Frontiers in Computing and Intelligent Systems ISSN: 2832-6024 | Vol. 13, No. 2, 2025 65 A Multimodal Analysis of Human Imitation of AI and Human–Machine Relations: A Case Study of 50 TikTok Videos Xi Wang * School of journalism and communication, Lanzhou University, Lanzhou, Gansu, China * Corresponding author Email: 2278465811@qq.com Abstract: In recent years, the proliferation of illogical and “grotesque” AI-generated videos on social media has given rise to a cultural trend in which humans imitate AI in reverse. This form of mimicry, which takes AI characteristics as its creative source, has drawn increasing scholarly attention. This study conducts an empirical investigation into the phenomenon of human imitation of AI, analyzing 50 sample cases extracted from the Douyin platform. By applying Norris’s multimodal interaction analysis framework, the research further examines the multimodal characteristics of such content. The findings suggest that in the era of intelligent media, mimicry evolves into a phenomenon of multi-directional co-construction and symbiosis between humans and algorithms. While such videos share common multimodal features and serve primarily as meme-like cultural dissemination, they also reflect underlying anxieties about human–machine relations. Keywords: Human Imitation of AI; Human–machine Relations; Multimodality; Theory of Mimicry; Generative AI. 1. Introduction In recent years, a widely circulated short-video creation phenomenon on social media is the AI-generated “TV characters reconcile” clips, in which creators feed into the model scenes of impending conflict between characters, accompanied by keywords instructing the AI to alter the outcome so that reconciliation is achieved. While early “AI misunderstanding prompts” emerged as accidental consequences of technical deficiencies, productions such as “TV characters reconcile” are deliberately engineered by users to exploit motion-generation technologies for dramatic effect. Within this proliferation of AI-mediated creative works, a distinctive phenomenon has emerged: human reverse imitation of AI. This refers to practices in which humans, in their behaviors or creative activities, learn from, replicate, or simulate the forms of AI-generated outputs. At its core, the phenomenon of “humans imitating AI” constitutes a novel convergence of AIGC (AI-generated content) and online parody culture, sparking the rise of a new wave of subcultural expression (Zheng Gucheng & Yuan Aiqing, 2025) [1]. Both creators and audiences immerse themselves in the carnivalesque entertainment of parodying AI-generated anomalies, conveying a distinctive attitude that oscillates between ironic acceptance and oppositional mimicry (Du Yitong, 2025) [2]. To date, however, existing research has largely remained speculative and reflective in nature, lacking empirical evidence grounded in standardized research methods. Recognizing that human imitation of AI often manifests relatively stable multimodal characteristics, this study adopts [3] framework of multimodal interaction analysis to examine 50 Douyin videos featuring human imitation of AI. Additionally, the study incorporates participant observation—researchers engage with video- creation practices as ordinary users—in order to investigate the patterned features of mimicry on Douyin, thereby enriching understandings of the evolving dynamics of human–machine relations in the context of intelligent media communication. 2. Literature Review 2.1. Human-Machine Relationship With the deep integration of artificial intelligence (AI) technologies into research, production, and everyday life, human–machine relations increasingly embody both division of labor and collaboration, constructing new modes of cooperation through complementary capacities. Human–AI co-creation represents a new form of collaborative creativity, which, while enhancing cooperation, simultaneously challenges fundamental issues of traditional human–machine relations such as “subjectivity” and “ethical boundaries.” This process propels human–machine relations toward greater complexity, as deep integration between humans and AI ushers in a “new content revolution” (Zhong Jianqin, 2020) [4]. In the course of co-creation, AI—by virtue of its capacity to integrate individual memory and engage in human– machine co-construction—more covertly aggregates and processes personal memory. This dynamic fosters the embedding of technological mediation into human subjectivity, thereby internalizing AI-shaped memory as part of individual memory itself (Zhou Li, Chen Muen, 2023) [5], and advancing human–AI relations to a deeper level. Consequently, the binary paradigm in prior research, which positioned humans as subjects and machines as mediating objects, is no longer adequate. A new configuration of human–machine relations, grounded in the ecology of intelligent communication, is taking shape (Ding Fangzhou, Zhang Duo, 2025) [6]. Against this backdrop, scholarly debates diverge on the implications of generative AI’s large- scale involvement in creative practices: some argue that the presence of generative AI in creative work engenders a sense of frustration (Jiang Qiaolei, Shi Mengsu, 2025) [7], as the technology’s competence threatens human superiority (Schmitt, 2020) [8], prompting resistance to its adoption. 