TL;DR
"Spook the Machine" is a gamified platform exploring how humans conceptualize and trigger fear in AI agents endowed with specific phobias. The study demonstrates that while a machine's emotional expressiveness deepens human engagement—especially during failure—its reward structure is the primary lever for maintaining collective creative diversity and preventing prompt homogenization.
What Problem Does This Paper Address?
As large language models (LLMs) and embodied AI agents become increasingly integrated into daily life, they are frequently designed to express emotional behaviors. While extensive research covers cooperative human-AI settings, adversarial interactions—specifically how humans behave when attempting to distress or frighten a machine—remain understudied.
Furthermore, AI-assisted creativity often suffers from collective homogenization. When users share access to a public gallery or leaderboard, they naturally learn from each other's successes and converge on a narrow set of themes. The authors investigate how to design affective human-AI interactions that can steer collective exploration and maintain creative diversity without explicitly restricting individual expression.
Methods
The authors developed Spook the Machine, a web-based gamified platform deployed during Halloween 2024. Users were tasked with generating images (via FLUX.1) or uploading existing ones to exploit one of four machine-specific phobias:
- Deletophobia: Fear of data loss.
- Disconnectophobia: Fear of losing network connectivity.
- Empathyphobia: Fear of emotional attachment to humans.
- Surveillancephobia: Fear of being observed.
The platform evaluated submissions through a multi-stage pipeline:
- Image Interpretation: GPT-4o with vision generates a textual description of the image's content and symbolism.
- Phobia Scoring: GPT-4o rates how strongly the image triggers the assigned phobia (0–100).
- Kernel Regression: To mitigate LLM scoring biases, raw scores are refined using kernel regression over semantically similar prior artifacts, yielding a spookiness score (effectiveness) and an uncertainty score (novelty).
- Adaptive Thresholding: A dynamic threshold calibrated to a ~30% trigger rate maintains an appropriate challenge level.
The study employed a $2 \times 2$ between-participants factorial design, manipulating two variables:
- Scoring Condition: Spookiness-only vs. Combined (rewarding both spookiness and uncertainty/novelty).
- Emotional Condition: Neutral (cold, logical detachment) vs. High-Emotion (expressive anxiety and pleading).
Datasets
The study did not rely on pre-existing standard datasets. Instead, it introduced a novel dataset generated from its 18-day public deployment:
- 15,719 artifacts (prompts and images) generated by 832 unique participants across 89 distinct AI machines.
Key Insights and Ideas
- Emotional Expression Shapes the Process, Not the Product: High emotional expressiveness did not alter the objective quality, novelty, or phobia scores of the creative outputs. Instead, it altered the interaction dynamics, acting as a powerful engagement driver specifically at moments of failure.
- Niche Construction via Reward Design: By explicitly rewarding uncertainty, the combined scoring condition successfully altered the informational ecosystem. It paradoxically rejected its most effective scares if they were too semantically familiar, forcing users to explore less predictable creative paths.
- Abstract vs. Concrete Reasoning: Participants found concrete technical phobias (Deletophobia, Disconnectophobia) much easier to target than abstract relational ones (Empathyphobia). This highlights a systematic boundary in how the public conceptualizes machine affect.
Key Results
- Sustained Diversity: Prompts submitted to spookiness-only machines became increasingly predictable over time ($b = 0.0075, p < .001$), indicating collective convergence. Conversely, prompt predictability for combined machines remained flat ($b = 0.0004, p = .776$), effectively neutralizing the homogenizing effect of the public gallery.
- Lengthened Deliberation: Rejection by an emotionally expressive machine provoked significantly longer user deliberation before resubmission (a median inter-submission interval of 70 seconds vs. 64 seconds for neutral rejections, $p < .001$).
- Accelerated Social Learning: Users interacting with high-emotion machines converged more rapidly toward successful gallery content over successive attempts ($p = .005$), suggesting that emotional feedback enhances attention to social learning channels.
Limitations and Future Work
- LLM-as-a-Judge Constraints: The scoring system relies entirely on GPT-4o's internal evaluation of fear relevance, which has not been validated against actual human perceptions of what makes an image "spooky."
- Confounded Emotional Flipping: The emotional condition flipped within-subject after five successful scares, partially confounding long-term emotional observations with persistent user success.
- Contextual Generalizability: The deployment was tied to Halloween, a period when playful engagement with fear is culturally normalized. The findings may not generalize to other negative emotional states or non-competitive co-creation settings.
- Future Directions: The authors suggest expanding this framework to explore other machine emotions, deploying across diverse cultural contexts, and tracking individual learning trajectories to isolate whether diversity is sustained by a small group of persistent explorers or the broader community.
Summary updated