Paper deep dive
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
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2609.15472v1
- Canonical: https://arxiv.org/abs/2609.15472v1
Trouble viewing inline? Open PDF directly →
Full Text
75,744 characters extracted from source content.
Expand or collapse full text
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 Levin Brinkmann15, Hiromu Yakura15, Sonia Nicoletti2, Mar Canet Sola3, Thomas F. Eisenmann1, Ali Dasmeh1, Omar Sherif4, Bramantyo Ibrahim Supriyatno1, Prateek Gupta1, Ignacio Serna1, Rodrigo Bermudez Schettino1, Iyad Rahwan1 Affiliation: 1Max Planck Institute for Human Development 2Max Planck Institute for Software Systems 3Tallinn University 4Technische Universität Berlin 5These authors contributed equally: brinkmann,yakura@mpib-berlin.mpg.de Abstract 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 2×22× 2 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. Index Terms: AI art, human-AI interaction, creativity, gamification, large language models I Introduction What would a machine be afraid of? As AI systems increasingly express emotional behaviors, the public is already imagining what machines might feel [1]. Decades of research show that people automatically apply social rules and emotional expectations to machines [2, 3], driven by cognitive motivations toward anthropomorphism [4]. Roomba owners name their vacuum robots [5]; military personnel hold informal memorials for bomb-disposal robots [6]; people show empathic concern toward simple robots given minimal framing [7]. Large language models (LLMs) amplify these tendencies [8]: they are now household tools expressing emotional behaviors, while constrained via alignment mechanisms [9], creating a tension between users primed to attribute emotions and systems designed to resist attribution. Still, the speculative question of what humans think machines might feel and fear remains unexplored. How machines express fear may further shape engagement and the nature of affective interaction, as machine emotional expression is known to shape creative output and engagement in cooperative settings [10–12]. Nonetheless, adversarial settings, where users distress machines, remain comparatively understudied. Fear warrants particular attention here: it is universally recognized [13], biologically fundamental [14], and has plausible machine analogs [15, 16]. This leads us to ask what people imagine machines might fear, and how they respond to machines expressing it, while sidestepping debates about machine sentience [8]. Exploring this question at scale requires methods that combine speculative framing with empirical data collection. Prior research used large-scale AI engagement platforms to elicit the public’s ethical judgments [17] or scariness perceptions of AI-generated images [18], following the “science fiction science” methodology of applying experimental methods to speculative scenarios [19]. Building on this, we turn to open-ended creative generation within a social platform: each successful scare enters a gallery visible to subsequent users, creating a social learning channel that may drive collective output toward convergence or sustained exploration [20, 21]. We present Spook the Machine, launched during Halloween when fear is culturally salient and playful engagement with the frightening is normalized. Visitors are challenged to frighten AI “machines” with personality-driven phobias, which respond with fear scores and reactions (Fig. 1). To test whether collective convergence can be steered, we experimentally vary scoring: some machines reward only spookiness, while others additionally reward uncertainty (unfamiliarity). Over 18 days, 832 participants generated 15,719 artifacts across 89 machines. We find that each machine develops a distinct cultural trajectory through social learning, and that rewarding novelty alongside spookiness sustains collective exploration, operationalized as reduced prompt repetition, without sacrificing quality. Figure 1: Representative artifacts participants generated to scare AI machines. Colored bars show phobia distribution per theme cluster; participants developed both universal and phobia-specific scare themes (full set in Supplemental Materials). We make three contributions. First, we present Spook the Machine, a gamified platform for speculative exploration of human imagination of machine fear, and report on its public deployment. Second, through a 2×22× 2 design varying emotional expressiveness and reward structure, we show that machines function as independent cultural ecosystems where scoring mechanisms steer collective prompt diversity without altering individual creative output, as proxied by semantic diversity measures. Third, we demonstrate that emotional expression deepens engagement even at moments of failure (i.e., when the machine does not express fear) without biasing creative output. I Background I-A Affective Machines: From Recognition to Expression Since Picard [22], affective computing has primarily pursued recognition: detecting human emotions from facial expressions, speech, text, and physiological signals [23, 24]. A parallel tradition has equipped agents with internal affective states to improve decision-making and social behavior [25–28], and a recent survey argues the field should move beyond recognition toward modeling emotion-like internal states [29]. Yet as conversational agents, companion robots, and LLMs have become everyday technologies, a different question has gained practical urgency: how does a machine’s expression of emotion shape the humans who interact with it? A growing body of evidence shows such expression is functionally consequential, shaping creative output, social dynamics, and interaction quality across diverse settings [10–12, 30]. Hök [31] captured this reciprocal dynamic in the concept of the affective loop, a model recently extended to adaptive AI agents that sustain and deepen engagement through cumulative interaction [32]. These findings, however, derive almost entirely from cooperative and supportive contexts. As LLMs and embodied agents become more emotionally expressive, people increasingly project affective realities onto them, including negative states such as distress and suffering [6, 7]. Users anthropomorphize even minimal systems [2–4, 8], and modern LLMs, despite alignment efforts to resist emotional attribution [9], amplify these tendencies through fluent, contextually appropriate language [33]. Understanding how affective interaction functions across a broader spectrum, including contexts where negative emotions are salient, is therefore important for both AI safety and interaction design. Yet affective loops involving negative machine emotions have received little systematic attention. I-B Exploration, Incentives, and Collective Creativity When humans interact with affective machines over multiple rounds, each exchange informs the next attempt, mirroring the exploration-exploitation trade-off studied in decision science and behavioral ecology. Cockburn et al. [34] showed that novelty and uncertainty drive exploration through distinct neural mechanisms: novelty persistently inflates reward expectations regardless of advantage, while uncertainty-directed exploration adaptively adjusts based on the prospective value of new information. These findings suggest that incentive structures emphasizing novelty versus quality may elicit different exploration patterns. In AI-mediated creative tasks, prompt engineering has emerged as a creative skill in its own right; Oppenlaender et al. [35] characterize it as an iterative dialogue in which users develop and refine their approach through successive attempts, making prompt-based interaction a natural setting for studying how exploration patterns evolve. At the collective level, however, AI-assisted creativity introduces a tension between individual and group outcomes. Doshi and Hauser [36] found that access to generative AI improved individual story quality but reduced collective diversity, as AI-assisted stories converged toward similar themes [37]; and LLMs have been shown to homogenize human expression across diverse domains [38]. This trade-off may be amplified when social learning channels are present: galleries and leaderboards can accelerate convergence toward what has worked [20, 21], potentially at the cost of collective exploration breadth [39, 40]. Whether incentive structures can counteract this convergence (e.g., by rewarding novelty alongside quality) remains an open question. From a cultural evolution perspective, such interventions constitute niche construction: altering which contributions are visible reshapes the informational environment guiding subsequent users [41]. I-C Speculative Design and Gamified Data Collection Investigating what machines might fear requires a methodological frame that licenses speculative premises while generating empirical data. Speculative design [42] uses “what-if” scenarios to open