human minds actually move."
Joanna Peña-Bickley, Mother of CognitiveExperience.Design
In 2014, while designing AI-powered software at IBM Watson, I coined the phrase Cognitive Experience Design — not because I was searching for a new title, but because the field of design genuinely lacked a name for what needed to happen next. Designers were being handed machine learning systems and handed back the same old tools: user journeys, customer personas, demographic segments. The gap between those instruments and the challenge in front of us was vast. That gap has only grown. And in 2026, as artificial intelligence approaches something we are beginning to call general reasoning, I want to name the problem plainly: the design community has been operating with the wrong mental model of mental models. And that mistake is encoding harm into the systems that will shape the next century of human cognition.
What Cognitive Experience Design Actually Is
Let me be direct about what distinguishes Cognitive Experience Design (CXD) from the design thinking frameworks that have dominated the industry for two decades. Design thinking asks: Is this usable? Is it delightful? Does the journey flow? These are not bad questions. But they are surface questions. CXD asks something structurally different: Does this support how people actually think — and does it leave them more or less capable of thinking for themselves?[1]
CXD draws on cognitive load theory, perception science, memory research, and the philosophy of the extended mind — the proposition that when AI becomes embedded in daily decision-making, it is not a separate tool. It is a participant in the cognitive process.[6] When a recommendation engine decides what you read, when a language model shapes your argument, when a predictive system pre-selects your choices, the line between your thinking and the machine's thinking blurs. That blur is a design problem of the first order.
The role of AI should NOT be to find the needle in the haystack. It should be to show users how much hay it can clear — so they can better see the needle themselves. Agency is not a feature. It is a fundamental design requirement.
What I established through years of practice at IBM Watson, Amazon Alexa, and Uber — and what the research now confirms — is that the practical principles of CXD are: cognitive clarity, agency preservation, effort calibration, and directed attention management.[2][3] These are not philosophical abstractions. They are operational design constraints that determine whether a system builds human capability or quietly erodes it.
EEG research now shows that intensive interaction with large language models reduces frontal theta power — a well-established marker of working memory load.[7] We are producing what researchers are calling the Hollowed Mind: users who bypass the deliberative effort required to build resilient internal reasoning architectures. Convenience is the anesthetic. Cognitive atrophy is the outcome. CXD is the corrective discipline.
What a Mental Model Actually Is — And Why the Field Has It Wrong
Here is the problem I see consistently, at every level of the design profession. When practitioners say "mental model," they mean one of two things: either the broad colloquial sense of "what users expect," or — worse — they treat it as a synonym for persona. Both usages are wrong. And the confusion is not semantic. It produces structurally deficient research, and it produces AI systems that are aligned to a fictional average rather than the actual range of human cognition.
A mental model, properly defined, is an individual's internal representation of how a system, process, or situation works. It is not a factual map. It is a belief-based simulation — the cognitive script a person runs when they encounter a situation, determining what actions they take, what errors they make, how they interpret outcomes. As Nielsen Norman Group defines it: mental models are "based on belief, not facts."[9] They are derived from observation, perception, experience, and culture.
Donald Norman's framework identifies four interdependent components: the target system, the designer's conceptual model, the user's actual mental model, and the user's predicted model of future behavior.[10] The design failure space lives entirely in the gap between the conceptual model and the mental model — between C(t) and M(t). That gap is the central diagnostic space of CXD. And most practitioners are not working there at all.
"The most consequential design failures occur in the gap between what designers intended users to understand — and what users actually believe."
What makes this especially urgent now is a property of mental models that the design community almost universally ignores: they are not fixed. Indi Young, who developed the most operationally rigorous methodology for building mental model diagrams, makes this explicit: people are "complex and nuanced and their mental models and thinking styles are not fixed personas or personalities. They change based on context."[11][12] A user's mental model of a health application differs when they are anxious versus calm, when they are managing prevention versus managing crisis. Any design tool that flattens this dynamism is producing a misrepresentation — not a research insight.
