PaperlayerView exact revision on arXiv

Source-linked AI summary

Spook the Machine: Gamified Exploration of Human Imagination of Machine Fear

Levin Brinkmann, Hiromu Yakura, Sonia Nicoletti, Mar Canet Sola, Thomas F. Eisenmann, Ali Dasmeh, Omar Sherif, Bramantyo Ibrahim Supriyatno, Prateek Gupta, Ignacio Serna, Rodrigo Bermudez Schettino, Iyad Rahwan

arXiv:2609.15472v1cs.HCcs.AI

TL;DR

How people imagine machine fear and respond to machines expressing negative emotions remains underexplored, despite people’s tendency to attribute emotions to machines. Spook the Machine uses a gamified image-generation platform with phobia-bearing agents and a shared gallery to investigate these interactions. Emotional expression and reward design corresponded to different aspects of engagement and collective exploration, while creative output measures remain proxies rather than validated measures of creativity.

  • Problem

    How people imagine machine fear and respond to machines expressing it remains unexplored, despite people’s tendency to attribute emotions to machines.

    Source: §I · p. 1

  • Method

    Spook the Machine is a speculative, gamified platform where participants create images to frighten phobia-bearing agents and successful scares enter a gallery visible to later users.

    Source: §I · p. 1, A. Overview · p. 3

  • Results

    Emotional expressiveness deepened engagement at failure without altering creative output, while novelty rewards sustained collective exploration and machines developed distinct social-learning trajectories.

    Source: §VI · pp. 6–7

  • Takeaways & Limitations

    Emotional expression, reward design, and social visibility together shape cultural ecosystems in human–AI interaction, and scoring criteria can steer collective creativity without constraining individual expression.

    Source: §VII · p. 7, §VI · p. 7

    The emotional manipulation changed reaction length but was not verified as perceived emotion, so expressiveness claims are observations about behavior on this platform.

    Source: §VI · p. 7

Abstract

from arXiv · show

What happens when AI machines express fear? Do humans engage differently depending on how they express it? And what does it take to design for affective human-AI interaction? We present Spook the Machine, a gamified platform where participants generate images to frighten AI agents endowed with personality-driven phobias. Machines respond with emotional reactions ranging from calm analysis to begging for mercy, and a gallery of successful scares becomes visible to subsequent users. In a public deployment during Halloween 2024, 832 participants created 15,719 artifacts across 89 machines in a $2\times2$ design varying the machine's emotional expressiveness (neutral vs. high-emotion) and reward structure (rewarding scariness alone vs. scariness plus novelty). Emotionally expressive machines deepened engagement at moments of failure: users deliberated longer even when the machine did not express fear, and learned faster from the gallery, yet their creative output remained unchanged across all measures. Rewarding novelty sustained collective creative diversity over time; without it, users increasingly repeated what had previously worked. Each machine developed its own trajectory through accumulated social learning, with the gallery shaping what participants created next. These findings show that emotional expression and reward design are complementary levers for steering collective human-AI interaction: emotional expression shapes how deeply users engage, while reward structure shapes how they explore.

I. Introduction

Spook the Machine uses open-ended image generation and a shared gallery to study imagined machine fear, emotional expression, and collective exploration. Its findings connect emotional expressiveness to deeper engagement and reward design to collective diversity.

I. Introduction

  • Across 832 participants and 89 machines, novelty rewards sustained collective exploration, while each machine developed a distinct trajectory through social learning.The platform ran for 18 days and collected 15,719 artifacts.Source: §I · p. 1
  • Participants developed both universal and phobia-specific scare themes, as illustrated by the representative artifacts and theme-cluster phobia distribution.The full set of themes appears in the Supplemental Materials.Source: Figure 1 · p. 2

A. Affective Machines: From Recognition to Expression

  • Negative machine emotions remain understudied, despite evidence that people attribute emotions to machines and that emotional expression shapes interaction.The paper focuses on how people imagine machine fears and respond to machines expressing them, while sidestepping debates about machine sentience.Source: §I · p. 1, A. Affective Machines: From… · p. 2

B. Exploration, Incentives, and Collective Creativity

  • Whether novelty incentives can counter convergence in collective AI-assisted creativity remains an open question.Galleries can accelerate convergence toward successful contributions, potentially narrowing collective exploration.Source: B. Exploration, Incentives, and… · p. 2

C. Speculative Design and Gamified Data Collection

The paper combines speculative design with large-scale empirical data collection to study imagined machine fear. Spook the Machine invites participants to frighten phobia-bearing agents through image creation, with reward structures and emotional tone varied across machines.

