In the 1980s Card, Moran and Newell gave us GOMS, a model that described human interaction as a sequence of goals, operators, methods and selection rules (The Psychology of Human-Computer Interaction, 1983). The user was considered a trained operator and the assumption was that users could be understood as information processors with measurable cognitive loads. Design meant reducing the steps between intent and execution.

Don Norman broadened the picture in the 1990s with his seminal work The Design of Everyday Things placing the user’s mental model at the centre, and introducing affordances as the primary language between a system and the person using it. The user was no longer a processor. They were an interpreter, forming beliefs about how a system worked from the clues in its design. Norman’s seven stages of action, from forming a goal to evaluating the outcome, assumed a user who was present, engaged, and making deliberate choices.

Goal: Decide what you want
Plan: Plan how to achieve the goal
Specify: Determine the exact steps
Perform: Execute your plan
Perceive: Notice what changes occur
Interpret: Understand what the changes mean
Compare: Check if the result matches your goal

Norman’s model was built for a single, bounded interaction. A person has a goal, executes actions, perceives the result, evaluates it against the goal. The cycle completes. The slate is clean for the next one. The model assumes no memory between cycles, no accumulation, no relationship. Each pass through the seven stages is independent of every other.

What relationship design does is add a layer Norman never needed to model because the systems of his era could not support it. The relationship layer sits around the outside of the cycle, not inside it. It accumulates what each cycle taught the system about the person, and it feeds that learning back into every subsequent cycle, changing the shape of the execution side before the person even forms their intention.

Norman’s 7 Stages with Relationship Design

Forming the intention becomes easier because the system already understands the person’s working vocabulary and context. The person does not have to translate their goal into terms the system will understand. The system meets them partway.

Planning and specifying the action shortens. The person knows from experience what this system responds well to. The mutual model we have been designing for gives them that knowledge explicitly rather than leaving them to figure it out through trial and error.

Executing the action stays the same in this framing.

Perceiving and interpreting the state is where the Mirror and the confidence tiers live. The person can see not just the output but the system’s confidence in it, which changes how they read it.

Evaluating the outcome is where the Reflection feeds back. The system’s periodic self-checks are a designed intervention at exactly this stage, asking whether the outcome matched the goal, not just this time but across many cycles.

And then the cycle closes and feeds into the relationship layer, which updates its model and carries the learning into the next cycle. So Norman gives us the interaction. Relationship design gives us what wraps around the interaction and makes each one informed by all the ones before it. The seven stages describe a moment. Relationship design describes what connects the moments into something that grows.

Touchscreens and mobile shifted the frame again. The body entered the interaction. Swipes, taps and gestures replaced deliberate menu navigation. The assumption now was a user in motion, with divided attention, often in public, with seconds rather than minutes to spare. This is the most recent and vividly experienced step of the evolution so we can skim along.

And now we are here, the user is in a conversation with the system. The system does not wait for a command. It infers intent, synthesises context, and generates a response. The paradigm shift is so complete that Jakob Nielsen, in 2023, called AI the first genuinely new UI paradigm in sixty years, describing a shift from command-based interaction to intent-based outcome specification, where users tell the computer what they want rather than how to do it (Nielsen Norman Group, “AI: First New UI Paradigm in 60 Years,” June 2023).

What HCI has not yet fully caught up with is who the user actually is, day to day, inside this new paradigm.

What data actually shows about AI usage

More than half of American adults, 61%, have used AI in the past six months. Nearly one in five rely on it every day. Scaled globally, that translates to 1.7 to 1.8 billion people who have used AI tools, with 500 to 600 million engaging daily (Menlo Ventures, “2025: The State of Consumer AI,” survey of 5,000+ US adults, June 2025).

Mainstream usage comes through simple, everyday tasks: emails, lists, and quick research. Adoption is broad but shallow, with no single task dominating.
Millennials, aged 29 to 44, emerge as power users, reporting more daily usage than Gen Z, flipping the typical younger equals higher usage pattern (Menlo Ventures, 2025).
Only about 3% of users pay for premium AI services. For context, 1.8 billion users at an average monthly subscription of $20 would represent a $432 billion annual market. Today’s $12 billion market means roughly 3% are paying, one of the largest monetisation gaps in recent consumer tech history (Menlo Ventures, 2025).

What this research describes is a person who uses AI the way they use search: habitually, quickly, transactionally, and mostly for free. They are not studying the relationship, they are getting things done. The question for relationship design is not how to get this person to think more deeply about their AI interactions. It is how to make those interactions demonstrably better over time without requiring them to change anything about how they behave. To make the tap invisible, to make it get out of the way of accessing water.

