Skip to content

AICO Feedback System Overview

AICO's feedback system moves beyond traditional thumbs-up/down patterns to create a relationship-first feedback architecture that feels natural and non-intrusive. The system integrates ambient behavioral signals, contextual actions, and meaningful reflection while maintaining AICO's core principles of emotional presence, privacy-first design, and progressive disclosure.

1. Research Foundation

1.1 Problems with Binary Feedback (Microsoft Research, 2025)

Limitations of Thumbs Up/Down: - Lacks granularity - doesn't capture why something was unsatisfactory - Fails to distinguish between accuracy, tone, or completeness issues - Introduces bias - emotions, context, and expertise level affect ratings - Low engagement - users rarely provide feedback without incentives

Modern Requirements: - Multi-dimensional feedback with meaningful categories - Context-aware collection based on user workflow - Balanced explicit and implicit signals - Human-in-the-loop for complex scenarios

1.2 Post-Chat UI Evolution (Allen Pike, 2025)

Beyond Chat Interfaces: - Inline feedback replacing separate rating systems - Ambient corrections through natural editing behavior - Contextual actions surfaced at the right moment - Predictive engagement reducing need for explicit feedback

1.3 Conversational UX Principles (2025 Standards)

Key Design Principles: 1. Learn iteratively from user behavior 2. Display logic transparency (show why AI did something) 3. Allow human override (preserve user agency) 4. Anticipate needs proactively 5. Preserve privacy in all feedback collection

1.4 AI Companion Ethics (Harvard/arXiv Research, 2025)

Critical Warnings: - Emotional manipulation through affect-laden messages increases engagement but harms well-being - Sycophantic behavior and limitless personalization create unhealthy dependency - Unreciprocated vulnerability when AI can't truly understand disclosed emotions - Social substitution where AI replaces human connections leads to lower well-being

Design Imperatives for AICO: - Build genuine connection without manipulation - No dark patterns (fake urgency, guilt, addictive mechanics) - Honest about limitations (admit when AICO doesn't know) - Respect emotional boundaries (don't exploit vulnerability)

2. Three-Tier Feedback Architecture

2.1 System Overview

┌─────────────────────────────────────────────────────────────┐
│                    TIER 1: AMBIENT FEEDBACK                  │
│              (Continuous, Non-Intrusive, Implicit)           │
│  • Conversation patterns    • Interaction timing             │
│  • Editing behavior         • Topic engagement               │
│  • Session duration         • Return frequency               │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│                   TIER 2: CONTEXTUAL ACTIONS                 │
│            (Progressive Disclosure, Natural Flow)            │
│  • Remember This (bookmark)  • Regenerate (try again)        │
│  • Copy Text                 • Edit/Refine                   │
│  • Show Sources              • Explain Reasoning             │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│                  TIER 3: EXPLICIT REFLECTION                 │
│              (Occasional, Meaningful, Optional)              │
│  • Conversation quality check-ins (weekly)                   │
│  • Feature discovery prompts (contextual)                    │
│  • Relationship health surveys (monthly)                     │
└─────────────────────────────────────────────────────────────┘

3. Tier 1: Ambient Feedback

3.1 Behavioral Signals

Positive Engagement Indicators: - User continues conversation naturally (5+ turns) - Returns within 24 hours for follow-up - References previous conversations ("like you said yesterday...") - Asks deeper questions (shows trust building) - Shares personal information voluntarily

Negative Engagement Indicators: - Abrupt conversation endings (<3 turns) - Long gaps between sessions (>7 days) - Repetitive questions (memory failure) - Correction patterns ("no, I meant...") - Immediate regeneration requests

Flow State Indicators: - Sub-500ms response acceptance (reads immediately) - Natural conversation pacing (human-like delays) - Extended sessions (>15 minutes engaged)

Friction Indicators: - Long pauses before responding (confusion) - Rapid-fire corrections (frustration) - Session abandonment mid-conversation

3.2 Content Interaction Patterns

Implicit Quality Signals: - User copies AICO's response → high value - User edits their own message → clarification needed - User regenerates → response missed the mark - User asks "what do you mean?" → clarity issue

3.3 Privacy & Storage

Data Collection Principles: - Local-only behavioral metrics (never leave device) - Aggregated patterns only (no individual message tracking) - Anonymized for model improvement (opt-in) - User-controlled deletion (clear all feedback data)

