> Feynman's favorite quip was: "If I think you're serious, I'll give you an answer. But I see you're here for a laugh." Today's paper seriously studies the funny — and finds that AI's sense of humor is structurally completely different from humans'.
---
Introduction: Does AI Have a Humor Style?
More and more people are using AI as a "human simulator" — having AI fill out psychological questionnaires and then analyzing its "personality traits." Psychologists want to know whether GPT is extroverted; sociologists want to know whether Claude has empathy.
This trend raises a deep question: Is the internal structure behind AI's questionnaire answers the same as the internal structure behind humans' answers?
If not, then all conclusions from psychological research that substitutes AI for humans are in doubt.
A paper from EMNLP 2025 answers this using a classic psychological instrument — the Humor Style Questionnaire (HSQ).
The answer is simple and brutal: No. And the gap is huge.
---
Chapter 2: The Four Faces of Humor
Some background. Psychologists divide humor into four styles:
- Affiliative humor: telling jokes to make people happy and strengthen social bonds
- Self-enhancing humor: viewing life humorously, staying optimistic in adversity
- Aggressive humor: sarcasm, mockery, ridiculing others
- Self-defeating humor: putting yourself down, self-deprecation to please others
- Title: Fingerprinting LLMs through Survey Item Factor Correlation: A Case Study on Humor Style Questionnaire
- Author: Simon Münker
- Venue: EMNLP 2025 (Main)
- Link: ACL Anthology
- Method: HSQ questionnaire × 6 LLMs, factor analysis, exploratory graph analysis
The first two are "healthy" humor; the last two are "unhealthy." In human respondents, these four styles show a stable factor structure — the healthy pair correlates positively, the unhealthy pair correlates positively, and the correlation between the two groups is weak or absent.
The researcher had 6 mainstream LLMs fill out the same HSQ and analyzed their factor correlation structure — how the four styles interrelate.
Humans' factor structure is stable and predictable. AI's factor structure — nothing like humans'.
---
Chapter 3: AI's Way of Being Funny Is Incomprehensible to Humans
Specific findings:
1. No AI successfully recovered the HSQ's four-factor structure. Zero. All six models failed.
2. Human subgroups are highly consistent internally — regardless of gender, age, or cultural background, humans' factor correlation structure is highly stable. But AI models showed near-zero similarity to humans.
3. Exploratory graph analysis (EGA) confirmed: no LLM could recover the HSQ's four theoretical constructs from its questionnaire answers.
What does this mean? When you ask AI to "pretend to be a humorous person" and fill out the questionnaire, the answers carry no human-like "psychological structure." They may be literally "correct" ("I like making people laugh"), but the internal correlation patterns among the answers are completely different from humans.
AI's humor is not a "personality" — it's a "statistical language game."
---
Chapter 4: Why?
The paper offers several possible explanations:
1. LLMs have no "self": humans answer questionnaires through introspection about their own behavioral patterns. AI "answers" based on statistical patterns of human questionnaire responses in training data. The former has a centralized "self-narrative"; the latter only distributed conditional probabilities.
2. Perspective bias in training data: web text describing humor may lean toward "humor expression in standard social situations" rather than private, introspective humor experiences. Humans' self-enhancing humor (an internal optimistic outlook) is rarely reflected in public text.
3. Factor structures require "experience integration": humans' four humor styles form a stable factor structure because these styles are integrated into a coherent self-concept through repeated real-life experience. LLMs have no life, hence no "integration."
---
Chapter 5: How Important Is This Finding?
It punctures a rapidly spreading illusion: that using LLMs as "human simulators" is viable.
A large volume of research now substitutes LLMs for human respondents — "GPT-4's Big Five scores," "Claude's performance on values questionnaires." But if LLMs' factor structure itself differs entirely from humans', those "scores" are meaningless — like saying someone has "an extroversion score of 85" when the scoring system itself doesn't apply to that entity.
This doesn't mean AI can't simulate human behavior — in narrow, behavior-cloning-based tasks, AI can be an excellent simulator. But in domains requiring internal psychological structural consistency (personality assessment, values measurement, clinical diagnosis), AI's "simulation" is currently just a skin of language — with no corresponding psychological "skeleton" inside.
---
*Paper info*