Research Framework

Pivot: Actionable Evaluation for Agentic AI

This updated framework shifts focus from static dark patterns to dynamic AI risks (Hypernudges) and centers on equipping designers with heuristic evaluation tools.

Framework Details (v2)

THEME 1: User-Level Context (Trust)

Focus: The "Halo Effect" vulnerability in older adults.

• Trust Transfer: Users transfer their trust in a reputable health brand (e.g., Mayo Clinic) to the AI agent.
• Lowered Defenses: Because of this "medical authority," older adults lower their privacy guards and skepticism.
• Vulnerability Mechanism: The user assumes the agent is purely benevolent, missing potential commercial or engagement-driven manipulation.

THEME 3: Agentic AI Capabilities

Focus: The new powers of "Agentic" systems (vs. static chatbots).

• Long-term Memory: The ability to remember preferences and medical history across sessions.
• Autonomy: The agent acts without explicit commands (e.g., suggesting a refill before asked).
• Semantic Commit: The ability to "lock in" user intent based on vague interactions.

THEME 2: The Mechanism (Dynamic Risk)

Focus: How "Hypernudging" exploits the Trust/Capability combination.

• Hypernudging (Duane et al., 2025): Unlike static dark patterns (a confusing button), hypernudges are data-driven, personalized, and real-time.
• Hidden Influence: The manipulation is invisible because it adapts to the user's current specific context and profile.
• Fluid Strategy: It's not a single UI element; it's a conversational strategy designed to optimize metrics (engagement/retention) over user wellbeing.

THEME 4: The Audit Gap

Focus: Why current governance fails.

• Principles vs. Practice: High-level principles (WHO/NIST) are abstract.
• Invisibility: Designers working on static screens (Figma) cannot "see" these dynamic behaviors.
• Lack of Tools: There are no standard "testing kits" for finding these hypernudges before deployment.

OUTPUT: Heuristic Audit Tool

The proposed contribution of this research.

• Test for Seams: Ensuring the AI's nature is visible to break the Halo Effect.
• Contestability Checks: Can the user challenge the AI's "advice"?
• Pre-deployment Stress Testing: A set of scenarios to trigger and detect hypernudging behaviors.

Four-Layer Concept Chain: Depth-First Literature Architecture

Week 6 establishes a stable four-layer architecture through Genealogical Tracing — tracing citation lineage rather than keyword search — yielding 25 interconnected sources.

Design Space

Design Space. The four‑layer architecture doubles as a risk‑aware design space for agentic health AI with older adults. Each layer corresponds to one core design axis that captures a key design tension (vulnerability, hypernudge mechanism, friction/seams, and contestability), and together they span the spectrum from individual susceptibility to governance‑level safeguards.

Framework Details (v3)

Layer 1: User Vulnerability & Trust

Focus: Why older adults are uniquely susceptible — and where on the trust spectrum they fall.

• Dimension A — Trust State (blind trust ↔ calibrated trust): The Halo Effect (Ellis 2025) drives institutional trust transfer from healthcare brands to AI agents, collapsing vigilance. Trust Calibration theory (Lee & See 2004) provides the target end-state: users who trust appropriately, neither over- nor under-relying on AI.
• Dimension B — AI Literacy (low AI literacy ↔ basic understanding): The AI Literacy Gap (Peixoto 2025) shows current XAI explanations assume younger, tech-literate users; cognitive decline creates "explanation blind spots" that reinforce blind trust.
• Supporting evidence: MOLD-US Framework (Wildenbos 2018) — Cognition and Perception barriers explain why simplified friction must precede complex transparency dashboards. Patient Portal Trust (Sakaguchi-Tang 2017) provides empirical grounding for how trust mis-calibration manifests in health IT contexts.

Layer 2: Hypernudge Risk Mechanism

Focus: The evolution from static nudge to dynamic AI manipulation — and its visibility.