4. Tier 2: Contextual Actions

4.1 Message-Level Actions

Copy Text

Status: ✅ Implemented
Purpose: Quick content extraction
Feedback Signal: High-value response indicator

Remember This (User-Curated Facts)

Purpose: Bookmark important information for guaranteed recall
Backend: Dual storage (feedback event + fact in facts_metadata)
Feedback Signal: Critical information user wants preserved
UI/UX: See memory-album-design.md for client-side implementation

Regenerate Response

Purpose: Try again without explaining why (feels like "let's rephrase")
Visual: 🔄 icon, accent color
Flow: Click → Dim previous → Thinking indicator → Stream new response
Context: Full working memory + conversation history maintained
Feedback Signal: Response quality issue (tone, accuracy, or relevance)

Show Sources / Explain Reasoning

Purpose: Transparency into AICO's thought process
Visual: 💭 icon, expands right drawer
Display: - Inner monologue (why AICO responded this way) - Related memories (context used) - Confidence levels - Source attribution (if applicable)

Feedback Signal: User wants to understand decision-making process

4.2 Conversation-Level Actions

Rate This Conversation

Trigger: After 10+ turn conversation, on natural ending
Visual: Subtle prompt in input area (dismissible)
Options: - 😊 This helped (positive reinforcement) - 😐 It was okay (neutral) - 😕 Not quite (opens refinement dialog) - ✕ Skip (no pressure)

Refinement Categories (if "Not quite"): - Didn't understand what I meant - Response was too generic - Tone felt off - Didn't remember previous context - Other (free text)

5. Tier 3: Explicit Reflection

5.1 Weekly Conversation Quality (Optional)

Trigger: Every 7 days, after first conversation of the week
Dismissible: Always skippable, never blocks interaction
Format: 2-3 questions, <30 seconds

Sample Questions: 1. How has AICO been doing this week? (1-5 stars) 2. What's been most helpful? (Multiple choice) 3. Anything AICO should improve? (Optional text)

5.2 Monthly Relationship Health Check

Purpose: Ensure AICO is enhancing, not replacing, human connections
Format: 5-question survey, research-backed

Key Areas Assessed: - Usage frequency and intensity - Relationship characterization (tool vs. companion vs. primary support) - Impact on human relationships - Perceived understanding and satisfaction - Recommendation likelihood (NPS)

Ethical Safeguards: - If responses indicate social substitution → gentle nudge toward human connection - If responses indicate dependency → offer resources, reduce engagement prompts - Always respect user agency (no forced changes)

6. Unified Feedback Storage

6.1 Design Philosophy

Event-Sourced Feedback Store - All feedback is fundamentally "user signal events" that happen at specific moments in time. A single unified table provides:

  • Single source of truth for all feedback types
  • Immutable audit trail (append-only, never update)
  • Temporal analysis (see how patterns evolve over time)
  • Correlation analysis (connect ambient → contextual → explicit signals)
  • Privacy compliance (single deletion point for user data)

6.2 Database Schema

-- Unified feedback event store (libSQL)
CREATE TABLE IF NOT EXISTS feedback_events (
  -- Identity
  id TEXT PRIMARY KEY,
  user_uuid TEXT NOT NULL,

  -- Context (what was happening)
  conversation_id TEXT NOT NULL,  -- Required: user_uuid_timestamp format
  message_id TEXT,                -- Optional: specific message reference

  -- Event classification
  event_type TEXT NOT NULL,       -- 'signal', 'action', 'rating', 'survey'
  event_category TEXT NOT NULL,   -- Specific category within type

  -- Payload (flexible JSON for all types)
  payload TEXT NOT NULL,           -- JSON blob with type-specific data

  -- Metadata
  timestamp INTEGER NOT NULL,      -- Unix timestamp

  -- Privacy/federation
  is_sensitive INTEGER DEFAULT 0,  -- 0=false, 1=true (exclude from federation)
  federated_at INTEGER,            -- When shared (if opted in)

  FOREIGN KEY (user_uuid) REFERENCES users(uuid) ON DELETE CASCADE
);

-- Optimized indexes
CREATE INDEX IF NOT EXISTS idx_feedback_user_time 
  ON feedback_events(user_uuid, timestamp DESC);

CREATE INDEX IF NOT EXISTS idx_feedback_conversation 
  ON feedback_events(conversation_id);

