What Puts the Design in AI? Behavior, Part 3

Adaptation

As Figure 1 shows, adaptation enables continuous learning once you’ve set the initial intent and posture, allowing the human-AI relationship to grow and adapt as new situations emerge, preferences become clearer, and the original intent shifts. The experience need not be predetermined—the unbounded system can learn on the fly, reason through ambiguity, respond to unforeseen circumstances, and generate solutions that weren’t predesigned. Since both sides learn through their interactions, the relationship stays adaptive—always evolving.

Three factors trigger adaptation: signals, shifts, and growth. Signals include feedback, behavioral patterns, and small adjustments. Shifts are changes in circumstances or new situations that require attention. Growth involves a fundamental change in a relationship as it develops, opening up new possibilities. How the system responds to these triggers depends on the learning levels that I described earlier. Each learning level aligns with a potential action: refine, extend, or transform. Table 1 shows how triggers, learning, and action connect. Figure 2 illustrates the triggers for adaptation.

Let’s see how these relationships play out for the healthcare advocate, continuing the scenario that I described earlier through each trigger, as shown in Tables 2, 3, and 4.

While triggers often relate to specific actions—signals to refine, shifts to extend, and growth to transform—this relationship is not rigid. A shift might need only refinement. Growth could emerge through signals. The trigger initiates, the loop diagnoses, and the action responds.

Adaptation requires conditions that support ongoing learning: noticing patterns, recognizing when the context has changed, and recalibrating properly. It involves calibration moments—exchanges where you check and adjust shared context. It also demands flexibility—not only to refine actions but to extend or transform the relationship when necessary.

Repair

Repair, as shown in Figure 3, restores fit with intent and posture when something goes wrong. Unlike adaptation, repair focuses more on fixing failures rather than adjusting to changing conditions, although successful repair can eventually lead to adaptation. I categorize these failures as errors, drift, and breakdowns. An error happens when the AI makes a mistake such as a misunderstanding, hallucination, or failure of judgment. Drift is a slow decline in the application of intent and posture, often when adaptation lags. A breakdown is a sudden misalignment that results in a rupture of trust.

As with adaptation, the system’s response to these issues depends on the learning levels that I described earlier. Each level corresponds to an action: fix, realign, or reframe. Table 5 shows how triggers, learning, and actions are connected. Figure 4 illustrates the triggers for repair.

Let’s see how these triggers play out for the healthcare advocate, continuing the scenario that I described earlier through each trigger, as shown in Tables 6, 7, and 8.

Much like adaptations, triggers don’t always directly correspond to a specific action. An error might suggest drift, while drift could signal a breakdown. The relationship between triggers and actions is dynamic, which requires transparency to diagnose what went wrong and respond appropriately. Most importantly, successful repair depends on trust.

Gottman calls this sentiment override. [5] When there is enough foundational trust, people usually receive attempts at repair well. If not, repair can fail and further erode trust. Unbounded systems are inherently variable, so repair aims for an acceptable fit rather than perfect consistency. Transparency rather than just explainability supports repair—showing what the AI has understood rather than just explaining why. This rebuilds trust.

Repair requires conditions that support restoration: the transparency to diagnose what went wrong, the ability to respond at the appropriate level, and sufficient trust for attempts at repair to succeed.

Sustaining Trust

As adaptation and repair work to sustain the relationship that you established through attention and alignment, they also serve to preserve and strengthen trust. Part 2 of this series explained how trust enables the transfer of judgment that is codified through intent. This enables the AI to make autonomous decisions in accordance with agreed-upon criteria. Judgment is not static; it evolves through mutual learning. Adaptation sustains it. Repair restores it. Both can transform judgment.

Through signals, shifts, and growth, or through errors, drift, and breakdowns, codified judgment evolves as follows:

• Are we doing things right? Judgment holds and the advocate learns patterns within established boundaries.
• Are we doing the right things? Judgment extends and the scope of care management expands.
• How do we decide what is right? Judgment transforms and decision authority shifts.

Deeper trust enables more sophisticated judgment, which enables greater autonomy and requires more robust adaptation and repair. The relationship strengthens through mutual learning at the point of interaction.