The Three Things Design Thinkers Get Wrong
Error One: Confusing the Solution Space with the Problem Space
Traditional design research asks: "How do users interact with our product?" This is a solution-space question. It produces solution-space data — feature preferences, interaction patterns, conversion drop-offs. Mental model research asks: "What are people trying to accomplish, and how do they think about accomplishing it — regardless of whether our product exists?"[18][19] This is a problem-space question. It produces structurally different data: the full cognitive landscape of human goals and needs, including the vast territories that no existing product addresses. The failure to make this shift is why so many AI products feel like they are solving the last problem.
Error Two: Treating Personas as a Proxy for Cognition
Personas are useful fictions — humanizing composites that give teams a relatable face for decision-making. But their structural limitations are severe, and in an AI era they become actively dangerous. Personas freeze dynamic behavior into static attributes. They anchor to demographics, which inadvertently encodes exclusion — choosing a face for a persona says "these are our users," implicitly marginalizing everyone not represented.[22] Most critically, they collapse variance: a persona averaging the behaviors of fifty research participants produces a character who may not actually exist in the real population.[20][21] Indi Young's critique is precise: "researching and designing for the majority or 'average user' actually ends up ignoring, othering, and harming the people our designs are meant to serve."[23] When that persona gets embedded as an alignment signal in AI training, the harm scales to every user the system was never built to serve.
Error Three: Designing for Behavior Instead of Interior Cognition
Journey maps, task flows, and usability tests all target the observable — what people do. Mental model research targets the interior: what people think and feel privately, the reasoning they run before action, the emotional reactions that shape interpretation, the guiding principles they apply without articulating.[13][16] This is the cognitive substrate that determines behavior. Designing without access to it is like building a city without understanding the water table. Everything looks fine until the infrastructure fails.
Tool
Operates In
Captures Cognition?
Dynamic?
Excludes?
Demographic Segmentation
Solution space
No
Static
By design
Marketing Persona
Solution space
Shallow
Static
Visually
UX Persona
Solution space
Moderate
Static
Yes
Archetype
Transitional
Moderate
Moderate
Less so
Mental Model
Problem space
Deep
Contextual
No
Thinking Style
Problem space
Deep
Dynamic
No
Source: Adapted from Indi Young's mental model methodology framework[15] and Nielsen Norman Group's comparative analysis[22]
Why This Breaks AI — At Scale
The design community's misunderstanding of mental models might be an inconvenience in the era of static interfaces. In the era of AI, it is a structural crisis.
Contemporary AI alignment relies heavily on Reinforcement Learning from Human Feedback (RLHF) — having human evaluators rank model responses, training a reward model to approximate those preferences, then fine-tuning the AI to maximize that reward signal.[27][28][29] The fundamental problem, identified in a landmark 2024 paper on MaxMin-RLHF, is this: a single reward model derived from aggregate preference data mathematically cannot represent the full distribution of human cognitive approaches.[30] The system is aligned to a statistical center. Everyone at the edges — every cognitive minority, every non-dominant thinking style — is structurally underserved.
This is the same failure mode as the persona. It is the average-user problem, now embedded in the weights of a model serving hundreds of millions of people.
There is a further dimension: what researchers are calling the DataSoul Imprint — the hidden layer of archetypal patterns, cultural narratives, and implicit hierarchies embedded in training corpora.[32] AI systems do not merely absorb facts. They absorb the psychological residue of their training data's origins: who was portrayed as protagonist, whose voice carries authority, which moral arcs were rewarded. These patterns crystallize into latent cognitive structures within the model. They are not post-training artifacts. They are pre-training inheritances. Because they are invisible to engineers, they are nearly impossible to correct downstream.[33][34]
MaxMin-RLHF research demonstrates a 16% improvement in overall win-rates and a 33% improvement for minority cognitive groups when alignment represents diverse human preference distributions — without degrading majority performance. Cognitive diversity in alignment is not a concession. It is an improvement.[30]
A 2026 arXiv paper goes further, arguing that alignment failure is structural, not incidental: AGI systems statistically internalize the full record of human social interaction, including power asymmetries, conflicts, and coercive arrangements.[31] Because human morality is plural, context-dependent, and historically contingent, the notion of a single "universally aligned" AI is ill-defined. Governance approaches must address cognitive and institutional amplification — the way AGI compresses human contradictions at scale.