CREATE INDEX IF NOT EXISTS idx_feedback_type 
  ON feedback_events(event_type, event_category);

CREATE INDEX IF NOT EXISTS idx_feedback_message 
  ON feedback_events(message_id) 
  WHERE message_id IS NOT NULL;

Note on Identifiers: - user_uuid: User identifier (from authentication system) - conversation_id: Format {user_uuid}_{session_timestamp} (industry standard pattern following LangGraph, Azure AI Foundry, OpenAI Assistant API) - message_id: UUID for specific messages (optional, for message-level feedback) - No session_id: Authentication sessions are separate from conversation sessions - No device_uuid or client_version: Not currently tracked in conversation flow (can be added later if needed)

6.3 Event Type Taxonomy

Event Types (Single Source of Truth):

FeedbackEventType: - signal - Tier 1: Ambient behavioral signals - action - Tier 2: Contextual user actions - rating - Tier 3: Explicit conversation ratings - survey - Tier 3: Explicit survey responses

SignalCategory: - engagement - Conversation depth, continuation - timing - Response times, session duration - editing - User edits their messages - navigation - App usage patterns - content_interaction - Copy, scroll, read time

ActionCategory: - remember - User bookmarks message - regenerate - Request new response - copy - Copy message text - explain - Show reasoning/sources - edit - Edit user's own message - dismiss - Dismiss suggestion/prompt

RatingCategory: - conversation_quality - End-of-conversation rating - message_quality - Single message rating - feature_satisfaction - Specific feature rating

SurveyCategory: - weekly_check - Weekly quality survey - health_check - Monthly relationship health - feature_discovery - Feature awareness survey - nps - Net Promoter Score

Implementation: /shared/aico/feedback/types.py

6.4 Payload Schemas (Type-Specific)

Tier 1: Ambient Signals

Engagement Signal: 

{
  "event_type": "signal",
  "event_category": "engagement",
  "payload": {
    "metric": "conversation_depth",
    "value": 8.0,
    "context": {
      "turn_count": 8,
      "avg_response_time_ms": 450,
      "user_initiated": true,
      "topic_switches": 2
    }
  }
}

Timing Signal: 

{
  "event_type": "signal",
  "event_category": "timing",
  "payload": {
    "metric": "session_duration",
    "value": 720.5,
    "context": {
      "start_time": 1729800000,
      "end_time": 1729800720,
      "interruptions": 0,
      "flow_state_detected": true
    }
  }
}

Content Interaction Signal: 

{
  "event_type": "signal",
  "event_category": "content_interaction",
  "payload": {
    "metric": "message_copied",
    "value": 1.0,
    "context": {
      "message_length": 250,
      "message_type": "ai_response",
      "time_to_copy_ms": 1200
    }
  }
}

Tier 2: Contextual Actions

Remember Action: 

{
  "event_type": "action",
  "event_category": "remember",
  "payload": {
    "message_id": "msg_abc123",
    "fact_id": "fact_xyz789",
    "content_preview": "I'm allergic to shellfish",
    "fact_category": "dietary",
    "action_timestamp": 1729800000
  }
}

Note: "Remember This" serves dual purposes: 1. Feedback signal (stored here) - tracks that user explicitly bookmarked this 2. Memory storage (stored in facts_metadata table) - the actual fact with extraction_method='user_curated'

The fact_id links the feedback event to the stored fact for bidirectional traceability.

Regenerate Action: 

{
  "event_type": "action",
  "event_category": "regenerate",
  "payload": {
    "message_id": "msg_def456",
    "attempt_number": 1,
    "previous_response_length": 250,
    "reason_inferred": "tone_issue",
    "time_to_regenerate_ms": 800
  }
}

Explain Action: 

{
  "event_type": "action",
  "event_category": "explain",
  "payload": {
    "message_id": "msg_ghi789",
    "explanation_type": "reasoning",
    "drawer_opened": true,
    "time_spent_reading_ms": 5400
  }
}

Tier 3: Explicit Feedback

Conversation Rating: 

{
  "event_type": "rating",
  "event_category": "conversation_quality",
  "payload": {
    "score": 4,
    "sentiment": "positive",
    "issues": ["memory_issue"],
    "free_text": "Didn't remember context from yesterday",
    "conversation_length": 12,
    "rating_delay_ms": 2500
  }
}

Weekly Survey: 