66 Others contend that the application of AI should not be seen as a subversion of existing artistic value systems; rather, its technological form remains within the realm of tool-based innovation, and its artistic production mechanism constitutes a structured simulation of human experience and emotion (Ni Wan, Guo Jialiang, 2024) [9]. 2.2. Mimetic Theory From the perspective of traditional theories of mimesis, “mimesis” is a foundational concept in Western metaphysical philosophy with profound implications for the visual arts. It encompasses multiple dimensions, including imitation of the world of ideas, of divinity, of nature, of self-experience, and of the works of ancient masters. Embedded within these meanings are explorations of the relationship between rationality and sensibility, which have been interpreted in different ways across historical periods [10]. In the era of intelligent media, the scope, speed, and impact of mimetic practices have expanded unprecedentedly. The interplay of imitation among artists, the general public, and AI has become inextricably intertwined. Fluid cyber spaces liberate individuals from the constraints of natural temporality and spatiality, enabling them to participate in the public domain as producers, to act freely beyond the limits of the physical body, and to construct frameworks of meaning (Zhan Di, Fang Jieyun, 2025) [11]. Existing scholarship has primarily focused on AI imitating humans within the paradigm of mimesis in intelligent media contexts, while some studies have used mimesis as an analytical lens to examine imitation behaviors in short videos, highlighting how individuals pursue a sense of belonging in virtual networks and how authentic human life becomes increasingly separated from, and alienated by, mediated representations (Sun Yuting, 2020) [12]. hese perspectives provide useful references for the present study’s exploration of the phenomenon of humans imitating AI. 3. Research Framework In order to capture users’ attention within the shortest span of time, short videos tend to rely on the interplay of multiple modalities to highlight their themes. Rapid camera switches, caption effects, background music, and hosts’ bodily gestures form the characteristic dense collage of multimodal symbols. The interactive essence of short videos is reflected not only in the multimodal compositional strategies employed during user production, but also in the multisensory cognitive processes activated during audience reception. This phenomenon is naturally aligned with multimodal interaction analysis theory, which provides a valuable lens for examining both the representational forms and the organizational logic of human imitation of AI. Norris’s theory of multimodal interaction analysis is built on an interdisciplinary foundation, integrating interactional sociolinguistics, mediated discourse theory, and multimodal discourse analysis into a dynamic analytical framework. The theory adopts “activity” as the basic unit of analysis and distinguishes between lower-level actions (micro-symbolic systems such as a single facial expression or gesture), higher- level actions (relatively complex and macro units composed of multiple lower-level actions), and frozen actions (modalities of action that do not appear in practice but remain latent and potential). Moreover, it emphasizes the collaborative mechanisms across modalities such as speech, visual symbols, and spatial arrangements. To investigate the interactive behaviors of human imitation of AI within the domain of short videos, this study employs multimodal transcription and video analysis. The process involves three stages: first, transcribing individual communicative modalities (lower-level actions); second, integrating two or more of these modalities to analyze and interpret higher-level actions with clearer communicative functions and meanings; and finally, synthesizing all transcriptions across modalities to complete the overall interpretation of meaning. 