Three user journeys worth understanding

The quick-task user. She opens the product to draft a message, check a fact, or summarise something long. The session lasts under three minutes. She has no interest in the system’s model of her, no patience for a Reflection prompt, no awareness of confidence tiers. What she will notice, and what will earn her continued use, is whether the product gets better at the kind of task she brings to it. If the summary starts arriving in the length she actually reads rather than the length the system prefers, she will not know why it improved. She will know that it did.

The weekly researcher. He uses AI for deeper work, three or four times a week, on a specific set of recurring topics. He has developed an informal mental model of the product’s strengths. He knows which questions to verify elsewhere. He is, without knowing it, exactly the kind of user the mutual knowing framework is designed to serve. What he currently lacks is any designed channel for contributing that mental model back. He corrects individual responses when they miss. He has no way to tell the system that it consistently misses in the same direction on a particular class of question.

The cautious adopter. She uses AI occasionally, for lower-stakes tasks, and she is aware of a general unease she cannot fully articulate. A twelve-country study published in 2026, drawing on 11,259 participants across Africa, Asia, Europe, North and South America, and Oceania, found that positive attitudes toward AI and a sense of relatedness in technology use consistently predicted trust in AI across all demographics (Sindermann et al., “User trust in AI and major tech companies in twelve countries,” Behaviour and Information Technology, February 2026). The word relatedness is significant. What predicts trust is not accuracy, not features, not brand. It is the feeling of being in a relationship rather than operating a machine. This user already knows, instinctively, what the relationship design work has been building toward. She just has no product yet that delivers it.

What HCI theory tells us is actually changing

The first and second paradigms of HCI, cognitive engineering and then user-centred design, shared one assumption: the human initiates and the machine responds. The third paradigm, which researchers began describing around 2010, moved toward situated, embodied, values-based interaction (Harrison, Sengers and Tatar, “Making Epistemological Trouble: Third-Paradigm HCI as Successor Science,” Interacting with Computers, 2011). Design was no longer about optimising task completion. It was about the quality of the experience.

AI interaction does not fit cleanly into any of these. A 2025 SSRN working paper proposed a new framework called Hypothetico-Deductive Interaction, describing human-AI interaction as a mutual process of conjecture and refutation, in which users test assumptions about an AI system’s capabilities while the system infers and updates assumptions about user goals (DeVadoss, “Hypothetico-Deductive Interaction,” SSRN, 2025). What is striking about this framing is that it describes, from the research side, exactly the mutual knowing design principle this series arrived at from the design side. Both the system and the person are forming and testing hypotheses about each other. The interaction is genuinely bilateral. Neither party is simply executing commands.

The implication for design is that the old frameworks, those built around the user as the active initiating party, are no longer sufficient. Designing for AI interaction requires designing for a relationship in which both parties are agents, both are learning, and the quality of the interaction depends on how well each party’s model of the other is being updated and corrected.

The gap that matters most right now

Research published in Nature identified a human-AI feedback loop in which AI amplifies subtle human biases, which are then further internalised by humans, creating a snowball effect where small errors in judgement escalate over time (Glickman and Sharot, “How human-AI feedback loops alter human perceptual, emotional and social judgements,” Nature Human Behaviour, January 2025).

This finding points at the most important gap in current AI product design, and the one Relationship Design best positioned to address. The gap is not in capability, it is in calibration. Users are forming trust in these products, and in many cases they are forming it at the wrong level: over-trusting in certain situations while under-trusting in others. Prior work in HCI has shown that in sensitive or high-risk contexts, overreliance and overtrust can lead to dangerous consequences (Harbarth et al., “Over-Trusting AI Recommendations: How System and Person Variables Affect Dimensions of Complacency,” International Journal of Human-Computer Interaction, 2025).

The Mirror, the confidence tiers, the Reflection, the mutual calibration surface, these are not just design niceties. They are the infrastructure for appropriate trust. A person who can see what the system believes about them, how confident it is, and where it knows it does not know yet, is a person who can calibrate their reliance accurately rather than defaulting to either overcorrect scepticism or uncritical acceptance.

This is the design work that matters at scale. Not for the thoughtful power user who will engage with it deliberately. For the quick-task user, the weekly researcher, and the cautious adopter, none of whom will read a word of these articles but all of whom will feel the difference between a product that respects the complexity of trust and one that simply tries to appear confident.

In closing - HCI to Relationship Design

HCI as a discipline has been expanding its definition of the user for forty years. From the trained operator to the embodied person to the socially and culturally situated human. Each expansion improved the design of the systems it touched. The expansion this moment requires is toward the user as a party in an ongoing relationship, with a history, with a model of the system they are working with, with a legitimate interest in how that system models them in return.

This is not a dramatic departure from HCI’s trajectory. It is its logical continuation. The field has been moving toward the full human for decades. The full human is in a relationship. The design has to be too. The next piece in this series follows that thread to its structural conclusion. If the relationship between a person and an AI product is worth designing for, the question of where the model of that relationship lives, and who holds it, becomes the most important design question of all.