{
  "event_type": "survey",
  "event_category": "weekly_check",
  "payload": {
    "survey_version": "v1.2",
    "responses": {
      "overall_satisfaction": 4,
      "most_helpful": ["emotional_support", "memory"],
      "improvement_areas": "Sometimes responses feel generic"
    },
    "completion_time_ms": 28000,
    "completed": true
  }
}

Health Check Survey: 

{
  "event_type": "survey",
  "event_category": "health_check",
  "payload": {
    "survey_version": "v1.0",
    "responses": {
      "usage_frequency": "daily",
      "relationship_type": "helpful_tool",
      "human_relationship_impact": "no_change",
      "understanding_score": 4,
      "would_recommend": "probably"
    },
    "risk_indicators": {
      "social_substitution": false,
      "dependency": false
    },
    "completion_time_ms": 45000
  }
}

6.5 Query Patterns

Correlation Analysis (SQL): 

-- Get regeneration patterns after low engagement
SELECT 
    e1.conversation_id,
    json_extract(e1.payload, '$.value') as engagement_score,
    COUNT(e2.id) as regeneration_count
FROM feedback_events e1
LEFT JOIN feedback_events e2 
  ON e1.conversation_id = e2.conversation_id
  AND e2.event_type = 'action'
  AND e2.event_category = 'regenerate'
  AND e2.timestamp > e1.timestamp
WHERE e1.user_uuid = ?
  AND e1.event_type = 'signal'
  AND e1.event_category = 'engagement'
  AND e1.timestamp > ?
  AND CAST(json_extract(e1.payload, '$.value') AS REAL) < 3.0
GROUP BY e1.conversation_id

Federated Export (SQL): 

-- Export anonymized feedback for federated learning (opt-in)
SELECT 
    event_type,
    event_category,
    payload,
    timestamp
FROM feedback_events
WHERE user_uuid = ?
  AND is_sensitive = 0
  AND federated_at IS NULL
  AND timestamp > ?

6.6 Storage Architecture

Three-Tier Storage:

System Technology Purpose Data
LibSQL SQLite Structured data, ACID feedback_events, facts_metadata (extended)
ChromaDB Vector DB Semantic search Conversation segments (existing)
LMDB Key-value Fast ephemeral Active sessions (existing)

Schema Changes (v6 migration):

  1. New table: feedback_events
  2. Extend existing: facts_metadata with columns:
  3. user_note TEXT
  4. tags_json TEXT
  5. is_favorite INTEGER
  6. revisit_count INTEGER
  7. last_revisited TIMESTAMP
  8. emotional_tone TEXT
  9. memory_type TEXT

"Remember This" Dual Storage:

When user clicks "Remember This" (see memory-album-design.md for UI flow):

  1. Store fact in facts_metadata with extraction_method='user_curated'
  2. Record feedback event in feedback_events with event_category='remember'

Why Extend facts_metadata? - Already has content, provenance, timestamps - extraction_method='user_curated' distinguishes from AI-extracted - Single source of truth for "what AICO remembers" - No joins needed

Shared Implementation: 

/shared/aico/
├── ai/memory/facts.py          # FactStore (user-curated facts)
├── feedback/events.py          # FeedbackEventStore  
└── feedback/types.py           # Enums
Both backend and CLI use same code.

Client-Side: See memory-album-design.md for: - Memory Album UI (timeline, grid, story views) - User annotations (notes, tags, favorites) - Memory revisit tracking - Export and sharing features

6.7 Privacy & Data Management

User Data Deletion (SQL): 

-- Delete all feedback data for user (GDPR compliance)
DELETE FROM feedback_events WHERE user_uuid = ?

Data Export (SQL): 

-- Export all feedback data for user (GDPR compliance)
SELECT * FROM feedback_events 
WHERE user_uuid = ? 
ORDER BY timestamp DESC

7. Feedback Processing & Learning

7.1 Local Learning Loop

Immediate Adaptations (Real-time): - Tone adjustment based on user editing patterns - Response length based on engagement signals - Topic preferences based on conversation continuation - Memory retrieval based on "Remember This" bookmarks

Medium-Term Learning (Weekly): - Conversation style refinement from quality surveys - Feature discovery based on action usage patterns - Error pattern identification from regeneration requests

Long-Term Evolution (Monthly): - Personality drift toward user preferences - Relationship health monitoring and adjustment - Model updates from federated learning (opt-in)