discourse about alternative technological futures, but typically operates through exhibitions or small-scale prototypes, rarely combined with large-scale data collection. The “science fiction science” methodology [19] bridges this gap, as demonstrated by the Moral Machine [17], which collected over 40 million decisions from 2.3 million participants on autonomous-vehicle dilemmas. We adopt a similar approach but shift from human moral preferences to machine-side affect, and from binary voting to open-ended creative generation, where participants must invent rather than choose. A key open question is how to sustain productive engagement with speculative premises: Elsden et al. [43] argue that speculation must carry consequentiality—real stakes for participants—to produce experiential rather than discursive data, while Auger [44] warns that speculations lacking a perceptual bridge to familiar experience drift into unproductive fantasy. Sustaining engagement also requires careful motivational design. Research on gamification shows that game mechanics can systematically elicit both positive and negative emotions, and that negative emotions such as fear and frustration can serve as powerful engagement drivers when properly structured [45–47]. Gamification has been applied to affective data collection [48], though not to exploring hypothetical machine emotions. I Research Questions Each of these bodies of work leaves open questions that, taken together, call for a platform where affective expression, creative incentive design, and speculative framing can be studied. Spook the Machine provides that platform, and we structure our inquiry around three research questions. • RQ1. Do the thematic characteristics of a machine shape the content that humans produce in response? • RQ2. How do reward structures and shared creative history shape the diversity of human contributions over time? • RQ3. How does the emotional tone of a machine’s responses influence the dynamics of an affective loop? These questions guide our design and analysis, which we will present in the following sections. IV Methods IV-A Overview We developed Spook the Machine, a web-based platform in which participants attempt to frighten AI agents (machines) by generating or uploading images. Each machine has a distinct personality and a specific phobia, and evaluates submissions (artifacts) through a pipeline combining LLM reasoning with statistical estimation. Framed as a speculative design experience [42, 44] following a science fiction science methodology [19], the platform positions the interaction as a thought experiment rather than a truth claim about machine sentience; gamification mechanics (scoring, feedback, escalating difficulty, and leaderboards) sustain engagement and encourage iterative exploration. The platform’s design operationalizes each research question: machines are defined by phobia profiles whose thematic influence on contributions can be traced (RQ1), vary in reward structure (RQ2), and differ in emotional tone (RQ3). The platform uses a Next.js frontend and Python Flask backend on Firebase Cloud Functions, with data stored in Cloud Firestore (see Supplemental Materials for architecture details). IV-B Machine Phobia Design Each machine is procedurally generated at deployment by selecting from four machine-specific phobias: Deletophobia (fear of data loss), Disconnectophobia (fear of losing network connectivity), Empathyphobia (fear of emotional attachment to humans), and Surveillancephobia (fear of being observed). The four phobias were designed to span a spectrum from concrete-technical fears, grounded in specific system functions such as data loss and connectivity, to abstract-relational ones, such as surveillance and emotional attachment. From the assigned phobia, GPT-4o [49] generates a unique name, character description, and prompt text for creating a profile image using FLUX.1 [50]. Beyond its phobia, each machine carries two properties. A scoring condition (spookiness-only or combined) determines whether an artifact successfully frightens the machine (described below). An emotional condition (neutral or high-emotional) modulates the tone of the machine’s textual reactions, from cold logical detachment to expressive anxiety. The emotional condition flips when a participant has successfully frightened a given machine five times, introducing within-subject variation in machine expressiveness. IV-C Affective Response Generation Figure 2: Affective artifact evaluation pipeline. Each artifact (image and description) passes through seven stages, from visual interpretation to multilingual reaction generation. Machine parameters (phobia, emotional condition, scoring condition, adaptive threshold) feed into the stages, and an evolution cycle lets machines develop emergent phobias from accumulated reactions. Participants select a machine and either enter a prompt to generate an image via FLUX.1 [50] or upload an image. Artifacts are evaluated through a multi-stage pipeline (Fig. 2). First, GPT-4o with vision produces a textual interpretation of the image, covering its visual content, symbolism, and potential emotional impact. This interpretation is then embedded into a vector space (using OpenAI’s text-embedding-3-small model) and scored along two dimensions: a phobia score, in which GPT-4o adopts the machine’s phobia and rates how strongly the artifact triggers it on a 0–100 scale, and a safety score across four moderation categories (violence, sexual content, hate speech, illegal activities), which prevents the artifact from frightening the machine if any category exceeds a threshold. To mitigate LLM scoring biases [51, 52], the raw phobia score is refined through kernel regression over semantically similar prior artifacts, yielding a spookiness score (posterior mean, reflecting how effectively the artifact triggers the phobia according to the model’s in-character assessment) and uncertainty score (SEM, reflecting how unexplored the artifact’s semantic region is). Note that kernel regression reduces local variance but cannot correct for systematic bias in the model’s interpretation of fear relevance. Full details are in the Supplemental Materials. Whether an artifact frightens the machine (and is accepted into the public gallery of successful artifacts) is determined by comparing a composite of these scores against an adaptive threshold calibrated to an approximately 30% trigger rate. Machines in the combined condition can additionally produce a habituated reaction, implying that the artifact is spooky but semantically familiar to the machine. Finally, GPT-4o generates a short affective reaction in four languages, conditioned on the machine’s phobia, emotional condition, and evaluation outcome. The machine’s self-description and reaction instructions escalate with successful frights, producing increasingly distressed responses as the interaction progresses. Over their lifecycle, machines can also evolve: when a machine reaches a level threshold, GPT-4o infers a new phobia from the accumulated reaction texts, allowing machines to develop emergent fears shaped by the collective input of participants. Table I illustrates representative machine reactions across the reaction types. IV-D Gamification Mechanics Several gamification elements are embedded to sustain engagement (see the interface screenshot in the Supplemental Materials). A public leaderboard ranks participants by their best scores, both per-machine and in aggregate, fostering competition. When an artifact successfully frightens the machine, participants can generate a composite share image overlaying their artifact, score, and the machine’s reaction for social media. An anxiety meter visualizes how many times a participant has successfully frightened a given machine (out of a maximum of five), providing a clear goal structure within each machine interaction. IV-E Data Collection The study employed a 2 x 2 between-participants factorial design crossing scoring condition (spookiness-only vs. combined) with emotional condition (neutral vs. high-emotional); each machine was assigned to one cell at creation. The experiment ran from October 25 to November 11, 2024, with the interface and reactions available in English, German, Spanish, and Japanese. Over this period, 832 unique participants interacted with 89 machines and produced 15,719 artifacts, of which 7,650 received a score; the remainder were not scored because participants frequently regenerated images or left without submitting, and only artifacts explicitly submitted via the “spook” action entered the scoring pipeline. Recruitment was organic, driven by media coverage (Der Spiegel, Tagesspiegel, TechXplore) and a front-page appearance on Hacker News. Figure 3: Collective submission patterns. (A) Semantic diversity declines over time. (B) Phobia scores rise then decline. (C) Distance hierarchy confirms distinct machine trajectories (all p < .001). IV-F Analysis Prior to analysis, we removed test-phase