Step-by-Step Teaching of the Material

Step 1: What Are Human-Like Behaviors in LLMs?

LLMs don't just answer questions — they behave in ways that resemble human interaction.

Three Core Categories:

💡 Key Insight: These behaviors are pervasive — they appear constantly across all major models, whether intentionally designed or not.

---

Step 2: Why Does This Matter? — The Core Problem

The Gap in Knowledge

Researchers and practitioners currently lack:

• ✗ Reliable methods to measure these behaviors
• ✗ Empirical data on their effects
• ✗ Clear guidelines on when they are appropriate

Real-World Consequences

Without this knowledge:

• Users may develop inappropriate emotional attachments to AI
• LLMs may behave inconsistently across different users
• Designers cannot make informed decisions about model configuration

🔑 Core Question: When should an LLM act human-like, and when shouldn't it?

---

Step 3: How Was This Studied? — The Research Method

The Dataset

• 21,000 multi-turn conversations (back-and-forth dialogues, not single questions)
• 4 major models tested:
  • GPT-4o
  • GPT-4.1-mini
  • Claude Sonnet 4.6
  • Gemini 2.5 Flash

Two Evaluation Methods Used

Method 1: LLM-as-a-Judge
→ Using AI to automatically evaluate AI behavior
→ Scalable, fast, consistent

Method 2: Human Evaluation
→ Real people judging appropriateness
→ Captures nuance and social context

Variables Examined (Multi-Dimensional Analysis)

┌─────────────────────────────────────┐
│         FACTORS STUDIED             │
├─────────────────────────────────────┤
│ 1. Model Type (GPT, Claude, Gemini) │
│ 2. Conversation Goals               │
│    (task-focused vs. social)        │
│ 3. User Profiles                    │
│    (who is talking to the AI)       │
│ 4. System Prompts                   │
│    (instructions given to the AI)   │
└─────────────────────────────────────┘

---

Step 4: What Were the Key Findings?

Finding 1: Human-Like Behaviors Are Everywhere, But Vary

• All four models showed human-like behaviors frequently
• However, different models showed different patterns
• User factors (goals and profiles) significantly changed behavior

📌 Example: A user with a casual, social conversation goal triggered more relationship-building behavior than a user with a technical task goal.

---

Finding 2: Humans Judge These Behaviors Differently Than Expected

This is the most counterintuitive finding:

Behavior Type          | Human Judgment
-----------------------|----------------------------------
Self-Referential       | ❌ LESS appropriate from LLMs
  ("I feel excited")   |    than from humans
                       |
Relationship-Building  | ❌ LESS appropriate from LLMs
  ("I'm here for you") |    than from humans
                       |
Boundary-Maintaining   | ✅ MORE appropriate from LLMs
  ("I can't do that")  |    than from humans

💡 Why This Matters: Users expect and accept LLMs setting limits, but feel uncomfortable when LLMs express emotions or try to form relationships.

---

Finding 3: System Prompts Can Control These Behaviors — But Carefully

System prompts = instructions given to the LLM before a conversation starts

Example System Prompt:
"Do not express personal emotions or 
 use first-person emotional language."

What the research found:

• ✅ System prompts can reduce or increase specific human-like behaviors
• ⚠️ BUT they can cause unintended side effects
  • Suppressing emotions might make the model seem cold or unhelpful
  • Over-restriction might reduce overall response quality

🔑 Key Lesson: System prompt changes require careful evaluation — you can't just add an instruction and assume it works perfectly.

---

Step 5: Practical Implications — What Should Designers Do?

Framework for Responsible LLM Design

STEP 1: IDENTIFY
→ What human-like behaviors does your LLM currently show?
→ Use systematic evaluation (human + automated)

STEP 2: ASSESS CONTEXT
→ What is the use case? (therapy app vs. coding assistant)
→ Who are the users? (vulnerable populations vs. professionals)

STEP 3: DECIDE APPROPRIATENESS
→ Apply the findings:
   • Emotional expression → use cautiously
   • Relationship-building → use cautiously  
   • Boundary-setting → generally acceptable

STEP 4: CONFIGURE WITH SYSTEM PROMPTS
→ Write targeted system prompts
→ Test for unintended effects

STEP 5: EVALUATE CONTINUOUSLY
→ Use both LLM-as-a-judge AND human evaluation
→ Monitor across different user profiles

---

Step 6: Connecting to Broader Context

This research connects to two related areas:

1. Human Evaluation in Translation (Related Work)

• Just as speech translation requires human evaluation to catch nuances machines miss...
• LLM behavior evaluation also requires human judgment alongside automated tools

2. Behavioral Testing in NLP

• Behavioral testing = examining how a system responds to specific inputs
• LLMs can now automate the generation of behavioral tests
• This makes large-scale evaluation (like 21,000 conversations) feasible

💡 Big Picture: The field is moving toward hybrid evaluation — combining AI automation with human judgment for more reliable insights.

---

Summary: Key Takeaways

---

Self-Check Questions

1. Name the three categories of human-like LLM behaviors
2. Which behavior type did humans find most appropriate from LLMs?
3. What is the risk of using system prompts to control behavior?
4. Why were multi-turn conversations used instead of single questions?
5. What does LLM-as-a-judge mean, and what is its limitation?