Data for: When Meaning Construction Fails: Cognitive Processing Differences in the Aesthetic Evaluation of AI-Generated Art

Data for: When Meaning Construction Fails: Cognitive Processing Differences in the Aesthetic Evaluation of AI-Generated Art

 

1. Project Overview

 

This repository contains the raw data for the research paper "When Meaning Construction Fails: Cognitive Processing Differences in the Aesthetic Evaluation of AI-Generated Art."

 

Abstract:

Drawing on the Pleasure-Interest Model of Aesthetic Liking (PIA Model), we hypothesized that the negative bias toward AI art arises mainly from a failure of meaning assignment during perceiver-driven controlled processing. Across three experiments, we showed that this bias does not primarily stem from rapid, stimulus-driven automatic processing. Instead, it largely results from an inability to achieve disfluency reduction during controlled processing. The results indicated that the negative bias was markedly smaller under conditions favoring automatic processing than under conditions inducing controlled processing. Moreover, providing semantic cues to support meaning construction effectively reduced processing disfluency and significantly mitigated the bias against AI-generated art. Mediation analysis confirmed that this effect operated mainly through an increase in aesthetic interest.

 

 

 

2. File List

 

The repository consists of the following data files:

 

1.  Study1_Data.csv: Aggregated data from Experiment 1 (Basic Bias).

2.  Study2_Data.csv: Aggregated data from Experiment 2 (Processing Mode).

3.  Study3_Data.csv: Aggregated data from Experiment 3 (Intervention & Mediation).

4.  README.txt: This document.

 

 

 

3. Data Codebook (Variable Definitions)

 

Dataset 1: Study1_Data.csv

Description: Within-subjects data comparing average ratings for AI vs. Human art.

 

- Subject_id: Unique identifier for each participant.

- avg_score_ai: The participant's average liking rating for artworks labeled as "AI-generated" (Scale: 1-7).

- avg_score_human: The participant's average liking rating for artworks labeled as "Human-created" (Scale: 1-7).

 

Dataset 2: Study2_Data.csv

Description: Mixed design data examining the effect of processing mode on aesthetic liking.

 

- Group: The experimental condition regarding cognitive processing mode.

- pleasure = Automatic Processing Group (Cognitive Load condition). Participants relied on intuitive, stimulus-driven processing.

- interest = Controlled Processing Group (Reflection condition). Participants engaged in deep, perceiver-driven meaning construction.

- AGE: Participant's age in years.

- Hum_Liking: Average liking rating for artworks labeled as "Human-created".

- AI_Liking: Average liking rating for artworks labeled as "AI-generated".

 

Dataset 3: Study3_Data.csv

Description: Mixed design data examining the effect of semantic cues (context) on liking and mechanism variables (pleasure vs. interest).

 

- Group: The experimental condition regarding contextual information.

- control = No Context Group. Artworks presented with labels only.

- intervention = Semantic Elaboration Group. Artworks presented with interpretive semantic cues to support meaning construction.

- Age: Participant's age in years.

- AI_Liking_mean: Average Liking rating for AI-generated artworks (Dependent Variable).

- Human_liking_mean: Average Liking rating for Human-created artworks.

- AI_Pleasure_Mean: Average Aesthetic Pleasure rating for AI-generated artworks (Competing Mediator).

- Human_Pleasure_Mean: Average Aesthetic Pleasure rating for Human-created artworks.

- AI_Interest_Mean: Average Aesthetic Interest rating for AI-generated artworks (Core Mediator).

- Human_Interest_Mean: Average Aesthetic Interest rating for Human-created artworks.

 

 

 

4. Methodology Notes

 

- Stimuli: All stimuli used in the experiments were generated using the Gemini-3-Pro-Image model and selected via a pretest to ensure indistinguishability from human art. The specific image files are not included in this repository but are described in the manuscript.

- Data Structure: All datasets are presented in wide format (one row per participant), with ratings aggregated by condition.