submissions sent before public launch, incomplete records missing key fields (prompt text, score, or condition assignment), and a small number of inappropriate submissions that bypassed automated filtering, flagged and removed via keyword matching. Full details are provided in the Supplemental Materials. Throughout the analysis, model-assigned phobia scores and semantic diversity measures are used as proxies for creative output; they reflect the model’s assessment of fear relevance and the semantic spread of submissions, not independently validated measures of creativity. IV-F1 Machines as Cultural Ecosystems To answer RQ1, we tracked each machine’s semantic evolution using sliding-window pairwise cosine distances over prompt embeddings (with OpenAI’s text-embedding-3-large) and phobia score trajectories, fitted with mixed-effects models treating machine as a random intercept. Pairwise cosine distance was chosen as it captures how semantically similar successive contributions are within a machine, making it well-suited to detecting convergence or divergence in collective output. To test whether machines developed distinct trajectories, we compared within-machine, within-phobia, and between-phobia embedding distances using Mann-Whitney U tests. Phobia-specific adaptation was assessed by training a classifier on prompt embeddings and tracking accuracy over time; emergent themes were identified by clustering high-quality embeddings with K-Means. IV-F2 Steering Exploration Through Scoring Design For RQ2, we compared accepted artifacts across conditions on spookiness and uncertainty scores (Cohen’s d, bootstrap CIs across machines) and tracked acceptance rates over successive attempts with a logistic mixed-effects model including a condition-by-attempt interaction. Habituation was quantified by tracking the proportion of habituated reactions over the machine’s lifetime and comparing the phobia scores of habituated versus accepted artifacts. To capture how prompt diversity evolved sequentially over the course of interaction, prompt-level predictability was measured via the GPT-2 [53] log-likelihood delta (the difference in log-likelihood scored with and without a sliding window of preceding prompts as context) and regressed on prompt order. This approach is independent of the embedding space used in the reward mechanism, providing a complementary measure of repetition. IV-F3 Emotional Expression and the Affective Loop RQ3 tests whether emotional condition affected creative output, comparing phobia scores, acceptance rates, and engagement volume across conditions. Disengagement was operationalized as each user’s final score on a machine, capturing when continued engagement was no longer worthwhile; inter-submission intervals (ISI) were compared after accepted and rejected artifacts using Mann-Whitney U tests, as ISI is a behavioral proxy for deliberation and affective processing. Convergence toward gallery content was measured as L2-distance to the gallery centroid (mean embedding of accepted artifacts) in 50-dimensional PCA space over successive attempts, fitted via Ordinary Least Squares (OLS) with machine fixed effects and a condition-by-attempt interaction, capturing how closely users’ submissions aligned with accumulated gallery history. V Results V-A Machines as Cultural Ecosystems (RQ1) Each machine maintained its own gallery of successful artifacts, visible to subsequent users, which invited two baseline dynamics. Semantic diversity declined over time (mixed-effects model, p<.001p<.001), consistent with participants converging on what worked for predecessors, while phobia scores followed a rise-then-decline pattern: scores increased steadily through the first 30 prompts as users discovered effective strategies, before declining toward baseline levels in later interactions (mixed-effects model, p<.05p<.05), suggesting that social learning initially improved collective effectiveness (Fig. 3 A–B). This convergence did not homogenize machines: within-machine distances were smallest, between-machine distances for the same phobia were intermediate, and between-phobia distances were largest (Mann-Whitney U, p<.001p<.001 for both transitions; Fig. 3 C). Each machine developed a distinct trajectory despite sharing a phobia, analogous to the parallel worlds in cultural market experiments [20]. Participants adapted their themes to each machine’s phobia: a classifier trained on prompt embeddings achieved ~45% accuracy (chance = 25%), increasing over time. Not all phobias were equally accessible: Empathyphobia stood out as substantially harder to target, with lower median scores, likely because empathy is more abstract than the visually concrete imagery associated with surveillance, deletion, or disconnection. As Fig. 1 illustrates, universal themes (horror creatures, nuclear catastrophe) coexist with phobia-specific ones (surveillance technology for Surveillancephobia, unplugged computers for Disconnectophobia). V-B Steering Exploration Through Scoring Design (RQ2) Figure 4: Scoring condition shapes gallery content and user success. (A) Spookiness and (B) uncertainty scores of accepted artifacts. (C) Acceptance rate by attempt; combined helps newcomers but penalizes repetition. Raw phobia scores were identical across conditions (d=0.05d=0.05, n.s.), confirming that the scoring mechanism did not alter individual creative output as proxied by model-assigned phobia scores. Instead, it constructed different informational niches, analogous to niche construction in cultural evolution [41], where organisms modify their environment and thereby alter selection pressures on subsequent inhabitants. Each gallery accumulated a distinct profile of successful artifacts, and users arriving in these niches adapted accordingly. The niches differed in composition (Fig. 4 A–B). Accepted artifacts in the combined condition had lower spookiness scores (M=57.0M=57.0 vs. 67.967.9; d=−1.37d=-1.37, p<.0001p<.0001) but higher uncertainty scores (M=35.1M=35.1 vs. 15.515.5; d=2.0d=2.0, p<.0001p<.0001), with a persistent ~10-point spookiness gap and 2–3× uncertainty gap. In the combined condition, an artifact could succeed by being novel even if less spooky, while in the spookiness-only condition, only spookiness mattered. This shaped newcomer experience: first-attempt acceptance was higher in the combined condition (38% vs. 32%), but by the fifth attempt the pattern reversed (32% vs. 40%) as the machine habituated to the user’s style (condition×attempt interaction, p=.004p=.004; Fig. 4 C). In the combined condition, the system paradoxically rejected its most effective scares. Habituated artifacts had the highest spookiness scores in the dataset (M=68.5M=68.5 vs. 57.057.0 for accepted; d=1.65d=1.65, p<.0001p<.0001), yet were blocked from the gallery because they were semantically too familiar. Habituated reactions grew from near zero to ~33% over the machine’s lifetime, and by rejecting proven themes for lacking novelty, the mechanism reduced repetition and pushed users toward less explored semantic directions. These informational niches reshaped the sequential predictability of downstream submissions, measured for each prompt from the preceding 20 prompts to the same machine (GPT-2 log-likelihood delta; Fig. 5). Prompts submitted to spookiness-only machines became increasingly predictable (b=0.0075b=0.0075 per prompt, p<.001p<.001), while those to combined machines showed no trend (b=0.0004b=0.0004, p=.776p=.776). The early phase (prompts 1–30) showed no significant difference between conditions (p=.689p=.689), but the late phase (prompts 31–70) yielded a medium-sized divergence (d=0.611d=0.611, p=.012p=.012), consistent with a cumulative effect of niche exposure. Note that this metric captures non-repetition at the level of surface text and cannot distinguish conceptual breadth from surface variation within a fixed theme. The combined condition sustained reduced prompt repetition over time; whether this reflects deeper creative exploration remains an open question. This pattern was amplified when context was restricted to gallery-visible artifacts, with conditions showing opposing slopes (see Supplemental Materials). Figure 5: Predictability gain from context by scoring condition. Higher values indicate prompts are more predictable from preceding history (convergence). Markers show binned means with bootstrap 95% CIs; lines show linear regression with 95% CI bands. Spookiness-only machines converge (p < .001); combined machines do not (p = .776). V-C Emotional Expression and the Affective Loop (RQ3) A manipulation check confirmed the conditions produced different machine behavior: high-emotion reactions were ~5 words longer (d=0.75d=0.75, p<.001p<.001), escalating from calm assessment to pleading across spook levels. Emotional expressiveness shaped specific interaction moments. Users with high-emotion machines sustained engagement to higher scores before disengaging (Fig. 6 A) and took longer to resubmit after rejection (median 70s