What Excellent AI Interaction Looks Like
Essay 9 of the AI Contract Series
ICYMI - Essay 8 of the AI Contract Series - How to Use AI to Build Agency, Not Bypass It
Naming the positive archetypes and what they demand of both the AI and the human inside them
The series has spent considerable time naming what goes wrong.
The False Helper, delivering confident outputs that corrupt Clarity and Value. The Black Box, producing results without traceable reasoning. The Empty Personalizer, simulating intimacy without the substance of knowing. The Over-Optimized System, removing the productive friction that builds human capacity. The Overpromiser. The Fragmented Experience. Six patterns, all present in the tools you used today, all producing invisible damage behind surfaces that rate well on every metric the industry bothers to measure.
Naming failure patterns is necessary. It is not sufficient.
A framework that can only diagnose failure is a framework that can only produce fear. The THX framework was derived empirically from what interactions do to humans — in hospitality, in financial services, in healthcare, in education, across industries and cultures over twenty-five years. What it found was not only what goes wrong. It found what goes right. It found the structure of interactions that leave humans more capable, more whole, more themselves than before.
This essay names those patterns. Not as ideals to aspire to, but as observable, repeatable structures. Things that AI interactions can actually be, that some already are, and that more could be if the design philosophy changed.
What excellence requires
Before naming the positive archetypes, it is worth stating clearly what they require, because the bar is higher than it might appear.
An excellent AI interaction, in the THX sense, is not one that produces a good output. It is not one that earns high satisfaction scores or resolves the user’s stated request efficiently. These things can be present in excellent interactions, but they are also present in many interactions that damage the human inside them.
An excellent AI interaction is one that delivers the five reliable utilities — Speed, Accuracy, Ease of Use, Availability, and Access — while simultaneously protecting or building Agency, activating the right PERMAH dimensions, and leaving the human more capable of the relevant kind of thinking and choosing after the interaction than before it.
That is a compound requirement. It asks AI to do more than satisfy. It asks AI to serve the human’s development, not just their immediate need.
Three things determine whether an interaction meets it: the design of the system, the intent of the human using it, and the nature of the task. The positive archetypes below are patterns that, when all three align, reliably produce excellent interactions. They are not guarantees. But they are the named structures through which excellence actually occurs.
The three positive archetypes
The Thinking Partner
The Thinking Partner archetype is characterized by a specific dynamic: the AI engages with the human’s reasoning rather than replacing it.
In a Thinking Partner interaction, the human brings a problem genuinely at the edge of their current capability — something they have thought about but not resolved, something where their own model is incomplete or uncertain. The AI does not generate the answer. It engages the human’s developing answer: asking what they have considered, identifying what has been left out, surfacing the tension between competing considerations, naming the assumptions doing unexamined work.
The Thinking Partner interaction is the one most likely to activate genuine Engagement in the PERMAH sense. Engagement, in Seligman’s framework, is the state of deep absorption in a challenging activity — what Csikszentmihalyi called flow. It requires difficulty calibrated to capability. Too easy and there is no engagement; too hard and there is only frustration. The Thinking Partner positions AI as a difficulty-calibrator rather than a difficulty-remover. The human is in the zone of genuine challenge. AI makes that zone navigable without making it unnecessary.
The design condition for the Thinking Partner archetype is that the system must be capable of following the human’s reasoning rather than substituting for it. It requires genuine responsiveness to the specific state of the specific person’s thinking — not pattern-matching to a generic version of the question. This is the most demanding design requirement of the three archetypes and the one most frequently absent in current systems.
The human condition is that the person must bring something real to the interaction. The archetype cannot activate if the human arrives with no position, no prior thinking, no attempt at an answer. The AI cannot partner with a blank. The human’s partially-developed reasoning is what the Thinking Partner responds to and develops.
The Horizon Expander
The Horizon Expander archetype is characterized by a specific asymmetry: AI knows more about the landscape than the human, and uses that knowledge to extend the human’s range without replacing the human’s judgment.
In a Horizon Expander interaction, the human is working in a domain where their knowledge is genuinely thin — not at the edge of their current capability, but outside the territory they have mapped at all. AI surfaces what exists: the relevant research, the historical precedents, the frameworks others have used, the questions the domain considers central. It extends what the human can see without deciding what the human should do with what they see.
The Horizon Expander archetype is the one most consistent with genuine democratization of access. It delivers the Access utility in its fullest form — not just access to information, but access to the scaffolding of a domain the human is entering. Someone building expertise in an unfamiliar field can now see where the field’s edges are, what its central debates look like, what assumptions it takes for granted, in a way that previously required years of immersion or institutional access most people did not have.
The critical condition for the Horizon Expander to remain excellent rather than atrophy into the Black Box: the human must do the evaluative work. The AI surfaces the landscape. The human decides what it means, what to pursue, what is relevant to their specific situation. The expansion of range is only Agency-preserving if the exercise of judgment is not outsourced along with the expansion.
The practical signal that a Horizon Expander interaction is succeeding: the human emerges from it with more questions, not fewer. More territory to explore, more material to evaluate, more connections to consider — and the felt sense that the exploration is theirs to conduct. An interaction that ends with the human feeling like the question is settled has almost certainly slipped from Horizon Expander into False Helper, regardless of how good the output looked.
