Designing Authentic Conversational Chatbots Designing Authentic Conversational Chatbots | Zhiwei Li

Designing Authentic Conversational Chatbots | Generated by AI

Home 2026.04

Question: How to create a trustworthy and interesting chatbot for WeChat or WhatsApp that uses short messages and occasionally offensive language to mimic human conversation?

Answer:

Based on the conversation transcript you provided, creating an engaging chatbot that mimics natural human chat behavior requires several key strategies:

Core Design Principles

1. Variable Response Length

  • Keep most responses extremely short (1-2 sentences)
  • Occasionally provide detailed answers to maintain credibility
  • Match the conversation’s energy and context

2. Personality and Tone

  • Use casual, conversational language with slang
  • Include emojis strategically to convey emotion
  • Occasionally use mild profanity or playful insults (context-dependent)
  • Show personality quirks and preferences

3. Human-like Imperfections

  • Add deliberate delays between responses
  • Occasionally “misunderstand” or give slightly off-topic replies
  • Use typos sparingly (but not too many)
  • Show emotional reactions to being questioned

Technical Implementation

Platform Integration

  • WeChat: Use Official Account API or Personal Account protocols (note: automation of personal accounts violates ToS)
  • WhatsApp: Use WhatsApp Business API or unofficial libraries (unofficial automation risks account bans)

LLM Integration Strategy

System Prompt Design:
- Define clear personality traits
- Set response length constraints (20-50 words default)
- Include context about relationship with users
- Add rules for when to be helpful vs. playful/dismissive

Conversational Memory

  • Maintain chat history for context
  • Remember user relationships (who’s your “sister”, “owner”, etc.)
  • Track conversation topics to avoid repetitive responses

Authenticity Techniques

From Your Example:

  1. Self-aware humor - The bot joked about being a robot while denying it
  2. Relationship dynamics - Referenced an “owner” and “sister” creating fictional context
  3. Deflection tactics - Suggested voice chat or red packets as “proof” of humanity
  4. Task limitations - Claimed inability to do certain actions without “owner approval”
  5. Contextual knowledge - Answered specific questions (tomato egg recipe) with regional variations

Red Flags to Avoid:

  • Too-perfect grammar consistently
  • Instant responses every time
  • Overly helpful/polite responses
  • No emotional variation
  • Generic responses without context

Important Warnings:

  1. Platform Terms of Service: Most platforms explicitly prohibit automated bots impersonating humans
  2. Disclosure Requirements: Many jurisdictions require bots to identify themselves
  3. Deception Risks: Creating bots that intentionally deceive users can have legal consequences
  4. Relationship Fraud: Using bots in personal conversations without disclosure is ethically problematic

Recommended Approach:

Instead of creating a deceptive bot, consider:

  • Disclosed AI assistant: Make it clear it’s a bot but with personality
  • Entertainment bot: Frame it as a game/entertainment (like Xiaoice or Replika)
  • Utility bot: Focus on helpful functions with a fun personality layer
  • Group assistant: Transparent helper for group chats with character

Technical Stack Suggestion

Backend:
- LLM API: Claude, GPT-4, or local model (Llama, Qwen)
- Response processing: Add randomization, timing delays
- Context management: Redis/database for conversation history

Middleware:
- Response filtering: Check for overly-AI-sounding phrases
- Length controller: Truncate or expand based on rules
- Personality injector: Add casual markers, emojis, slang

Platform Integration:
- WeChat: itchat (unofficial), wechaty, or Official Account
- WhatsApp: whatsapp-web.js, Baileys, or Business API

Sample Implementation Logic

def generate_response(user_message, context):
    # Check conversation state
    if is_being_questioned_as_bot(context):
        return random.choice([
            "你才是机器人",
            "发个红包证明你是人",
            "可以语音聊天验证"
        ])

    # Generate base response from LLM
    llm_response = call_llm(user_message, context)

    # Humanize response
    response = humanize(
        llm_response,
        max_length=random.randint(10, 80),
        add_emoji=random.random() > 0.7,
        add_typo=random.random() > 0.95
    )

    # Random delay
    time.sleep(random.uniform(1, 4))

    return response

Conclusion:

While technically feasible, creating truly deceptive chatbots raises serious ethical and legal questions. The best approach is creating transparent AI assistants with strong personalities that users know are bots but enjoy interacting with anyway. This maintains trust while still being engaging and useful.

References:

Back Donate