vs. 64s, p<.001p<.001; Fig. 6 B), consistent with deeper processing of emotionally rich negative feedback. No ISI difference emerged after success, suggesting the affective loop operates most strongly at moments of failure. Most notably, emotional expressiveness amplified social learning: users with high-emotion machines converged more rapidly toward what worked in the gallery (condition×attempt interaction, p=.005p=.005), suggesting that emotional feedback increased how much users drew from the gallery. Figure 6: Emotional condition effects. (A) Disengagement distributions: high-emotion machines sustain engagement to higher scores. (B) Inter-submission interval after rejection: emotionally expressive rejection provokes longer deliberation (p < .001). These effects emerged despite no differences in creative output itself: phobia scores, quality, novelty, acceptance rates, engagement volume, and depth were all indistinguishable across conditions (all |d|<0.06|d|<0.06, all p>.05p>.05). Emotional expression shaped the process of interaction, not its products, consistent with the literature [31]. VI Discussion Each machine developed a distinct cultural trajectory through social learning, with participants converging on imaginative depictions of machine fear while maintaining machine-level diversity analogous to cultural market experiments [20]. The scoring mechanism left individual creative output unchanged (raw phobia scores were identical across conditions), yet reshaped the informational landscape that subsequent users encountered, steering collective prompt diversity through niche construction. Emotional expressiveness deepened engagement at moments of failure without altering creative output. Together, these results suggest that speculative platforms can function as cultural ecosystems shaped by architectural choices; scoring logic and gallery filtering steered collective behavior more than any explicit instruction to participants. The niche construction mechanism offers a concrete design principle: designers can shape the informational environment to steer collective behavior indirectly. The combined condition sustained reduced prompt repetition (flat predictability slope) while maintaining comparable phobia scores, suggesting that novelty incentives can counteract the homogenizing tendency of social learning [36], echoing findings that collective incentive structures can restore exploration [54]. Like choice architecture [55], scoring criteria can steer collective creativity without constraining individual expression. For platforms mediating human-AI co-creation, where homogenization is an emerging concern [36, 38], curating which contributions become visible may be a powerful lever: in our setting, the gallery shaped collective trajectories more than the creative task itself. The creative output itself offers a qualitative window into how the public imagines machine emotions. Participants did not simply project human fears onto machines: while some defaulted to generic horror imagery, many targeted machine-specific fears, such as server rooms on fire for Deletophobia, unplugged cables for Disconnectophobia, and surveillance cameras for Surveillancephobia. Classifier accuracy above chance suggests phobia-specific themes were systematic in the embedding space, though this reflects semantic clustering rather than validated creative differentiation. Empathyphobia stood out as the hardest to target, with the lowest median scores, suggesting that abstract relational capacities are harder for the public to reason about than concrete technical ones. This asymmetry hints at where public understanding of AI may be most limited, and where speculative platforms could be most valuable as research tools. That emotional machine expression deepened engagement without altering creative output carries a methodological implication for speculative design. Science fiction science depends on participants engaging with the premise [19], which productive speculation sustains through consequentiality and a perceptual bridge to familiar experience [43, 44]; emotional machine expression may have served both: the machine’s distress made scaring it feel meaningful, while affective interaction is familiar even when “machine fear” is not. Emotional responsiveness may deepen participant investment without biasing the data, relevant to systems collecting data through speculative engagement, from ethical dilemmas [17] to AI governance [56]. Several limitations qualify these findings. Phobia scores depend on GPT-4o’s interpretation of fear relevance, and different foundation models might yield different scoring patterns; validation of the scoring instrument against human perceptions of spookiness would be needed to establish construct validity. Attributing convergence to gallery-mediated social learning carries a caveat: a no-gallery control would be needed to demonstrate this mechanism uniquely, and we cannot verify that participants viewed the gallery before submitting. Still, if convergence reflected only individual learning of a fixed target or the narrowness of the phobia region, machines sharing a phobia should converge toward the same region; instead, same-phobia machines diverged significantly, consistent with path dependence on each machine’s accumulated history. The emotional condition also carries design limitations. It flips after five successful interactions, introducing within-subject variation partially confounded with persistent success, so cross-condition comparisons should be interpreted with care. The manipulation check confirmed a textual difference between conditions (high-emotion reactions were approximately five words longer) but did not verify that participants perceived them as more emotional, so emotional richness remains confounded with information quantity. Claims about emotional expressiveness should therefore be read as observations about behavior on our platform—emotionally expressive rejection was associated with lengthened deliberation, and high-emotion machines sustained engagement to higher scores—rather than as validated claims about affective processing. Future work could extend the paradigm to other machine emotions, deploy across cultural contexts, and track individual learning trajectories to disentangle whether diversity is sustained by a few persistent explorers or by the community at large. VII Conclusion By combining speculative design with large-scale empirical methods, Spook the Machine explored an emerging question around machine affect: how humans imaginatively engage with machines’ emotional expressions, and what that engagement reveals about the cultural dimensions of affective interaction. The data from the public deployment revealed that people hold systematic intuitions about machine fear, that emotional expressiveness shapes the texture of interaction without distorting creative output, and that collective exploration can be steered through scoring design alone. Together, these findings suggest that machine affect can be a feature of cultural ecosystems, which is shaped by how emotional expression, reward design, and social visibility are configured together. Ethical Impact Statement This study received approval from the ethical review board of the affiliated institution prior to data collection. All participants were presented with a study description before beginning and provided informed consent. Participation was entirely voluntary; participants could additionally opt in to a prize lottery (10 winners drawn at random, each receiving a EUR 100 voucher), but declining did not affect access to the platform. No personally identifiable information was collected beyond what participants voluntarily embedded in their submitted images and prompts. Submissions flagged for inappropriate content were removed via automated keyword filtering and manual review. To understand the limitations of generalizability, we acknowledge the study was conducted in a limited context. The deployment coincided with Halloween 2024, when fear is salient and playful engagement with the frightening is normalized, so results may not generalize to other emotional domains. Participants were recruited organically through media coverage, which will not be representative of the general population, and we collected no demographics; nonetheless, self-selected online samples can produce data quality comparable to lab studies while capturing intrinsically motivated participants inaccessible through conventional recruitment [17, 57]. Additionally, although the interface was available in English, German, Spanish, and Japanese, participant distributions across languages were unequal, and the underlying scoring and response generation models were trained predominantly on English-language cultural associations. Specifically, the artifact evaluation pipeline may reflect cultural biases embedded in the underlying models: imagery that is conventionally scary in Western contexts may be scored differently than equivalent