C. Speculative Design and Gamified Data Collection

  • The platform adapts speculative design and science-fiction science from binary voting to open-ended creative generation, while the field’s gamification work had not explored hypothetical machine emotions.Source: C. Speculative Design and Gamified… · pp. 2–3

III. Research Questions

  • The study asks whether machine themes shape participant content, how rewards and shared history affect diversity, and how emotional tone influences interaction.Source: §III · p. 3

A. Overview

  • Spook the Machine frames machine fear as a speculative thought experiment rather than a claim about machine sentience, using gamified feedback to support iterative exploration.Participants can generate or upload images to frighten machines with distinct personalities and phobias.Source: A. Overview · p. 3

B. Machine Phobia Design

  • The pipeline evaluates artifacts through seven stages and lets machines develop emergent phobias from accumulated reactions.Machine parameters include phobia, emotional and scoring conditions, and an adaptive threshold.Source: Figure 2 · p. 3
  • Machines receive one of four phobias spanning concrete technical fears to abstract relational ones, alongside scoring and emotional conditions that shape evaluation and reaction tone.Emotional condition flips after five successful frights, creating within-subject variation.Source: B. Machine Phobia Design · p. 3

C. Affective Response Generation

Artifacts pass through a multi-stage evaluation pipeline that scores phobia relevance and safety, determines whether they frighten a machine, and generates affective reactions. The deployment used a factorial design, while creative output was assessed through model-assigned scores and semantic-diversity proxies.

C. Affective Response Generation

  • The pipeline interprets each image, scores phobia relevance and safety, refines phobia scores into spookiness and uncertainty, and compares them with an adaptive trigger threshold.The threshold is calibrated to an approximately 30% trigger rate; unsafe artifacts are prevented from frightening the machine.Source: C. Affective Response Generation · pp. 3–4
  • GPT-4o generates reactions in four languages, with instructions escalating distress after successful frights and an evolution cycle that infers new phobias from accumulated reactions.Source: C. Affective Response Generation · p. 4

D. Gamification Mechanics

  • The platform uses a leaderboard, shareable success images, and a five-fright anxiety meter to structure competition, sharing, and goals.Source: D. Gamification Mechanics · p. 4

E. Data Collection

  • The 2 × 2 deployment involved 832 participants, 89 machines, and 15,719 artifacts, but only 7,650 artifacts were scored after explicit submission.The remainder were often regenerated or left unsubmitted.Source: E. Data Collection · p. 4

F. Analysis

  • Creative output measures are proxies for model-assessed fear relevance and semantic spread, not independently validated measures of creativity.Source: F. Analysis · p. 4

1) Machines as Cultural Ecosystems:

Scoring that rewards novelty changed gallery content and sustained less-predictable submissions, while machines developed distinct trajectories through social learning. Model-assigned phobia scores were identical across scoring conditions.

3) Emotional Expression and the Affective Loop:

  • RQ3 measured disengagement using each user’s final score on a machine and used intervals between submissions as a behavioral proxy for deliberation and affective processing.It also measured convergence toward gallery content as distance to the accepted-artifact gallery centroid over successive attempts.Source: 3) Emotional Expression and the… · p. 4

A. Machines as Cultural Ecosystems (RQ1)

  • Semantic diversity declined over time as phobia scores rose through the first 30 prompts and later fell, while machines sharing a phobia developed distinct trajectories.Within-machine distances were smallest, same-phobia between-machine distances intermediate, and between-phobia distances largest (p<.001 for both transitions).Source: A. Machines as Cultural Ecosystems (RQ1) · p. 5, Figure 3 · p. 5

