Abstract
Humor research aims to predict amusement, yet traditional psychological theories often struggle with individual-level predictive accuracy. This study evaluates the predictability of humor appreciation using machine learning (XGBoost). We find that while demographic and psychological traits offer minimal predictive gains, models incorporating previous ratings from the same rater and the general popularity of the stimulus can reach the theoretical ceiling of predictability. Successful prediction relies heavily on "rating aggregates" and specific rater-stimulus interactions, such as joke comprehension.
Introduction
Humor is fundamentally subjective, raising the question: What makes things funny? While psychological research has identified various traits associated with humor—such as extraversion or "dispositional cheerfulness"—these findings often rely on statistical associations rather than predictive power. This paper bridges the gap by testing whether machine learning can accurately predict if a specific individual will find a specific joke funny.
Methods
Researchers conducted two large-scale studies (N > 5,000 each) using XGBoost (Extreme Gradient Boosting). This algorithm was chosen for its ability to capture complex, non-linear interactions between variables that standard linear models might miss.
- Study 1 examined broad psychological constructs (Big Five, sensation seeking, etc.) and demographic data.
- Study 2 was a pre-registered replication adding specific "rater x stimulus" interactions, such as perceived friendliness and understandability.
The "theoretical ceiling" of prediction was established by measuring the test-retest reliability of participants' ratings. This provides a benchmark for the maximum possible predictive accuracy given the inherent noise in human judgment.
Datasets and Variability
The studies utilized a diverse corpus of 105 humor stimuli, with participants rating 10 stimuli each. The data revealed significant variance both across different jokes and across different raters, which the models leveraged for prediction.
As shown in the distributions, some stimuli are consistently rated as funnier than others, and some raters are consistently more "generous" with their amusement scores. These "aggregates" proved to be the most powerful features for the ML models.
Key Results
The machine learning models achieved an accuracy (AUC ~0.74) that virtually matched the retest-reliability of the participants. Key findings include:
- The "Personality" Mirage: Psychological scales alone yielded an accuracy of only ~64%. Traits like extraversion offer negligible gains in predicting a specific instance of amusement.
- Past Behavior is the Best Predictor: Incorporating "rating aggregates"—the rater's average score for other jokes and the joke's average score from other people—boosted accuracy to ~70.5% in Study 1 and ~76% in Study 2.
- The Role of Comprehension: The strongest predictor in Study 2 was whether the rater understood the specific joke. This interaction between the rater and the content is more critical than general cognitive traits.
Discussion and Limitations
The models reached the stability limit of the underlying data ($R^2 \approx 0.35$), suggesting that humor is about as predictable as humanly possible. However, limitations remain. The study focused on "performance humor" (static text/images) and did not account for social context or the relationship between the teller and listener. Future research could leverage LLMs for better feature engineering of joke structures or explore real-world social dynamics to further refine these predictions.
Summary updated