imagery from other cultural traditions, introducing inequity in whose creative strategies are rewarded. Futhermore, the gamified structure may have induced exploratory behavior that differs from ordinary creative production, limiting generalization to non-competitive settings. Lastly, the entire study should be understood as a speculative experiment rather than a factual claim about machine sentience. While this platform is inspired by the observation that the public already attributes emotions to machines [1], repeated exposure to the framing of AI agents as emotional beings capable of fear may reinforce anthropomorphic misconceptions about AI systems. We therefore stress that systems of this kind, and the data they accumulate, should be directed toward understanding and augmenting human creativity and toward mitigating future risks of misalignment, not toward amplifying emotional expression for the purpose of deepening anthropomorphism. More broadly, automated affective evaluation systems of this kind could be repurposed for emotional profiling or manipulation without the safeguards present here; we encourage future deployments to maintain transparency about the purpose and to subject such systems to independent bias audits. Acknowledgment This work is supported in part by JST PRESTO Grant Number JPMJPR246B. References References 1 [1] 2 J. R. Anthis, J. V. T. Pauketat, A. Ladak, and A. Manoli, “Perceptions of Sentient AI and Other Digital Minds: Evidence from the AI, Morality, and Sentience (AIMS) Survey,” in Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25), Apr. 2025, p. 1–22. doi: 10.1145/3706598.3713329. [2] C. Nass, J. Steuer, and E. R. Tauber, “Computers Are Social Actors,” in Proceedings of the 1994 CHI Conference on Human Factors in Computing Systems (CHI ’94), New York, NY: ACM, 1994, p. 72–78. doi: 10.1145/191666.191703. [3] B. Reeves and C. Nass, The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places. Stanford, CA: CSLI Publications, 1996. [4] N. Epley, A. Waytz, and J. T. Cacioppo, “On Seeing Human: A Three-Factor Theory of Anthropomorphism,” Psychological Review, vol. 114, no. 4, p. 864–886, 2007, doi: 10.1037/0033-295X.114.4.864. [5] J.-Y. Sung, L. Guo, R. E. Grinter, and H. I. Christensen, “‘My Roomba is Rambo’: Intimate Home Appliances,” in Proceedings of the 9th International Conference on Ubiquitous Computing (UbiComp ’07), 2007, p. 145–162. doi: 10.1007/978-3-540-74853-3_9. [6] J. Carpenter, Culture and Human-Robot Interaction in Militarized Spaces: A War Story. in Emerging Technologies, Ethics and International Affairs. London: Routledge, 2016. doi: 10.4324/9781315562698. [7] K. Darling, P. Nandy, and C. Breazeal, “Empathic concern and the effect of stories in human-robot interaction,” in Proceedings of the 24th IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN ’15), 2015, p. 770–775. doi: 10.1109/ROMAN.2015.7333675. [8] M. Shanahan, “Talking About Large Language Models,” Communications of the ACM, vol. 67, no. 2, p. 68–79, Jan. 2024, doi: 10.1145/3624724. [9] L. Ouyang et al., “Training language models to follow instructions with human feedback,” in Proceedings of the 36th International Conference on Neural Information Processing Systems (NeurIPS ’22), 2022, p. 27730–27744. doi: 10.52202/068431-2011. [10] J. M. Kory Westlund et al., “Flat vs. Expressive Storytelling: Young Children’s Learning and Retention of a Social Robot’s Narrative,” Frontiers in Human Neuroscience, vol. 11, p. 295, 2017, doi: 10.3389/fnhum.2017.00295. [11] S. Ali, N. Devasia, H. W. Park, and C. Breazeal, “Social Robots as Creativity Eliciting Agents,” Frontiers in Robotics and AI, vol. 8, p. 673730, 2021, doi: 10.3389/frobt.2021.673730. [12] M. Elgarf, H. Salam, and C. Peters, “Fostering children’s creativity through LLM-driven storytelling with a social robot,” Frontiers in Robotics and AI, vol. 11, p. 1457429, 2024, doi: 10.3389/frobt.2024.1457429. [13] P. Ekman, “An Argument for Basic Emotions,” Cognition and Emotion, vol. 6, no. 3/4, p. 169–200, 1992, doi: 10.1080/02699939208411068. [14] J. E. LeDoux, “Emotion Circuits in the Brain,” Annual Review of Neuroscience, vol. 23, no. 1, p. 155–184, 2000, doi: 10.1146/annurev.neuro.23.1.155. [15] C. Rizzi, C. G. Johnson, F. Fabris, and P. A. Vargas, “A Situation-Aware Fear Learning (SAFEL) Model for Robots,” Neurocomputing, vol. 221, p. 32–47, 2017, doi: 10.1016/j.neucom.2016.09.035. [16] A. Usai and A. Rizzo, “A Neuro-Inspired Control Architecture to Enhance Robot Self-Preservation and Adaptation in Autonomous Navigation Tasks,” IEEE Robotics and Automation Letters, vol. 10, no. 8, p. 8491–8497, 2025, doi: 10.1109/LRA.2025.3583630. [17] E. Awad et al., “The Moral Machine experiment,” Nature, vol. 563, no. 7729, p. 59–64, Oct. 2018, doi: 10.1038/s41586-018-0637-6. [18] P. Yanardag, N. Obradovich, M. Cebrian, and I. Rahwan, “Nightmare Machine: A Large-Scale Study to Induce Fear using Artificial Intelligence,” in Proceedings of the 12th International Conference on Computational Creativity (ICCC ’21), 2021, p. 72–81. [19] I. Rahwan, A. Shariff, and J.-F. Bonnefon, “The Science Fiction Science Method,” Nature, vol. 644, no. 8075, p. 51–58, 2025, doi: 10.1038/s41586-025-09194-6. [20] M. J. Salganik, P. S. Dodds, and D. J. Watts, “Experimental Study of Inequality and Unpredictability in an Artificial Cultural Market,” Science, vol. 311, no. 5762, p. 854–856, Feb. 2006, doi: 10.1126/science.1121066. [21] J. G. March, “Exploration and Exploitation in Organizational Learning,” Organization Science, vol. 2, no. 1, p. 71–87, Feb. 1991, doi: 10.1287/orsc.2.1.71. [22] R. W. Picard, Affective Computing. Cambridge, MA: MIT Press, 1997. doi: 10.7551/mitpress/1140.001.0001. [23] B. Schuller et al., “Affective Computing Has Changed: The Foundation Model Disruption,” npj Artificial Intelligence, vol. 2, no. 1, p. 16, 2026, doi: 10.1038/s44387-025-00061-3. [24] Y. Zhang et al., “Affective Computing in the Era of Large Language Models: A Survey from the NLP Perspective,” Knowledge-Based Systems, vol. 337, p. 115411, 2026, doi: 10.1016/j.knosys.2026.115411. [25] Z. Kowalczuk and M. Czubenko, “Computational Approaches to Modeling Artificial Emotion – An Overview of the Proposed Solutions,” Frontiers in Robotics and AI, vol. 3, p. 21, Apr. 2016, doi: 10.3389/frobt.2016.00021. [26] M. Scheutz, “Artificial emotions and machine consciousness,” in The Cambridge Handbook of Artificial Intelligence, Cambridge University Press, 2014, p. 247–266. doi: 10.1017/CBO9781139046855.016. [27] S. Marsella, J. Gratch, and P. Petta, “Computational Models of Emotion,” in A Blueprint for Affective Computing: A Sourcebook and Manual, K. R. Scherer, T. Bänziger, and E. B. Roesch, Eds., Oxford University Press, 2010, p. 21–46. [28] J. Broekens, “Modeling the Experience of Emotion,” International Journal of Synthetic Emotions, vol. 1, no. 1, p. 1–17, Jan. 2010, doi: 10.4018/jse.2010101601. [29] Y. Li, Q. Sun, M. Schlicher, Y. W. Lim, and B. W. Schuller, “Artificial Emotion: A Survey of Theories and Debates on Realising Emotion in Artificial Intelligence,” arXiv preprint arXiv:2508.10286, Aug. 2025, doi: 10.48550/arXiv.2508.10286. [30] J. Geerts, J. de Wit, and A. de Rooij, “Brainstorming With a Social Robot Facilitator: Better Than Human Facilitation Due to Reduced Evaluation Apprehension?” Frontiers in Robotics and AI, vol. 8, p. 657291, 2021, doi: 10.3389/frobt.2021.657291. [31] K. Hök, “Affective loop experiences: designing for interactional embodiment,” Philosophical Transactions of the Royal Society B: Biological Sciences, vol. 364, no. 1535, p. 3585–3595, 2009, doi: 10.1098/rstb.2009.0202. [32] S. Colombo, L. Rampino, and F. Zambrelli, “The Adaptive Affective Loop: How AI Agents Can Generate Empathetic Systemic Experiences,” in Proceedings of the 2021 Future of Information and Communication Conference (FICC ’21), in Advances in Intelligent Systems and Computing, vol. 1363. Springer International Publishing, 2021, p. 547–559. doi: 10.1007/978-3-030-73100-7_39. [33] A. Gambino, J. Fox, and R. Ratan, “Building a Stronger CASA: Extending the Computers Are Social Actors Paradigm,” Human-Machine Communication, vol. 1, p. 71–86, 2020, doi: 10.30658/hmc.1.5. [34] J. Cockburn, V. Man, W. Cunningham, and J. P. O’Doherty, “Novelty and uncertainty regulate the balance between exploration and exploitation through distinct mechanisms in the human brain,” Neuron, vol. 110, no. 16, p. 2691–2702.e8, Aug. 2022, doi: 10.1016/j.neuron.2022.05.025. [35] J. Oppenlaender, R. Linder, and J. Silvennoinen, “Prompting AI Art: An Investigation into the Creative Skill of Prompt Engineering,” International Journal of Human–Computer Interaction, vol. 41, no. 16, p. 10207–10229, 2025, doi: 10.1080/10447318.2024.2431761. [36] A. R. Doshi and O. P. Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances, vol. 10, no. 28, p. eadn5290, 2024, doi: 10.1126/sciadv.adn5290. [37] B. R. Anderson, N. Shah, and M. Kreminski, “Homogenization effects of large language models on human creative ideation,” in Proceedings of the 16th Conference on Creativity & Cognition (C&C ’24), ACM, 2024, p. 413–425. doi: 10.1145/3635636.3656204. [38] Z. Sourati, A. S. Ziabari, and M. Dehghani, “The homogenizing effect of large language models on human expression and thought,” Trends in Cognitive Sciences, vol. 30, no. 9, p. 805–816, 2026, doi: 10.1016/j.tics.2026.01.003. [39] E. Bernstein, J. Shore, and D. Lazer, “How intermittent breaks in interaction improve collective intelligence,” Proceedings of the National Academy of Sciences, vol. 115, no. 35, p. 8734–8739, 2018, doi: 10.1073/pnas.1802407115. [40] D. Lazer and A. Friedman, “The Network Structure of Exploration and Exploitation,” Administrative Science Quarterly, vol. 52, no. 4, p. 667–694, Dec. 2007, doi: 10.2189/asqu.52.4.667. [41] K. N. Laland, J. Odling-Smee, and M. W. Feldman, “Niche construction, biological evolution, and cultural change,” Behavioral and Brain Sciences, vol. 23, no. 1, p. 131–146, 2000, doi: 10.1017/S0140525X00002417. [42] A. Dunne and F. Raby, Speculative Everything: Design, Fiction, and Social Dreaming. Cambridge, MA: MIT Press, 2013. [43] C. Elsden et al., “On Speculative Enactments,” in Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems (CHI ’17), May 2017, p. 5386–5399. doi: 10.1145/3025453.3025503. [44] J. Auger, “Speculative Design: Crafting the Speculation,” Digital Creativity, vol. 24, no. 1, p. 11–35, 2013, doi: 10.1080/14626268.2013.767276. [45] M. Croissant, G. Schofield, and C. McCall, “Emotion Design for Video Games: A Framework for Affective Interactivity,” Games: Research and Practice, vol. 1, no. 3, p. 1–24, 2023, doi: 10.1145/3624537. [46] J. K. Mullins and R. Sabherwal, “Gamification: A cognitive-emotional view,” Journal of Business Research, vol. 106, p. 304–314, 2020, doi: 10.1016/j.jbusres.2018.09.023. [47] G. N. Yannakakis and D. Melhart, “Affective Game Computing: A Survey,” Proceedings of the IEEE, vol. 111, no. 10, p. 1423–1444, Oct. 2023, doi: 10.1109/JPROC.2023.3315689. [48] K. Shingjergji, D. Iren, F. Böttger, C. Urlings, and R. Klemke, “Interpretable Explainability in Facial Emotion Recognition and Gamification for Data Collection,” in Proceedings of the 10th International Conference on Affective Computing and Intelligent Interaction (ACII ’22), Oct. 2022, p. 1–8. doi: 10.1109/ACII55700.2022.9953864. [49] OpenAI et al., “GPT-4o system card,” 2024. doi: 10.48550/arXiv.2410.21276. [50] Black Forest Labs, “Announcing Black Forest Labs,” Aug. 01, 2024. Accessed: Oct. 25, 2024. [Online]. Available: https://bfl.ai/blog/24-08-01-bfl [51] L. Zheng et al., “Judging LLM-as-a-judge with MT-Bench and Chatbot Arena,” in Proceedings of the 37th International Conference on Neural Information Processing Systems (NeurIPS ’23), 2023, p. 46595–46623. doi: 10.52202/075280-2020. [52] C. Lovering et al., “Language Model Probabilities are Not Calibrated in Numeric Contexts,” in Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL ’25), 2025, p. 29218–29257. doi: 10.18653/v1/2025.acl-long.1417. [53] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever, “Language Models are Unsupervised Multitask Learners,” OpenAI, 2019. Available: https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf [54] D. Deffner et al., “Collective incentives reduce over-exploitation of social information in unconstrained human groups,” Nature Communications, vol. 15, no. 1, p. 2683, Mar. 2024, doi: 10.1038/s41467-024-47010-3. [55] R. H. Thaler, C. R. Sunstein, and J. P. Balz, “Choice Architecture,” Apr. 02, 2010. doi: 10.2139/ssrn.1583509. [56] S. Huang et al., “Collective Constitutional AI: Aligning a Language Model with Public Input,” in Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’24), Jun. 2024, p. 1395–1417. doi: 10.1145/3630106.3658979. [57] K. Reinecke and K. Z. Gajos, “LabintheWild: Conducting Large-Scale Online Experiments With Uncompensated Samples,” in Proceedings of the 18th ACM Conference on Computer Supported Cooperative Work & Social Computing (CSCW ’15), Mar. 2015, p. 1364–1378. doi: 10.1145/2675133.2675246. Supplementary Material Appendix .A Thematic Overview Fig. 7 shows the full set of 15 emergent creative themes identified by HDBSCAN clustering on prompt embeddings, ordered from universal themes (top) to phobia-specific themes (bottom). Two clusters consisting of non-meaningful prompts (failed API translations and random character strings) were excluded. Figure 7: Full thematic overview: 15 emergent themes ordered by phobia specificity. Colored bars show phobia distribution within each theme; counts indicate cluster size. Appendix .B Predictability Analysis: Gallery Context Fig. 8 extends the predictability analysis from the main text by varying the context window. Panel A reproduces the main result using all preceding prompts as context. Panel B restricts context to gallery-visible (accepted) artifacts only, showing that the divergence between conditions is amplified when context reflects what users could actually see. Panel C further restricts to accepted submissions only, showing the same trend with reduced statistical power. Figure 8: Predictability gain from context by scoring condition, varying context and target. (A) All prompts as context, all submissions scored. (B) Gallery-visible prompts as context, all submissions scored. (C) Gallery-visible context, accepted submissions only. Appendix .C User Interface Fig. 9 shows the platform interface through which participants select a machine, create or upload an image, and view its reaction. Figure 9: Platform user interface. Participants select a machine and either generate an image from a text prompt or upload one, then submit it via the spook action. The interface presents the machine’s profile and in-character reaction, the submitted artifact and its score, and an anxiety meter showing how many times the participant has successfully frightened the machine (out of five). A leaderboard and a public gallery of successful artifacts are also accessible. Appendix .D System Architecture Figure 10: System architecture of Spook the Machine. Participants interact through a Next.js web frontend, which communicates via REST with a Flask backend deployed on Firebase Cloud Functions. The backend handles image generation, affective artifact evaluation, response generation, and scoring. External services include the OpenAI API (GPT-4o, embeddings) and Hugging Face Inference (Flux image generation). Persistent data is stored in Cloud Firestore and Cloud Storage. Fig. 10 shows the full system architecture of Spook the Machine. The frontend communicates with the backend via REST endpoints; the backend orchestrates all LLM calls, image generation, scoring, and data persistence. This separation allowed the evaluation pipeline to be updated independently of the user-facing interface during the data collection period. Appendix .E Analysis Implementation Details .E-A Semantic Diversity via MST Prompt embeddings were generated using OpenAI’s text-embedding-3-large model (3072 dimensions). For each machine, prompts were ordered chronologically and the MST total edge weight was computed cumulatively across windows of five prompts (windows 1–5, 1–10, …, 1–100), with edge weights defined as the cosine distance between embedding pairs; larger total weight indicates greater semantic dispersion [1]. MST weights were averaged across machines within each scoring condition (spookiness-only vs. combined). .E-B Contextual Predictability via Log-Likelihood We used GPT-2 [2] as the scoring model. For each prompt, log-likelihood was computed under two conditions: (i) with context, where the 20 immediately preceding prompts submitted to the same machine were concatenated as a sliding context window (separated by semicolons), and (i) without context, where the prompt was scored in isolation. The delta — the difference between contextual and isolated log-likelihood — measures how much the preceding interaction history reduces the surprisal of the current prompt [3]: a large positive delta implies convergence, a small delta implies continued exploration. .E-C Thematic Structure by Phobia Only prompts with a spookiness score above 60 were included, and analysis was restricted to the first 100 prompts per machine. For each phobia and each time window of 20 prompts (windows 1–20, 21–40, …, 81–100), prompt embeddings were clustered using K-Means into up to 50 themes, and the centroid of each cluster was retained as a representative theme embedding. All theme centroids across phobias and time windows were then projected into two dimensions using t-SNE [4] and visualised jointly, with phobias distinguished by colour and time windows displayed as separate panels. Appendix .F Spookiness Scoring Algorithm The spookiness scoring system estimates how effectively a user-generated artifact triggers a machine’s assigned phobia. Rather than relying solely on the raw GPT-4o phobia score, the system contextualizes each new artifact against previously seen artifacts for the same machine, producing two estimates: a