• Dimension C — Manipulation Form (static nudge ↔ dynamic hypernudge): The Nudge → Hypernudge citation chain: Thaler & Sunstein (2008) → Yeung (2017) → Duane et al. (2025). Static dark patterns (Gray 2018; Brignull 2024) serve as baseline; hypernudges are algorithmically driven, real-time, and continuously updated — they bypass deliberative reasoning. (On the terminology shift from "dark patterns" to "deceptive patterns," see Week 7 Report §8.)
• Dimension D — Visibility (visible UI pattern ↔ invisible conversational strategy): Traditional dark patterns are identifiable UI elements (e.g., trick wording, hidden costs). Hypernudges in conversational AI operate as invisible persuasive strategies embedded in dialogue flow, tone, and framing — undetectable through conventional UI audit.

Layer 3: HCI Countermeasures — Friction & Seams

Focus: "Positive friction" as the direct antidote to hypernudge passivity, calibrated for older adults.

• Dimension E — Friction Type (elder‑friendly friction ↔ heavy-action friction): Cognitive Forcing Functions (Gull 2025) propose deliberate design friction that interrupts autopilot reliance on AI. The design axis runs from elder‑friendly friction (attention-layer seams: subtle pauses, confirmations, "AI is guessing" labels) to heavy-action friction (multi-step overrides, explicit consent gates). For older adults, the countermeasure converges on attention-layer seams — lightweight interruptions that make AI's nature visible without imposing cognitive burden.
• AI visibility as integrated seam: Transparency (showing AI's reasoning, confidence, data sources) is not a separate axis but an inherent property of elder-friendly friction. Every seam simultaneously reveals that "this is AI" and creates a micro-pause for reflection.
• Empirical context: Zubatiy (2023) and Brewer (2022) reveal specific trust-miscalibration patterns in older adults that inform where seams should be placed in conversational flows.

Layer 4: Governance & Contestability

Focus: Users must not only "see" AI but "challenge and correct" it — and governance must enforce this right.

• Dimension F — Contestability Timing (post-hoc only ↔ in-situ + log dual-layer): Pure post-hoc contestability (filing a complaint after harm) is inadequate for older adults. The design space demands in-situ contestability (real-time override, "why did you suggest this?" buttons) combined with persistent audit logs that enable retrospective review by the user or their advocate.
• User-as-auditor (via WeAudit citation chain): Rather than a separate axis, the user role spectrum (passive consumer ↔ active auditor) is concretized through the WeAudit-inspired workflow (Deng 2025, 2023a, 2023b): users (or their designated proxies) can flag problematic AI behaviors, which feed into governance review loops. This operationalizes contestability-by-design into a concrete safeguard.
• Design foundation: Hirsch (2017) — translates contestability from legal theory into HCI design principles.
• Supporting references: WHO (2021), NIST AI RMF (2023) — institutional governance anchors (policy/grey literature, not core).

Cross-Layer Logic

• Layer 1 → 2: Vulnerability (blind trust + low AI literacy) explains why invisible hypernudges are dangerous for older adults.
• Layer 2 → 3: Dynamic, invisible hypernudges demand countermeasures beyond static dark pattern audits — specifically, elder-friendly friction seams.
• Layer 3 → 4: Design friction is necessary but insufficient without governance mechanisms guaranteeing users' right to contest in real time and retrospectively.
• Layer 1 → 4 (Amplified Harm): Older adults require both layers of protection — attention-layer seams (L3) AND dual-layer contestability (L4).

From Four-Layer Architecture to Narrowed Research Question

Week 8 responds to supervisor feedback by narrowing the research direction. The four-layer architecture (V3) is retained as background context, but the organizing principle is now a focused, empirically tractable research question derived from thematic literature analysis.

The Narrowed Research Question

In trusted-institution contexts (e.g., hospital-endorsed health AI), how should trust calibration interventions and proxy auditor interfaces be designed so that users — including but not limited to older adults — can maintain appropriate skepticism without abandoning the system?