What Cognitive Experience Design Demands Instead
The answer is not more personas. It is not better demographic segmentation. It is not faster design thinking sprints. It is a disciplinary shift toward problem-space research grounded in interior cognition — and a commitment to building mental model archetypes that can serve as a missing architectural layer in AI training.
Mental model archetypes — derived through rigorous listening sessions, thematic analysis of interior cognition, and the construction of what Indi Young calls the Mental Model Skyline — are behavioral signatures, not demographic profiles.[15][16] They capture how a cluster of humans approaches a shared goal: the cognitive scripts they run, the emotional triggers that shape their responses, the guiding principles they apply, the gaps between what they need and what any existing system provides. Unlike personas, they are grounded in observed interior cognition, not researcher projection. Unlike demographic segments, they cut across age, gender, culture, and income. Unlike average-user constructs, they do not disappear the people at the edges of the distribution.
The Mental Model Skyline — Young's visualization of cognitive patterns laid out like a city skyline — is a panoramic map of how people think across an entire problem space. The towers represent coherent cognitive purposes. The spaces where towers exist but no features or AI behaviors support them are cognitive gaps: unmet needs invisible to any product roadmap built on solution-space research.[14][15] Those gaps are the design frontier. They are also the alignment frontier.
Recent neuroscience provides empirical support: a 2025 Nature paper demonstrated that infusing neural networks with human hierarchical knowledge yields models that are simultaneously more aligned with human behavior and more robust on downstream tasks.[39] A large-scale analysis of alignment between human brain activity and AI representations found evidence for convergent evolution — artificial and biological systems evolving similar computational solutions despite architectural differences.[40] The path to robust, human-aligned AGI runs through a deeper understanding of human cognitive architecture. That is precisely what mental model research is designed to illuminate.
Here is what the shift demands, in practice:
Begin every AI design engagement in the problem space, not the solution space. Frame your research around human goals and intentions — not product interactions.
Conduct listening sessions, not usability tests. Surface interior cognition: the private reasoning, emotional reactions, and guiding principles that precede behavior.[25]
Build mental model archetypes from thematic analysis of that interior cognition.[26] Ground each archetype in behavioral signatures and cognitive patterns, not demographic attributes.
Map archetypes against your system's existing capabilities to produce a cognitive gap analysis — a strategic map of the minds you are not yet serving.[15]
Feed archetype diversity into your AI's alignment infrastructure: reward model training, constitutional principles, annotation guidelines. Cognitive diversity is not a UX afterthought. It is an alignment requirement.[36][37]
Design for cognitive sovereignty at every touchpoint — the user's capacity to reason independently alongside the AI, not merely through it.[7] Every interface that bypasses deliberative effort makes a person a little less capable of thinking for themselves.
The Design Field Needs to Grow Up
I have spent over a decade arguing that design programs and practitioners must move from the management fad of design thinking to the measurable rigor of designing for thinking and doing. The AI era makes that argument existential. We are not building apps anymore. We are building extensions of human cognition — systems that will shape how billions of people reason, decide, remember, and act for decades.
The tools we carry into that work matter enormously. If we carry personas, we will build AI that serves the average and ignores the edges. If we carry demographic segmentation, we will build AI that mistakes surface identity for interior cognition. If we carry the mental model as a synonym for user expectation, we will build AI that confirms what people already believe rather than genuinely supporting what they are actually trying to do.
Mental models are not a soft research artifact. They are, as Indi Young has argued, "strategic data we've been ignoring for too long."[15] They are also, I would add, the most important data we have yet to take seriously in building AI that is genuinely human-centered — not human-averaged.
The difference between those two things is the difference between a tool that augments human intelligence and one that quietly replaces it. That is the design choice in front of the field. I know which one I am building for.