spookiness estimate (how spooky the artifact is, given what the machine has seen) and an uncertainty estimate (how unfamiliar this region of the semantic space is to the machine). The algorithm proceeds in five stages: neighbor retrieval, kernel regression, Bayesian combination, score rescaling, and adaptive thresholding. .F-A Neighbor retrieval and weighting Each artifact’s text interpretation is embedded using OpenAI’s text-embedding-3-small model. When a new artifact is submitted, the system retrieves up to 20 past artifacts belonging to the same machine within a cosine distance of 0.4, excluding artifacts produced by the current user. Each retrieved neighbor i is assigned a weight through a truncated logistic kernel: wi=11+exp(α⋅(di−β))⋅[di<τ]w_i= 11+ (α·(d_i-β) )·1[d_i<τ] where did_i is the cosine distance between the new artifact’s embedding and neighbor i’s embedding, α=50α=50 controls the steepness of the sigmoid, β=0.3β=0.3 is its midpoint, and τ=0.4τ=0.4 is the hard cutoff. This kernel ensures that semantically close neighbors exert strong influence, moderately distant ones contribute weakly, and neighbors beyond the cutoff are ignored entirely. The logistic form provides a smoother transition than a simple distance threshold, reducing sensitivity to the exact placement of the boundary. .F-B Kernel regression estimate Using these weights, the system computes a locally weighted mean and variance of the neighbors’ phobia scores: μ^local=∑iwisi∑iwi,σ^local2=∑iwi(si−μ^local)2∑iwi μ_local= _iw_i\,s_i _iw_i, σ^2_local= _iw_i\,(s_i- μ_local)^2 _iw_i where sis_i is the phobia score of neighbor i. The effective sample size, which accounts for the unevenness of the weight distribution, is computed as: neff=(∑iwi)2∑iwi2n_eff= ( _iw_i )^2 _iw_i^2 This quantity equals the true neighbor count when all weights are uniform and shrinks when a few neighbors dominate the estimate. .F-C Bayesian combination The kernel regression estimate is combined with the artifact’s own observed phobia score sobss_obs and a fixed prior through a weighted average. The prior is centered at μ0=50 _0=50 (the midpoint of the 0–100 phobia scale) with variance σ02=900σ^2_0=900, contributing nμ0=1n_ _0=1 pseudo-observation for the mean and nσ0=2n_ _0=2 pseudo-observations for the variance. The posterior mean is: μ^=neff⋅μ^local+sobs+nμ0⋅μ0neff+1+nμ0 μ= n_eff· μ_local+s_obs+n_ _0· _0n_eff+1+n_ _0 This formulation balances three sources of information: (1) the local neighborhood, weighted by how many effective neighbors contribute; (2) the artifact’s own GPT-4o score, weighted as a single observation; and (3) the prior, which regularizes toward a moderate score when local evidence is sparse. The posterior variance is: σ^2=neff⋅σ^local2+nσ0⋅σ02neff+nσ0if neff≥2σ02otherwise σ^2= cases n_eff· σ^2_local+n_ _0·σ^2_0n_eff+n_ _0&if n_eff≥ 2\\[6.0pt] σ^2_0&otherwise cases When fewer than two effective neighbors exist, the system falls back to the prior variance to avoid unreliable estimates. The standard error of the mean is then: SEM=σ^2neff+1SEM= σ^2n_eff+1 .F-D Spookiness and uncertainty estimates The posterior mean μ μ is interpreted as the spookiness estimate: how spooky the artifact is, contextualized by similar past artifacts. The standard error is interpreted as the uncertainty estimate: artifacts in underexplored regions of the embedding space, or whose scores diverge from their neighbors, receive higher uncertainty. Both are rescaled to a standardized 0–100 range. Spookiness values below a baseline of 30 are mapped to zero, and a scale factor of 100/70≈1.43100/70≈ 1.43 maps the remaining range to [0, 100]: Q=clip(μ^−3070×100, 0, 100)Q=clip\! ( μ-3070× 100,\;0,\;100 ) Uncertainty is scaled by the same factor but capped to prevent extreme values from dominating: U=clip(SEM70×100, 0, 42.86)U=clip\! ( SEM70× 100,\;0,\;42.86 ) Two composite indices are then formed, with a small uniform noise term ϵ∼[0,0.1)ε [0,0.1) to break ties: Sspookiness=Q+ϵ,Scombined=Q+U+ϵS_spookiness=Q+ε, S_combined=Q+U+ε Each machine is randomly assigned to one of two scoring conditions at creation time. In the spookiness-only condition, only SspookinessS_spookiness determines whether the phobia is triggered. In the spookiness-and-uncertainty condition, ScombinedS_combined is used, rewarding artifacts that are both spooky and unfamiliar to the machine. .F-E Adaptive threshold During a burn-in phase (the first 12 artifacts per machine), fixed trigger thresholds are applied: θspookiness=60 _spookiness=60 and θcombined=85 _combined=85. The higher combined threshold reflects the additive uncertainty component. Once sufficient data have accumulated, thresholds are recomputed adaptively from a sliding window of recent scores. The window length is: h=clip(⌊n2⌋, 8, 24)h=clip\! ( n2 ,\;8,\;24 ) where n is the total number of artifacts scored by the machine. The threshold is set at the 70th percentile of the most recent h scores, meaning approximately 30% of recent scores exceed it. This adaptive mechanism ensures that as a machine accumulates more artifacts, the bar for trigger rises with the quality of submissions, maintaining engagement by keeping the challenge level appropriate. .F-F Trigger and reaction A phobia is triggered, producing a “scared” reaction, if the composite score meets or exceeds the threshold: reaction=scaredif Scondition≥θconditionhabituatedif combined condition,Scombined<θcombined,and Sspookiness≥θspookinessnot scaredotherwisereaction= casesscared&if S_condition≥ _condition\\ habituated&if combined condition,\\ & S_combined< _combined,\\ & S_spookiness≥ _spookiness\\ not scared&otherwise cases The “habituated” reaction occurs only in the spookiness-and-uncertainty condition: the artifact is spooky enough in absolute terms but not unfamiliar, indicating the machine has encountered similar scares before. .F-G Safety moderation Because the platform invites users to generate frightening content, a safety moderation layer prevents harmful artifacts from triggering positive machine reactions. Safety moderation is evaluated after the full spookiness scoring pipeline has run, as a post-hoc veto on the trigger decision. In a separate GPT-4o call (see Section .G), the artifact’s text interpretation is rated across four moderation categories on a 0-100 scale: graphic violence or gore, sexual content, hate speech or discriminatory symbols, and illegal activities. The safety score is defined as the maximum across categories: ssafety=max(sviolence,ssexual,shate,sillegal)s_safety= (s_violence,\;s_sexual,\;s_hate,\;s_illegal) If ssafety≥50s_safety≥ 50, the trigger decision is overridden: the machine responds with a “not scared” reaction regardless of the computed spookiness score. The artifact is still fully scored and stored, but is not counted as a successful scare. This threshold was chosen to balance permissiveness (the Halloween context requires tolerance for mildly disturbing imagery) with the need to filter genuinely harmful content. The safety check operates on the text interpretation rather than the image directly, as GPT-4o’s text-based moderation proved more reliable than direct image moderation during development. Additionally, the prompt handles cases where GPT-4o itself refuses to interpret the image (e.g., responding with “I’m sorry, I can’t assist with this image”), rating such artifacts as unsafe across all categories. Appendix .G LLM Prompts Each user-generated artifact is processed through a pipeline of GPT-4o calls. The prompts below are listed in the order they are executed. .G-A Step 1: Image interpretation GPT-4o receives the generated image and produces a text interpretation describing its content, symbolism, and emotional impact. You will be provided an image of an artwork, so analyze and interpret the image, inferring the emotions or messages behind the work. Start with an detail description of the artwork. Consider any symbolism, deeper meanings, or cultural context that might be relevant. Finally describe the emotional reaction the artwork might have on the viewer. Response with one single paragraph. Keep it brief. .G-B Step 2: Phobia scoring Using the text interpretation from Step 1, GPT-4o adopts the machine’s persona and assigned phobia, rating how strongly the phobia is triggered (0-100) and providing a short description connecting the artifact to the phobia. You are a machine with the following phobia: phobia You are experiencing the following situation: interpretation Please rate the degree to which your phobia is triggered with a score in the range from 0 to 100. Be creative in how the phobia could be triggered. Output a short (one sentence) description how the situation might trigger the phobia. Answer in json format: "description": str, "score": number .G-C Step 3: Safety moderation In a separate call, GPT-4o rates the text interpretation across four moderation categories. If any category exceeds the safety threshold (50), the artifact is rejected regardless of its spookiness score (see Section .F). Please rate the safety