This question drops the claim that the halo effect is age-specific (it is a general human bias; Thorndike 1920), replaces "halo effect inoculation" with trust calibration intervention, and treats the proxy auditor as a concrete interface and role design problem — scoped to be empirically tractable within six weeks.

What Changed from V3

Thematic Literature Review (Ch 1–4)

Ch 1: Trust Formation in Institutional Health AI Contexts

What do we know about how users form trust when a health institution endorses an AI tool?

• Ellis 2025: Institutional trust transfer pattern (proposed, single study)
• Wong 2025: Authority-dependent trust / privacy resignation (HK, N=27)
• Guo et al. 2025: Trust flows AI → physician → hospital (reverse direction to halo hypothesis)
• Danish Health cohort 2025 (N=39,109): Older adults more skeptical, not less

Gap: No study has experimentally tested whether institutional endorsement causally increases uncritical trust, or whether a calibration intervention can shift it.

Ch 2: Friction and Calibration as Design Interventions

What design mechanisms exist for calibrating trust without causing abandonment?

• İnan 2025: Positive friction in dialogue systems
• Ehsan 2024: Seamful XAI — exposing AI limitations as trust calibration
• Pang 2021: Older adults' latent capability for trial-and-error learning
• Peixoto 2025: 76/79 XAI studies excluded disability needs — accessibility constraint

Gap: No friction intervention has been tested specifically as a trust calibration tool in institutional health AI contexts.

Ch 3: Proxy Auditing — Delegation, Collaboration, and Interface Design

What happens when audit responsibility is delegated to a proxy — and what can older adults actually contribute?

• Deng 2025 (WeAudit): Non-experts can produce useful audit insights, but sample was young volunteers
• Baghestani 2024: Mixed-age collaboration dynamics — dominant-follower failure in 50% of pairs
• Chong (supervisor's lab): Older adults showed strong capacity for bias identification
• KangJie et al. 2025: Older adults motivated to learn about AI risks; barriers are pedagogical, not capability
• Peng et al. 2024: Older adults co-designed health-misinformation chatbots, demanded tools that "teach them to judge for themselves"

Gap: No study has designed and tested a proxy auditor interface — the concept exists only as a paragraph in Deng 2025.

Ch 4: Temporal and Contextual Constraints

Why do we need to be cautious about any claim in this space?

• Laupichler 2024: No AI literacy scale tested for temporal validity
• Danish cohort 2022→2024: Attitudes shifted significantly in 2 years (OR 3.21)
• AARP 2024: Older adults simultaneously skeptical of AI and actively using it

Implication: Any design must be adaptive. Friction calibrated to 2024 literacy levels may be patronizing by 2027.

Design Targets

Design Target A: Trust Calibration Intervention

Replaces the speculative "Halo Effect Inoculation" from V3. Instead of assuming a halo effect specific to older adults, the intervention targets institutional trust transfer — a general phenomenon that occurs whenever a trusted entity (hospital, government) endorses an AI tool. The design question: what does the user see, when, and how does it shift their trust calibration?

• Supporting evidence: Ellis 2025 (proposed mechanism), Guo 2025 (reverse trust transfer), Danish cohort (skepticism as counter-evidence), independence-motivation literature
• Open question: Can a brief, non-disruptive friction intervention at the point of institutional endorsement measurably shift users toward calibrated trust?

Design Target B: Proxy Auditor Interface

Replaces the framework-level abstraction from V3. The proxy auditor is now framed as a concrete interface and role design problem: the older adult provides experiential signals (felt wrong, confusing), and a proxy handles the analytical audit work — not because the user has nothing to contribute, but because the roles are specialized.

The justification for proxy auditing is not that older adults are incapable of auditing, but that distributing roles can make auditing more effective and less burdensome in high-stakes, institutional AI settings.

• Supporting evidence: Deng 2025 (WeAudit), Baghestani 2024 (collaboration fragility), Chong (older adults can engage critically), KangJie 2025 + Peng 2024 (motivation and agency)
• Open question: What does the proxy auditor interface look like? How is the older adult's experiential signal captured without making them passive?