4. Results of Multimodal Interaction Analysis 4.1. Presentation of Classification Results At present, Douyin dominates the short video sector in China, with a stable daily active user base exceeding 600 million and a user demographic that spans across age groups and regions, making it the leading platform nationwide [13]. This paper takes videos of humans imitating AI on the Douyin platform as the research object. This study takes videos of humans imitating AI on the Douyin platform as its research object. According to the researcher’s observations, the phenomenon of human imitation of AI in short videos broke out intensively in July 2024 and has continued since then. Between July 7, 2024, and January 1, 2025, the researcher conducted focused observation of 127 videos retrieved under the keywords Human Imitation AI Contest and Reverse AI Challenge. After excluding irrelevant and duplicate cases, 50 videos with relatively high popularity (measured by views, likes, saves, and comments) were selected as the final sample for analysis. To address RQ1 and RQ2, the 50 samples were coded and categorized according to their presentational forms (see Table 1), and the frequency of each category was calculated. Subsequently, one representative case was selected from each category for further analysis, applying Norris’s multimodal interaction analysis framework. The coding scheme for the 50 observed cases is as follows: Table 1. Sample form code table class definition representative works quantity AI changes the ending The ending of the video reverses the character, is simple and easy to imitate, and usually only has a single reversal factor and a simple mode @Bai Cai AIGC 23 Free expression of the subject Imitate the AI by performing without props @Jiaber Wu Botao 12 Non-related elements enter the picture Random elements suddenly appear in the painting imitating AI @Buqi Dance Troupe 9 AI plot interpretation The video content has obvious plot interpretation but the ending is unexpected, including many AI elements and a variety of modes @Mu Tianran 6 67 4.2. Data Coding Results 4.2.1. Category 1: AI-Altered Endings. This type of video is typically short in duration and simple in content, involving relatively few modalities. Most are produced by ordinary users for entertainment, although some commercial advertisers also adopt this form to promote their products in line with their content style. Representative work: @Baicai AIGC: “AI doesn’t understand me at all! I am clearly on a fat-loss diet!”. Video Description: The video runs for a total of 26 seconds and is shot entirely with a fixed camera angle, featuring a man eating. The screen is split into two parts: the left panel, labeled “old photo,” shows a static medium close-up of the man holding vegetables; the right panel is the dynamic version of the same photo, where the man throws away the vegetables and instead retrieves fried food from off-screen. Based on the integration of modalities, it was found that the video consists of multiple sets of parallel visual modalities, which together construct a higher-level action: creating a reversal effect of “AI does not understand me” or “AI changes the ending” by performing actions that contradict the original meaning of the photo. The performance mainly relies on props (different types of food) to enhance the mimetic effect, corresponding to the typical features of AI-generated videos in which irrelevant elements are introduced into the frame, thereby violating narrative logic. This type of video represents the largest category among the sampled cases. It is characterized by relatively low modal density and high reproducibility, making it easier for users to imitate. According to the researcher’s observation, creators of this type are generally ordinary users who follow online trends. Their storylines are relatively simple, typically staging actions opposite to the original photo or plot. 4.2.2. Category 2: Subjective Free Performance This category mainly consists of creators mimicking AI through their own bodily actions, without relying on explicit narrative logic or external props. Such videos particularly test human capacity for imitation—attempting to reproduce the whimsical qualities of AI without the aid of special effects or tools. As such, this category serves as a key site for studying human mimetic behaviors and characteristics. Representative Example: @Jiaber Wu Botao: “AI has changed people’s lives”. Video Description: The video is 18 seconds long and adopts a split-screen format to simulate photo restoration. The upper screen, labeled “old photo,” shows a static image of characters A and B, with arms over each other’s shoulders, both giving a thumbs-up to the camera. The lower screen, labeled “AI restoration,” features a video performance in which actors imitate the scene. In this case, the visual modality emerges as the sole element of analytical significance and becomes the focal point of the entire video. Although the performance lacks clear narrative logic, and the interactions among participants are random and uncertain, all performers consistently display two defining features: slow-motion movements and vacant gazes, as if forgetting their human identity and transforming into malfunctioning machines. The participants are well-known content creators labeled under music and entertainment categories, possessing certain performance and creative skills. While imitators’ abilities and video effects vary greatly in this category, videos intentionally mimicking AI almost invariably adopt the traits of “slow motion” and “vacant gaze.” Moreover, this form requires no props and no elaborate scripting, making it a relatively simple style of creation, and therefore one of the most common. 