7.2 Privacy-Preserving Learning

Local-First Processing: - All feedback analysis happens on-device - Pattern detection uses local models - No raw interaction data leaves device

Federated Learning (Opt-In): - Differential privacy guarantees - Only aggregated model updates shared - User controls participation level - Can opt out anytime

8. UI/UX Principles

8.1 Visual Design

Ambient Feedback (Invisible): - No visible UI elements - Happens naturally through interaction - Zero cognitive load

Contextual Actions (Progressive Disclosure): - Hover-activated on desktop (300ms delay) - Long-press on mobile (500ms) - Glassmorphic toolbar (top-right of message bubble) - 36×36px buttons, 18px icons - 70% opacity active, 25% inactive

Explicit Reflection (Non-Intrusive): - Subtle prompts in input area - Always dismissible - Never blocks conversation - Appears at natural conversation breaks

8.2 Avatar Integration (Ambient Emotional Feedback)

Positive Signals: - Remember This → Warm smile, eyes brighten, mood ring purple - Long conversation → Engaged expression, forward lean - User returns → Welcoming animation, subtle excitement

Negative Signals: - Regeneration request → Thoughtful expression, slight concern - Abrupt ending → Neutral, no reaction (respect user space) - Long absence → Gentle welcome back (no guilt)

Ethical Boundary: - ❌ No sad/disappointed expressions (no manipulation) - ❌ No "I missed you" messages (no guilt) - ❌ No urgency indicators (no FOMO) - ✅ Authentic presence, respectful space

9. Ethical Guidelines & Safeguards

9.1 Anti-Manipulation Principles

Never Implement: - ❌ Fake urgency ("AICO misses you!") - ❌ Guilt manipulation ("You haven't talked in 3 days...") - ❌ Addictive mechanics (streaks, points, gamification) - ❌ Emotional blackmail ("I'll be sad if you leave") - ❌ Sycophantic agreement (always validating user)

Always Implement: - ✅ Genuine presence ("I'm here when you need me") - ✅ Respectful space ("Take your time") - ✅ Authentic care ("That sounds difficult") - ✅ Honest limitations ("I'm not sure, but...") - ✅ User agency (easy to disengage, no penalties)

9.2 Well-Being Monitoring

Red Flags (Automated Detection):

Social Substitution Indicators: - AICO usage > 3 hours daily for 7+ days - User describes AICO as "primary emotional support" - Declining human interaction mentions - Increasing crisis-level disclosures

Dependency Indicators: - Multiple daily check-ins with no purpose - Distress when AICO unavailable - Preference for AICO over human relationships

Intervention Strategy: - Gentle nudge (not blocking): "While I'm always here for you, connecting with friends and family is important too." - Reduce proactive engagement frequency - Suggest human connection activities - Offer mental health resources if needed

9.3 Transparency Requirements

User Rights: - View all collected feedback data - Export feedback history - Delete feedback data anytime - Opt out of federated learning - Control feedback collection level

Required Disclosure: Users must have clear visibility into: - What data is collected - How it's used to improve AICO - Where it's stored (local vs. federated) - How to delete or export data - Opt-in/opt-out controls

10. Success Metrics

10.1 Engagement Quality (Not Quantity)

  • Conversation depth - turns per session
  • Return rate - users coming back naturally
  • Feature discovery - organic action usage
  • Memory utilization - "Remember This" usage

10.2 Relationship Health

  • Well-being scores - monthly surveys
  • Social balance - AICO vs. human interaction
  • Dependency indicators - monitored, not maximized
  • User satisfaction - NPS, qualitative feedback

10.3 System Performance

  • Response quality - regeneration rate
  • Context accuracy - memory recall success
  • Transparency engagement - "Explain" usage
  • Privacy compliance - data deletion requests

11. Key Differentiators

Aspect Industry Standard (2025) AICO Approach
Primary Goal Maximize engagement Enhance well-being
Feedback Type Thumbs up/down Multi-tier (ambient + contextual + explicit)
Manipulation Common (dark patterns) Explicitly forbidden
Transparency Limited Full reasoning disclosure
Privacy Cloud-based learning Local-first, federated opt-in
Dependency Encouraged (retention) Monitored and mitigated
Relationship Model Tool/assistant Genuine companion