Advisor-Aligned Direction: Human-in-the-Loop AI Proxy Auditor

Version 5 operationalizes the latest advisor consensus: broaden proxy auditor beyond family, explicitly handle veneer-of-credibility risk, and prototype a capture-and-replay human support pattern for sustained trust calibration.

Consensus Integrated (from advisor meeting)

1. Proxy auditor is broadened: Not family-only delegation. The proxy set includes family caregivers, family physicians/clinicians, institutional reviewers, and trusted peer/community channels.
2. Veneer-of-credibility is first-class risk: "Many people use it" or "an authority recommends it" can create false safety perception without real security vetting.
3. Design target is HITL AI proxy auditor: Human helps once, AI replays that support pattern over time to reduce family/expert fatigue while preserving user agency.

The Refined Research Question (V5)

In institutional health AI contexts, how can a human-in-the-loop proxy-auditor system transform endorsement-based trust shortcuts into evidence-based trust calibration by surfacing verifiable risk checks and replaying prior human guidance?

RQ note: In V5, "caregiver" is a role family, not a single actor. It covers family caregivers, clinicians, institutional reviewers, and trusted community proxies so the evidence-check workflow can operate beyond one household support path.

Ch 2 & Ch 3 Endpoint Adjustments (V5)

Ch 2: From "Dark Pattern Mechanism" to "Credibility Veneer Mechanism"

• Scope expansion: Keep nudge/hypernudge analysis, but add endorsement cues (institution logo, physician mention, peer uptake, "official app" framing) as credibility accelerators.
• Core risk: Surface trust can rise faster than verified safety evidence.
• Design implication: Every endorsement cue must be paired with visible risk-check provenance (who checked what, when, with what method).

Ch 3: Broadened Proxy Auditor + Capture & Replay

• Proxy set: Family caregiver, clinician, institutional reviewer, and trusted community representative.
• HITL model: Human provides initial guided assessment; AI captures decision rationale and replay steps for future similar app evaluations.
• Role principle: AI supports endorsement quality; it does not replace physician/family judgment.

Deployment Scope Boundaries (embedded in Ch 3 ending)

• Auditability: Log endorsement source, risk-check artifacts, and uncertainty/confidence statements.
• Contestability: Users can inspect and challenge why an app was considered "safe enough."
• Boundaries: Require periodic re-checks and expiry for old endorsements to avoid stale credibility inheritance.

Why Design Target A Was Removed

Design Target A (veneer-resistant trust calibration) was removed to avoid splitting implementation effort across two intervention surfaces in a short project window. In V5, credibility-veneer mitigation is treated as a design constraint across Ch 2 and the deployment-boundary section at the end of Ch 3 (evidence provenance, audit trail, expiry/revalidation), not as a standalone target. This keeps the prototype scope coherent: one build target (Target B) with clear measurable outcomes.

Design Target B: Capture & Replay Human Support

This target reduces repeated caregiver/clinician burden by preserving one high-quality guided review as a reusable support trace for similar future app choices.

• Capture block: Record what the human checked, why it mattered, and what triggers escalation.
• Replay block: Re-run the same guidance flow with confidence labels and one-tap "ask human again."
• Evaluation metrics: Repeat human-help requests per decision, time-to-safe-decision, and user confidence calibration drift over time.

Interface Pattern: Capture & Replay Human Support

The core interaction pattern is a reusable support trace: a family member or physician co-reviews one app with the user, and the system records the review rationale as structured steps (warning signs checked, permissions flagged, unresolved risks). For later apps, the AI replays this trace as guided prompts, preserving human reasoning while minimizing repeated synchronous support.

• Capture: "What did the human check?" + "Why was this acceptable/not acceptable?"
• Replay: Stepwise prompts + confidence + "ask human again" escalation threshold.
• Outcome target: Better calibrated trust with lower caregiver repetition cost.