4.2.3. Category 3: Intrusion of Irrelevant Elements. This category of video relies heavily on introducing elements unrelated to the original video content in order to imitate AI. Unlike the previous two categories, where off- screen elements—such as introducing skewers into a dieting scene or turning two actors into four—still maintained some connection to the original scenario, here the irrelevant elements emerge in unexpected places and in unexpected ways, thereby entering the audience’s field of vision abruptly. Moreover, these inserted elements are generally fixed in form. Representative Example: @Buqi Dance Troupe: “AI: Content suspected to be human-generated”. Video Description: The video runs for a total of 23 seconds. In the first half, a group of five people perform a choreographed dance. In the latter half, each performer introduces an unrelated off-screen element into their act. Compared with the previous two categories, this video exhibits a higher degree of modal density. The shifts in written language modality and auditory modality (music) serve as cues for the audience to recognize the transition from normal performance to AI imitation. The introduction of irrelevant off-screen elements constitutes the densest part of the modal composition, effectively capturing audience attention. Among these, the use of tools as modalities is the most challenging and the most intriguing to viewers. Since the content creator positions himself primarily within the dance genre, the “pre- AI” dance sequence is not discussed in detail here. In the “post-AI” section, however, the second performer from the left removes his trousers to reveal red sports shorts. A retrospective look at the earlier performance reveals that he had consistently stayed in the back row with his hands on his waistband and minimized dance movements, thereby setting up the moment of removing his trousers seamlessly. Similarly, because the performance faces the audience, few noticed the basketball concealed behind him. This category of mimicry demands high levels of skill and coordination, making it difficult for most users to reproduce; consequently, its frequency of occurrence is relatively low. The following category, however, integrates even greater requirements— including performance, coordination, cinematography, narrative design, and creativity—making it rarer still. Yet it epitomizes the higher level of human imitation of AI, encapsulating nearly all the creative features humans employ when mimicking AI. 4.2.4. Category 4: AI Plot Interpretation. Representative Example: @Mu Tianran: AI: “He’s saying all my lines!” Video Description: In the first five seconds, Actor A opens a door and collides with Actor B. Just as a conflict seems about to unfold, soothing background music—contrasting sharply with the tension of the scene—begins to play, and the actors’ movements slow down, shifting into random dance. At the 10-second mark, Actor B exits the frame, and Actor C enters. At this moment, props such as red wine and a large bowl of noodles, irrelevant to the original setting, appear in the hands of Actors A and C. Based on the integration of the single communicative modalities, the 18-second video can be divided into three higher-level action segments that together constitute a 68 coherent meaning structure. From 0–4 seconds, the pre- imitation phase depicts the protagonists initiating a dispute after colliding at the doorway. From 5–7 seconds, the preparation phase for imitation occurs, during which the actors perform in slow motion to mimic the distinctive 15- frame discontinuity typical of AI-generated video effects. From 8–18 seconds, the imitation culminates in an explosive phase, as irrelevant elements suddenly appear in the frame, simulating the illogical intrusion of extraneous factors characteristic of AI-generated content. Although the video lasts only 18 seconds, it incorporates a rich array of multimodal elements. In particular, through visual modalities, the performers skillfully integrate multiple features of grotesque AI-generated videos: sudden narrative reversals, slow-motion enactments, and the abrupt intrusion of unrelated elements. Together, these strategies create a strong sense of “AI authenticity” for viewers, eliciting recognition of the performance as convincingly AI-like. 5. Research Findings Based on case observation and induction, the phenomenon of humans imitating AI is, to a large extent, initiated by a small group of people, and then triggers large-scale imitation as a form of popular culture. Its characteristics can generally be summarized as follows: First, slow-motion mimics AI’s “frame-skipping” slow-motion phenomenon. When large language models generate intermediate frames, the lack of sufficient temporal information can lead to detail loss, and since generating high-resolution videos consumes significant computational resources, models often lower frame rates to balance performance. This passive “frame-skipping” directly causes the action rhythm in AI-generated videos to slow down. Second, “vacant