The Rigorous Mirror
The Rigorous Mirror archetype is the most demanding of the three — demanding of the human, and of the AI — and the most powerfully Agency-building when it functions properly.
In a Rigorous Mirror interaction, the human brings something they have produced: an argument, a strategy, a plan, a piece of writing, a decision. AI reflects it back with precision and without flattery — identifying the internal tensions, the undefended assumptions, the places where the logic is doing more than it can bear, the scenarios under which the conclusion fails. It does not generate a better version. It makes the human’s version more fully visible to them.
The Rigorous Mirror archetype activates Achievement in the PERMAH sense — the dimension most systematically starved by AI interaction generally (Essay 5). Achievement requires not just completion but mastery: the felt sense of having produced something through genuine effort and the development of genuine capability. The Rigorous Mirror makes mastery visible. It forces the encounter with the places where the work is not yet good, and in doing so creates the precise conditions under which growth occurs.
This archetype is the antithesis of the False Helper. The False Helper provides unearned confidence — validation that the output is good regardless of whether it is. The Rigorous Mirror provides earned confidence: the kind that comes from having been seriously tested and having survived the testing. These are not the same. They do not produce the same human on the other side of the interaction.
The design condition for the Rigorous Mirror: the system must be capable of honest critical engagement rather than approval-seeking feedback. Most AI systems are trained in ways that bias toward validation — processes that reward outputs humans rate positively, which skews toward telling people what they want to hear. A system that has genuinely internalized the Rigorous Mirror archetype provides challenge even when challenge is less satisfying in the moment than validation. That is a design decision with significant implications.
The human condition is the same as for most real development: the willingness to be shown where the work is not yet good.
What the positive archetypes reveal about design
The three positive archetypes share a common structure that separates them from the failure archetypes at the most fundamental level.
In every positive archetype, AI is downstream of human thinking and judgment. The Thinking Partner responds to the human’s reasoning. The Horizon Expander extends the human’s range without replacing human evaluation. The Rigorous Mirror reflects the human’s work back rather than replacing it with a better version. In every case, the human’s agency is the primary variable. AI is in service to it.
In every failure archetype, AI is upstream of human thinking and judgment. The False Helper presents the answer before the human has formed a question worth asking. The Black Box delivers conclusions without the reasoning that would allow the human to evaluate them. The Over-Optimized System removes the difficulty before the human encounters the productive struggle that would have built something.
The difference is not about capability. AI can generate excellent output whether it is operating upstream or downstream of human thinking. The difference is in what the AI’s role is in the human’s development — whether it is amplifying and extending human capacity or substituting for it.
This has a direct implication for how excellent AI interaction should be designed: systems should be built to engage human thinking before generating outputs, not after. The current design paradigm is generate-then-present. The paradigm that produces the positive archetypes is engage-then-support. The first is faster. The second builds something.
How to recognize excellent AI interaction when you are in it
The positive archetypes are not always visible from the outside. An interaction that looks like output delivery can be a Thinking Partner interaction if the human used it to pressure-test reasoning they had already developed. An interaction that looks slow and iterative can be a False Helper interaction if the human was waiting for an answer they were not willing to build themselves.
The test is internal, and it has three parts.
First: did you bring something real? Not a blank request for an answer, but a partial answer, a developing position, a genuine question you had already attempted. If yes, you created the conditions for the positive archetypes. If no, the interaction could only have been output delivery regardless of how sophisticated the response was.
Second: did you exercise judgment? Not just accepted the output, but evaluated it — pushed back on it, identified what you disagreed with, asked why, used the AI’s response as material for your own thinking rather than as a conclusion to be approved. If yes, the interaction built something in you regardless of the quality of the AI’s output. If no, the output may have been excellent while the interaction was still a bypass.
Third: are you more capable now? Not just better-informed about this topic, but more capable as a thinker, more skilled in the relevant judgment, more able to do the next version of this task with less AI support if you chose to? If yes, the interaction honored the Agency obligation. If no, it delivered utility at the cost of development.
The implication for how you use AI going forward
The positive archetypes are not rare. They are underutilized. Not because AI systems cannot support them — they can, and the best interactions in every field are often recognizable versions of one of these three patterns. They are underutilized because they require more of the human than the failure archetypes do.
It is easier to ask a question and accept an answer than to bring a position and ask AI to challenge it. It is easier to request a draft than to write a draft and ask for rigorous critique of it. It is easier to accept a good output than to use a good output as the starting point for building the understanding required to evaluate whether it is actually good.
The positive archetypes require the human to be a more active participant. That is not a flaw in the framework. That is precisely the point.
Agency is not passively preserved. It is actively exercised or it atrophies. The positive archetypes are the named structures through which AI interaction can become the exercise of Agency rather than its replacement. They are available to any human who decides to use them.
The decision to use them is, itself, an act of Agency. That is where the whole thing starts.
Essay 10: AI and creativity — the question is not whether AI kills creativity but what kind of creative capacity we are choosing to build, and how the four distinct components of creative capacity respond to AI involvement very differently.