I’m treating adaptation as technology-agnostic at its core, but also using it as a scaffold for LLM-related patterns. I’ve also chosen to blend adaptation with prediction/generation, since they’re tightly connected in what I’m trying to do here. Does this make sense? I’m not sure.
A system’s capacity to modify its behaviour, presentation, or output in response to user input, context, or inferred intent. It spans both deterministic (rule-based) and probabilistic (prediction- or generation-based) responses.
Types of adaptation
Deterministic
UI changing based on application state, user preferences, roles, etc. Systems reacting to user inputs or past behaviours in predictable ways. If X happens, then Y will happen. You can still miss when a user has different expectations, but at least the system is consistent.
Probabilistic
Predictions, inference, or generated responses rather than fixed rules. It is driven by 🤖 models and produces context-sensitive results that may vary—even given the same input.
Integration patterns
Adaptation is connective tissue—it modulates the other layers. It introduces contextual variability into systems that usually prefer stability.
Layered adaptation
Hard on the outside, soft on the inside. Deterministic rules define boundaries, constraints, or guardrails. Probabilistic layers operate within those boundaries to personalise, infer, or generate.
Soft proposes, hard disposes
The probabilistic system suggests. The deterministic system decides. This is common when safety or trust is critical—predictions are filtered, ranked, or vetoed by fixed logic.
Confidence-based fallback
Adaptation paths vary based on model confidence. High-confidence outputs are used directly. Low-confidence predictions trigger a fallback—either deterministic logic or human intervention.
Agency spectrum mapping by confidence:
- 0-40% Passive: fallback to deterministic
- 40-70% Reactive: suggest with caveats
- 70-90% Proactive: act with notification
- 90%+ Co-operative: act autonomously
User control boundaries
Let users see or influence which parts of the system adapt. This builds trust and improves UX predictability. UI may expose overrides, lock behaviours, or show when outputs are model-driven.
Control preservation strategies:
- Intervention points: Explicit moments where users can modify system behaviour
- Granularity settings: Choose between high-level goals vs fine-grained control
- History: Track and revert system changes
Adaptive scaffolding
Adjust system support based on user expertise and context. More guidance for novices (see onboarding), more autonomy for experts (see mastery).
Progressive adaptation modes:
- Learning mode: Heavy deterministic guardrails, explicit explanations, limited AI autonomy
- Assisted mode: Balanced deterministic/probabilistic, contextual hints, shared control
- Expert mode: Minimal constraints, subtle suggestions, maximum flexibility
The system may track proficiency indicators (speed, error rates, feature usage) to automatically suggest mode transitions, though users always retain manual override.
Feedback loops
Adaptation improves through iterative refinement. The system learns from corrections, preferences, and outcomes.
Feedback mechanisms:
- Implicit: Observing which suggestions are accepted/rejected
- Explicit: Direct ratings, corrections, or preference settings
- Contextual: Environmental signals and task outcomes
Things to consider
- User model: the requirements for understanding the user and their context of use are much higher in adaptive systems
- Personalised response: adaptive systems go beyond static functionality, tailoring their responses to individual users and their specific needs
- Beware the bubble: adaptive systems should be able to handle new situations and user needs, even when these deviate from established patterns
- Recovery mechanisms: when adaptation fails, users need clear paths to correct or override system decisions
- Skill preservation: over-adaptation can lead to skill atrophy—maintain opportunities for users to exercise judgement