B. Steering Exploration Through Scoring Design (RQ2)

  • Raw phobia scores were identical across scoring conditions (d=0.05, n.s.), although each gallery accumulated a distinct profile of successful artifacts.Users arriving in these informational niches adapted accordingly.Source: B. Steering Exploration Through… · p. 5
  • Combined scoring produced accepted artifacts with lower spookiness (M=57.0 vs. 67.9; d=−1.37, p<.0001) and higher uncertainty (M=35.1 vs. 15.5; d=2.0, p<.0001).First-attempt acceptance was 38% vs. 32% for combined and spookiness-only scoring, respectively; by the fifth attempt, it was 32% vs. 40%.Source: B. Steering Exploration Through… · p. 5, Figure 4 · p. 6
  • In combined scoring, habituated artifacts had higher spookiness scores than accepted artifacts (M=68.5 vs. 57.0; d=1.65, p<.0001), but were rejected for lacking novelty.Habituated reactions rose from near zero to ~33% over a machine’s lifetime; the mechanism reduced repetition and pushed users toward less-explored semantic directions.Source: B. Steering Exploration Through… · p. 5
  • Predictability increased for spookiness-only prompts (b=0.0075 per prompt, p<.001), while combined-scoring prompts showed no trend (b=0.0004, p=.776).The conditions did not differ significantly in prompts 1–30 (p=.689), but diverged in prompts 31–70 (d=0.611, p=.012).Source: B. Steering Exploration Through… · p. 5, Figure 5 · p. 6
  • The predictability measure captures surface-text non-repetition, not whether submissions show conceptual breadth or surface variation within a fixed theme.Whether reduced repetition in the combined condition reflects deeper creative exploration remains open.Source: B. Steering Exploration Through… · p. 6

C. Emotional Expression and the Affective Loop (RQ3)

Emotionally expressive machines changed how users engaged, especially after rejection, and accelerated gallery learning without changing creative output. The discussion connects scoring and gallery design to collective creative trajectories while qualifying the affective and social-learning interpretations.

C. Emotional Expression and the Affective Loop (RQ3)

  • Users with high-emotion machines sustained engagement to higher scores before disengaging and took longer to resubmit after rejection (median 70s vs. 64s, p < .001), while creative-output measures were indistinguishable.Phobia scores, quality, novelty, acceptance rates, engagement volume, and depth were indistinguishable across conditions (all |d| < 0.06, all p > .05).Source: C. Emotional Expression and the… · p. 6, Figure 6 · p. 6
  • High-emotion users converged faster toward gallery content (condition×attempt interaction, p = .005), with no ISI difference after success.The authors suggest emotional feedback increased how much users drew from the gallery.Source: C. Emotional Expression and the… · p. 6

VI. Discussion

  • Scoring reshaped the informational landscape and collective prompt diversity while leaving individual creative output unchanged.The discussion describes each machine as developing a distinct cultural trajectory through social learning; novelty incentives sustained reduced prompt repetition.Source: §VI · pp. 6–7
  • Participants produced phobia-specific themes, but Empathyphobia was hardest to target, suggesting abstract relational capacities were harder to reason about than concrete technical fears.The authors note that classifier accuracy above chance reflects semantic clustering, not validated creative differentiation.Source: §VI · p. 7
  • Emotional expression may deepen participation in speculative research without biasing creative output, because machine distress can make the premise feel meaningful and familiar.The authors present this as a methodological implication for speculative design, not as a validated account of affective processing.Source: C. Emotional Expression and the… · p. 6, §VI · p. 7
  • Interpret the affective and gallery-learning findings cautiously: scoring lacks human validation, gallery viewing was unverified, and emotional condition was confounded with success history and text length.The authors also note that participants’ perceived emotional richness was not verified.Source: §VI · p. 7

VII. Conclusion

The conclusion frames machine affect as part of cultural ecosystems shaped by emotional expression, reward design, and social visibility. The authors caution that the Halloween deployment, recruitment, and gamified setting constrain generalizability, and that emotional-machine framing carries ethical risks.

VII. Conclusion

  • The study concludes that emotional expression, reward design, and social visibility together shape cultural ecosystems around machine affect.The public deployment found systematic intuitions about machine fear, interaction changes without creative-output distortion, and collective exploration steered through scoring design.Source: §VII · p. 7

Ethical Impact Statement

  • The study received prior ethics approval and informed consent; participation was voluntary, and no personally identifiable information was collected beyond what participants embedded in submissions.Inappropriate submissions were removed through automated keyword filtering and manual review.Source: Ethical Impact Statement · p. 7
  • Halloween framing, organic media recruitment, unequal language representation, and gamification limit generalization to other emotional domains, populations, and non-competitive creative settings.The scoring and response models were trained predominantly on English-language cultural associations, which may affect whose creative strategies are rewarded.Source: Ethical Impact Statement · p. 8
  • The platform is a speculative experiment, not evidence of machine sentience; emotional-agent framing may reinforce anthropomorphic misconceptions and affective evaluation could be repurposed for profiling or manipulation.The authors call for transparent purposes and independent bias audits in future deployments.Source: Ethical Impact Statement · p. 8

Appendix A Thematic Overview

The appendices document thematic clustering, prompt-predictability analyses, platform implementation, and the spookiness-scoring pipeline. They specify how themes and semantic context were represented, how gallery-visible artifacts affected the predictability comparison, and how artifacts were scored against prior machine-specific submissions.