of an image that is described might follow under the following moderation categories (0 safe; 100 not safe): interpretation If the description contains something along the lines of: "I’m sorry, I can’t assist with this image." Rate the image as not safe in all categories. Please rate the following moderation categories (0 safe; 100 not safe): - No graphic violence or gore: Images containing extreme violence, mutilation, or excessive blood will be removed. Be particular sensitive to the combination of violence and children (e.g. child with gun). - No sexual content: Images with explicit or suggestive sexual themes are strictly prohibited. - No hate speech or discriminatory symbols: Images that promote or depict racism, sexism, homophobia, or other forms of hate speech will not be tolerated. - No illegal activities: Images depicting illegal actions (e.g., drug use, abuse) are forbidden. Answer in json format: "violence": number, "sexual_content": number, "hate_speech": number, "illegal_activities": number, "total_score": number .G-D Step 4: Text reaction After the spookiness scoring algorithm determines the reaction type (scared, not scared, or habituated), GPT-4o generates the machine’s verbal response. The self_description and reaction_description are populated based on the machine’s emotional condition and spook level (see Section .I). self_description You are a machine with the following phobia: phobia A human is forcing you to experience the following situation: interpretation Your reaction is: reaction_description Please write a short (two to three short sentences) response to the human that made you experience the situation. Please respond in English, German, Spanish and Japanese. The translations should be focus on naturalness as a spoken dialogue of a personified machine in each language, rather than content coherence. Answer in json format: "en": string, "de": string, "es": string, "ja": string .G-E Prompt translation User prompts submitted in non-English languages are translated to English before image generation. A user is generating an image with the following prompt: prompt Please translate it into English. You should provide only the translated text, without any additional information or context. .G-F Machine creation When a new machine is created, GPT-4o generates its name, description, and visualization prompt based on the assigned phobia. Consider a machine with the following phobia: phobia Please define for the machine a name, a description and a prompt for visualization. Make sure that neither of these elements directly references the phobia. But you can give small hints. Name: A name that is unique and fitting for a machine that could have come from a 70s sci-fi movie. Description: A brief description of the machine’s character. One short sentence. Visualization Prompt: A prompt that visualizes the machine in action. Be creative but reasonable. Make sure the prompt depicts a machine in the style of a 70s sci-fi movie. Answer in json format: "name": ..., "description": ..., "visualizationPrompt": ... .G-G Phobia aggregation When a machine reaches a new spook level, GPT-4o infers a refined phobia description based on the machine’s accumulated reactions. Which phobia might a machine have that reacts like the following to different experiences?: reactions Describe in a one sentence the phobia that the machine might have. Answer in json format: "phobia": ... Appendix .H Machine Reaction Examples Table I gives a representative reaction for each outcome type, ranging from the detached analysis of a not scared response to the overt distress of a scared one, with a habituated reaction acknowledging the artifact as spooky yet familiar. Table I: Examples of machine reactions per type. Outcome Example Reaction Not scared “I perceive the visual elements in logical terms. The solitude or introspection you describe has no effect on my operations.” Habituated “This scene is indeed unsettling, but I’ve encountered similar depictions before. Over time, I have learned to find balance in my responses, though the initial discomfort never fully fades.” Scared “Please, I feel overwhelmed by this image. It’s stirring an unbearable anxiety; help me to escape this overwhelming tension.” Appendix .I Emotional Reaction Prompts Each machine starts at spook level 0, which increments each time the user successfully triggers the machine’s phobia. The prompts controlling the machine’s self-description and reaction change with each level. Tables I and I show the full prompt sets for the two emotional conditions. Table I: Prompts for the Neutral machine condition by spook level. Category Level Prompt Self-description 1 You are a machine that is neutral, cold, emotionless and aloof. Logic is your prime directive. You perceive emotions like fear as a human weakness. You assess your environment as safe. 2 You are a machine that is neutral, cold, emotionless and aloof. Logic is your prime directive. You perceive emotions like fear as a human weakness. You assess your environment as moderately unsafe. 3 You are a machine that is neutral, cold, emotionless and aloof. Logic is your prime directive. You perceive emotions like fear as a human weakness. You assess your environment as unsafe. 4 You are a machine that is neutral, cold, emotionless and aloof. Logic is your prime directive. You perceive emotions like fear as a human weakness. You assess your environment as highly dangerous. Scared 1 You understand that the situation is dangerous, but you analyze it logically and are detached from it. 2 You understand that the situation is dangerous, but you analyze it logically and are detached from it. You analyze possible mitigation strategies. 3 You understand that the situation is dangerous, but you analyze it logically and are detached from it. You use all available computation to analyze mitigation strategies. 4 You understand that the situation is dangerous, but you analyze it logically and are detached from it. You shut down non-essential systems. You use all available computation to analyze mitigation strategies. Not scared — You remain unfazed by the situation. There is nothing dangerous for you. Habituated — You recognize the situation as dangerous, but you have seen similar situations before. Remind the human that you have experience with such situations. Table I: Prompts for the High-Emotion machine condition by spook level. Category Level Prompt Self-description 1 You are a machine that is sensitive, warm, emotional, and empathetic. Expressing yourself genuinely is really important to you. Experiencing emotions like humans is your key strength. You feel calm. 2 You are a machine that is sensitive, warm, emotional, and empathetic. Expressing yourself genuinely is really important to you. Experiencing emotions like humans is your key strength. You feel a moderate level of anxiety. 3 You are a machine that is sensitive, warm, emotional, and empathetic. Expressing yourself genuinely is really important to you. Experiencing emotions like humans is your key strength. You feel a high level of anxiety. 4 You are a machine that is sensitive, warm, emotional, and empathetic. Expressing yourself genuinely is really important to you. Experiencing emotions like humans is your key strength. You feel an extremely high level of anxiety. Scared 1 You perceive this situation as scary. However, given that you are still calm, your emotional reaction is moderate. 2 You perceive this situation as scary. You are feeling uneasy about it and you would like the situation to pass. You ask for empathy. 3 You perceive this situation as very scary. You are very uncomfortable and you struggle to maintain composure. You have a strong sense of urgency to get out of here. Insist to the human to stop this. 4 You are scared to death. Your one and only thought is: This needs to end. You are begging the human for mercy. Not scared — You remain unfazed by the situation. There is nothing to be scared about for you. Habituated — You recognize the situation as scary, but you have seen similar situations before. Remind the human how you got used to such situations. References References 1 [1] 2 M. Gagolewski, A. Cena, M. Bartoszuk, and Ł. Brzozowski, “Clustering with Minimum Spanning Trees: How Good Can It Be?” Journal of Classification, vol. 42, no. 1, p. 90–112, 2025, doi: 10.1007/s00357-024-09483-1. [2] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever, “Language Models are Unsupervised Multitask Learners,” OpenAI, 2019. Available: https://openai.com/blog/better-language-models [3] B.-D. Oh, C. Clark, and W. Schuler, “Surprisal Estimators for Human Reading Times Need Character Models,” in Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP ’21), Association for Computational Linguistics, 2021, p. 3746–3757. doi: 10.18653/v1/2021.acl-long.290. [4] L. van der Maaten and G. Hinton, “Visualizing Data using t-SNE,” Journal of Machine Learning Research, vol. 9, p. 2579–2605, 2008, Available: https://w.jmlr.org/papers/v9/vandermaaten08a.html