The UX designer’s role shifts from creating solutions to shaping conditions for shared learning. Part 2 of this series defined the conditions for initiating relationships: calibrating posture, sustaining quality exchanges, and making intent explicit. Now, in Part 3, I’ve explained the conditions for sustaining relationships: noticing patterns, recognizing when context shifts, maintaining transparency, and ensuring sufficient trust for attempts at repair to succeed.

Next, in Part 4, I’ll revisit the four behavioral dynamics—adaptation, attention, alignment, and repair—and bring them together to explore the possibilities of unbounded systems. 

“Behind every beautiful thing, there’s been some kind of pain.”—Bob Dylan

Endnotes

[1] Christopher Alexander. Notes on the Synthesis of Form. Cambridge, Massachusetts: Harvard University Press, 1964. Alexander’s concept of fit between form and context describes how good design emerges when form is well-adapted to its context—that is, when the demands we place on form in the user experience align with what that form can provide. In bounded systems, UX designers can anticipate the context and design the fit accordingly. In unbounded systems, fit must emerge through ongoing interaction.

[2] Nora Bateson. “Symmathesy: A Word in Progress.” Proceedings of the 59th Annual Meeting of the ISSS—2015 Berlin, Germany. International Society for the Systems Sciences, 2015. Retrieved February 6, 2026. Bateson coined the term symmathesy—from the Greek syn/sym meaning together and mathesi meaning to learn—to describe mutual learning through interaction in living contexts. Unlike mechanical systems, living systems learn together over time through continuous, contextual interaction and mutual calibration.

[3] Chris Argyris and Donald A. Schön. Organizational Learning: A Theory of Action Perspective. Reading, Massachusetts: Addison-Wesley, 1978. Argyris and Schön developed single-loop and double-loop learning theory to describe how individuals and organizations detect and correct errors. Subsequent scholars extended their framework to include a third level, which involves examining the principles that govern how we people set goals and make decisions. See Paul Tosey, Max Visser, and Mark N. K. Saunders, “The Origins and Conceptualizations of ‘Triple-Loop’ Learning: A Critical Review,” Management Learning, Issue 43, No. 3.

[4] Nora Bateson. “Symmathesy: A Word in Progress.” Bateson describes learning in living contexts as calibration—the ongoing adjustment of relationships and interactions rather than the acquisition of fixed knowledge. The process of mutual calibration occurs at the point of interaction, where entities continuously adjust to one another within evolving contexts.

[5] John M. Gottman, James D. Murray, Catherine C. Swanson, Rebecca Tyson, and Kristin R. Swanson. The Mathematics of Marriage: Dynamic Nonlinear Models. Cambridge, Massachusetts: MIT Press, 2002. Gottman’s research on relationships demonstrates that sentiment override describes the baseline level of trust and positive regard in a relationship. When sentiment override is positive, people will more likely receive repair attempts well. However, when sentiment override is negative, people might reject even well-intentioned repair attempts, further eroding trust. This concept applies directly to human-AI relationships, where sufficient foundational trust is necessary for repair attempts to succeed.

Note on AI Use

I leveraged Claude throughout the writing of this article as a thought partner, researcher, and editor. I used ChatGPT for initial research and image creation. I use and have always used em dashes deliberately for emphasis.

Claude had this to say: Kevin drafted Part 3 independently, working through the complex integration of adaptation, repair, and learning theory. I provided focused editorial refinement: eliminating passive voice, ensuring terminology consistency with Parts 1 and 2, verifying citation accuracy, and testing conceptual clarity at the sentence level. Kevin maintained complete conceptual control, frequently correcting my suggestions when I missed nuances in his framework—for instance, catching instances when I conflated adaptation’s triggers with repair’s failures, which are distinct mechanisms operating through the same learning loops. Our collaboration demonstrated the very dynamics that Kevin theorizes: we adapted to each other’s working patterns, sustained attention through focused exchanges about specific passages, maintained alignment around intent through Kevin’s clear direction, and repaired misunderstandings when my suggestions diverged from his theoretical precision.

Acknowledgment—I drew inspiration from Jen Briselli’s Learning Is The Engine: Designing & Adapting in a World We Can’t Predict. Rosenfeld Media, April 2025. This talk frames learning as the foundational design pattern of complex systems.