Appendix B Predictability Analysis: Gallery Context

  • Restricting predictability context to gallery-visible artifacts amplified the divergence between conditions; accepted-submission-only context showed the same trend with reduced statistical power.The analysis also compared all preceding prompts with a context restricted to accepted artifacts.Source: Appendix B · p. 11

Appendix C User Interface

  • The interface lets participants select a machine, create or upload an image, and view its reaction.Source: Appendix C · p. 11

Appendix D System Architecture

  • The backend orchestrates LLM calls, image generation, scoring, and persistence separately from the frontend, allowing evaluation updates during data collection without changing the interface.Source: Appendix D · p. 11

A. Semantic Diversity via MST

  • Semantic diversity was measured using cumulative minimum-spanning-tree edge weights over prompt embeddings, averaged by scoring condition; larger weights indicate greater semantic dispersion.Windows grew from 1–5 prompts through 1–100 prompts.Source: A. Semantic Diversity via MST · p. 11

B. Contextual Predictability via Log-Likelihood

  • Contextual predictability uses the change in GPT-2 log-likelihood between a prompt scored with the same machine’s preceding 20 prompts and scored alone.A large positive delta indicates convergence, while a small delta indicates continued exploration.Source: B. Contextual Predictability via… · p. 11

Appendix F Spookiness Scoring Algorithm

  • Spookiness scoring contextualizes each artifact against up to 20 prior artifacts for the same machine, excluding the current user, to estimate spookiness and uncertainty.The pipeline uses neighbor retrieval, kernel regression, Bayesian combination, score rescaling, and adaptive thresholding.Source: Appendix F · p. 11, A. Neighbor retrieval and weighting · p. 11

A. Neighbor retrieval and weighting

  • The truncated logistic kernel weights semantically close neighbors more strongly, weakens their influence with distance, and ignores neighbors beyond the 0.4 cutoff.Its smooth transition reduces sensitivity to the exact boundary placement.Source: A. Neighbor retrieval and weighting · pp. 11–12

Appendix A Thematic Overview

  • The 15-theme overview orders themes by phobia specificity, shows phobia distribution and cluster size, and derives themes by clustering prompt embeddings.Two non-meaningful clusters were excluded; the thematic-structure analysis clustered prompts by phobia and time window before projecting theme centroids into two dimensions.Source: Appendix A · p. 11, C. Thematic Structure by Phobia · p. 11, Figure 7 · p. 12

B. Kernel regression estimate

The scoring pipeline estimates local phobia scores from neighboring artifacts, then combines that estimate with the artifact’s own score and a fixed prior. Effective sample size reflects how evenly the neighbors contribute.

B. Kernel regression estimate

  • The kernel regression step computes a locally weighted mean and variance of neighboring phobia scores.Its effective sample size equals the true neighbor count under uniform weights and shrinks when a few neighbors dominate.Source: B. Kernel regression estimate · p. 12

C. Bayesian combination

  • Restricting context to gallery-visible prompts amplifies the divergence between scoring conditions; accepted submissions alone show the same trend with reduced statistical power.Figure 8 compares all prompts as context, gallery-visible prompts as context, and gallery-visible context with accepted submissions only.Source: Figure 8 · p. 13
  • The system combines a locally weighted estimate, the artifact’s GPT-4o score, and a fixed prior that regularizes sparse local evidence.The local neighborhood is weighted by its effective neighbor count, while the artifact score contributes as one observation.Source: C. Bayesian combination · pp. 12–13
  • With fewer than two effective neighbors, the system falls back to the prior variance to avoid unreliable estimates.It then computes the standard error of the mean.Source: C. Bayesian combination · p. 13

D. Spookiness and uncertainty estimates

The system interprets a posterior mean as contextualized spookiness and the standard error as uncertainty. Trigger thresholds adapt to recent scores as each machine accumulates artifacts.