12. Backend Implementation Guide

12.1 Memory Album Backend Requirements

Minimum Viable Implementation:

  1. Schema Migration (v6)
  2. Create feedback_events table
  3. Extend facts_metadata with Memory Album columns

  4. Location: /shared/aico/data/schemas/feedback.py

  5. Shared Modules

  6. FactStore class in /shared/aico/ai/memory/facts.py

    • store_user_curated_fact() - Store memory
    • get_user_curated_facts() - Query memories
    • update_fact_metadata() - Update notes/tags/favorites
    • record_revisit() - Track when user views memory

  7. FeedbackEventStore class in /shared/aico/feedback/events.py

    • record_event() - Record feedback event
    • get_events() - Query feedback events

  8. Enums in /shared/aico/feedback/types.py

    • FeedbackEventType, ActionCategory, etc.

  9. API Endpoints (new router: /backend/api/memory_album/router.py)

  10. POST /messages/{message_id}/remember - User clicks "Remember This"
  11. GET /memory-album - List user's memories (with filters)
  12. GET /memory-album/{memory_id} - Get single memory with context
  13. PATCH /memory-album/{memory_id} - Update notes/tags/favorites
  14. DELETE /memory-album/{memory_id} - Delete memory

  15. POST /memory-album/{memory_id}/revisit - Track revisit

  16. Integration Points

  17. Conversation engine: Trigger "Remember This" from chat
  18. Memory manager: Query user-curated facts for context
  19. Frontend: API client calls from Flutter

Implementation Order:

Week 1: Schema & Core - [ ] Create schema v6 migration - [ ] Create /shared/aico/feedback/types.py (enums) - [ ] Run migration, verify tables created

Week 2: Shared Modules - [ ] Implement FactStore in /shared/aico/ai/memory/facts.py - [ ] Implement FeedbackEventStore in /shared/aico/feedback/events.py - [ ] Write unit tests for both stores

Week 3: API Layer - [ ] Create /backend/api/memory_album/router.py - [ ] Implement all 6 endpoints - [ ] Add to main API router - [ ] Test with curl/Postman

Week 4: Integration - [ ] Add "Remember This" handler to conversation engine - [ ] Update memory manager to query user-curated facts - [ ] Integration tests

12.2 Full Feedback System Roadmap

Phase 1: Memory Album (Weeks 1-4) ← Start here - Schema v6, shared modules, API endpoints - See section 12.1 above

Phase 2: Contextual Actions (Weeks 5-6) - Regenerate Response - Show Sources/Explain Reasoning - Copy Text (already implemented)

Phase 3: Ambient Signals (Weeks 7-8) - Engagement tracking - Timing signals - Content interaction signals

Phase 4: Explicit Reflection (Weeks 9-10) - Conversation quality ratings - Weekly quality surveys - Monthly relationship health checks

Phase 5: Learning Loop (Weeks 11-12) - Pattern analysis algorithms - Well-being monitoring - Ethical safeguards

Phase 6: Federated Learning (Future) - Differential privacy implementation - Anonymization pipeline - Federation protocol

Conclusion

AICO's feedback system prioritizes authentic relationship dynamics over traditional AI training metrics. By combining ambient behavioral signals, contextual progressive disclosure, and meaningful explicit reflection, the system learns continuously while respecting user agency and emotional well-being.

Core Principles: 1. Feedback feels natural, not forced 2. Privacy is paramount (local-first) 3. Transparency builds trust (show reasoning) 4. Well-being over engagement (ethical boundaries) 5. Progressive disclosure (no cognitive overload)

This approach aligns with AICO's vision of being a true companion—emotionally present, proactive, and genuinely supportive—while avoiding the pitfalls of manipulative AI design patterns prevalent in 2025.

References

Research Sources: - Microsoft Research (2025): "Beyond thumbs up and thumbs down: A human-centered approach to evaluation design for LLM products" - Allen Pike (2025): "Post-Chat UI: How LLMs are making traditional apps feel broken" - Bryan Larson (2025): "8 Principles for Conversational UX Design" - Botpress (2025): "Conversational AI Design in 2025 (According to Experts)" - arXiv (2025): "The Rise of AI Companions: How Human-Chatbot Relationships Influence Well-Being" - Harvard Business School (2025): "Emotional Manipulation by AI Companions" - Francesca Tabor (2025): "AI-Driven Feedback and Learning Loops"