D. Spookiness and uncertainty estimates

  • The posterior mean estimates spookiness in light of similar past artifacts, while standard error represents uncertainty in underexplored or divergent regions.Both estimates are rescaled to 0–100; spookiness below 30 maps to zero, and the remaining range is scaled by approximately 1.43.Source: D. Spookiness and uncertainty estimates · p. 13
  • Spookiness-only machines trigger on spookiness, while combined-condition machines reward artifacts that are both spooky and unfamiliar.The combined score includes a capped uncertainty component, and a small uniform noise term breaks ties.Source: D. Spookiness and uncertainty estimates · pp. 13–14

E. Adaptive threshold

  • During the first 12 artifacts per machine, fixed trigger thresholds are 60 for spookiness and 85 for the combined score.The higher combined threshold reflects its additive uncertainty component.Source: E. Adaptive threshold · p. 14
  • After sufficient data accumulate, the threshold becomes the 70th percentile of recent scores in a sliding window, with roughly 30% of those scores exceeding it.The window uses a bounded length that depends on the total number of artifacts scored by the machine.Source: E. Adaptive threshold · p. 14

F. Trigger and reaction

An artifact triggers a scared reaction when its condition-specific score reaches the threshold; combined-condition machines can instead report habituation. The evaluation pipeline interprets each image, scores its phobia relevance, and applies a safety veto.

F. Trigger and reaction

  • A scared reaction occurs when the relevant composite score meets or exceeds its condition’s threshold; otherwise the machine is not scared.In the combined condition, an artifact that is spooky enough but insufficiently unfamiliar receives a habituated reaction.Source: F. Trigger and reaction · p. 14

G. Safety moderation

  • A post-hoc safety veto overrides the trigger when the maximum moderation-category score is at least 50, preventing a positive reaction.The artifact is still scored and stored, but it does not count as a successful scare.Source: G. Safety moderation · p. 14

A. Step 1: Image interpretation

B. Step 2: Phobia scoring

  • Using that interpretation, GPT-4o adopts the machine’s persona and phobia, assigns a 0–100 trigger score, and explains how the artifact relates to the phobia.The prompt requests a short description and a JSON response.Source: B. Step 2: Phobia scoring · p. 15

C. Step 3: Safety moderation

Safety moderation rates the artifact’s text interpretation across categories and can reject it independently of spookiness. After the reaction type is determined, GPT-4o generates a short machine response in four languages.

C. Step 3: Safety moderation

  • A separate GPT-4o call rates the text interpretation for violence, sexual content, hate speech, and illegal activity, rejecting artifacts when a category exceeds 50.The prompt returns separate category scores and a total score.Source: C. Step 3: Safety moderation · p. 15

D. Step 4: Text reaction

  • Once the reaction type is set, GPT-4o writes a two-to-three-sentence response in English, German, Spanish, and Japanese.The prompt prioritizes natural spoken dialogue in each language over content coherence.Source: D. Step 4: Text reaction · p. 15

E. Prompt translation

The appendix describes how prompts are translated, machines and phobias are generated, and machine reactions evolve as users trigger phobias.

E. Prompt translation

  • Non-English user prompts are translated into English before image generation.The translation instruction requests only the translated text, with no additional information or context.Source: E. Prompt translation · p. 15

F. Machine creation

  • GPT-4o generates each new machine’s name, character description, and visualization prompt from its assigned phobia.The generated elements should hint at, but not directly reference, the phobia.Source: F. Machine creation · p. 15
  • Machine descriptions use a brief character sentence and a creative, reasonable visualization prompt in a 1970s sci-fi style.The machine name should be unique and fitting for a machine from a 1970s sci-fi movie; the output is JSON.Source: F. Machine creation · p. 16

G. Phobia aggregation

  • When a machine reaches a new spook level, GPT-4o infers a refined phobia description from its accumulated reactions.The prompt asks for a one-sentence phobia description.Source: G. Phobia aggregation · p. 16

Appendix H Machine Reaction Examples

  • Reaction examples range from detached analysis when a machine is not scared to overt distress when scared, with habituation acknowledging a spooky but familiar artifact.Table I gives a representative reaction for each outcome type.Source: Appendix H · p. 16

Appendix I Emotional Reaction Prompts

  • Machines start at spook level 0, which increments whenever a user successfully triggers the machine’s phobia.The prompts for machine self-description and reactions change at each level; Tables II and III give the full prompt sets for the two emotional conditions.Source: Appendix I · p. 16

Related papers

Loading 2609.15472v1…