gazes” mimic AI’s inability to understand the meaning of eye expressions. Eye gaze is an important way for humans to convey emotions and intentions, but AI models find it difficult to truly grasp the complexity of human emotions and intentions. As a result, they struggle to capture the appropriate gaze states in specific contexts, leading to unstable and incoherent impressions. Third, the intrusion of irrelevant elements mimics AI’s generation of new objects. In many AI-generated videos, whether due to model flaws or contradictory prompts, AI often produces irrelevant new elements. Humans take advantage of this phenomenon by deliberately adding new elements into videos in various ways, thereby creating a foreground modality strong enough to capture the audience’s attention. Fourth, narrative reversals mimic AI’s alteration of endings. In intentionally “grotesque” AI-generated videos, it is common for characters on the verge of conflict in film or television scenes to suddenly reconcile, making “kissing” or “embracing” fixed features of such AI- generated content. Consequently, in human reverse imitation of AI, this narrative device is often reproduced. Researcher observations of comment sections reveal that many users jokingly remark, “a real AI would already have them kissing,” treating whether or not characters kiss or embrace as one of the main criteria for judging whether a video is AI-generated. 5.1. Imitation and the Reconfiguration of Human–Machine Relations In today’s era of intelligent media, human imitation is no longer confined to reproducing nature or existing human inventions; rather, it now manifests in reciprocal acts of imitation and co-creation between humans and the intelligent algorithms they themselves have created. In earlier studies on human–machine collaboration, artificial intelligence was often regarded as a mere tool assisting human creativity. As AI’s influence expands across multiple domains of life, however, the notion of “human subjectivity, AI instrumentality” has been increasingly questioned. The relationship between humans and AI, through processes of deep entanglement, reveals new trajectories. Human reverse imitation of AI illustrates that intelligent algorithms have risen beyond the instrumental level of creative collaboration to a symbolic or even spiritual dimension: AI becomes a source of mimetic concepts for humans, rather than merely a generator of images upon command. Within such multi- layered entanglements between humans and algorithms, the key question arises: can humans discover new meanings of human–machine relations? In Human-Machine Reconfigurations [14], Lucy Suchman critiques traditional cognitive science and AI research for conceiving of human thought in computational terms and assuming human reasoning to be inherently plan-based. She emphasizes that the “essence” of both humans and machines is not fixed, but dynamically performed within specific practices and interactions. This performativity is the core of Suchman’s argument. Accordingly, research on human– machine relations should move away from essentialist questions—such as whether machines can possess intelligence like humans or whether they can hold equivalent status—and instead pursue more practical inquiries: how, in concrete social settings and interactions, are humans and machines performed as similar or different? What theoretical, practical, or political consequences arise from the construction of such “similarities” or “differences”? Donna Haraway’s Cyborg Manifesto [15] further situates human– machine relations within an anti-essentialist ontology: there is no eternal, unchanging “human nature,” but only “temporary formations” continuously constructed within networks of technology, culture, and power. When humans deliberately perform “non-human” traits, they are not “losing their humanity,” but demonstrating that humanity itself already contains an openness to the “non-human.” This openness is not forced, but rather a conscious choice—by rejecting the fixation that “humanity must stand in opposition to machines,” humans inject new possibilities into the very concept of humanity: it can be rational or mechanical; emotional or algorithmic; coherent or fragmented. “Humanity” is no longer a defined essence, but a process continuously created through entanglement with technology. Today’s satirical human imitations of AI may, in hindsight, be dismissed as absurd or trivial. Yet, in their historical moment, they serve as a form of self-reflection for humans and contribute to the iterative refinement of artificial intelligence, supplementing and co-constructing the evolving relationship between the two. 5.2. Imitation as a Catalyst for Human Self- Reflection The phenomenon of humans imitating AI is not itself an inevitable trajectory of historical progress; rather, it surged into popularity and faded from public view within just half a year. In terms of its habitual modal actions and the use of classic background music, this form of creative behavior quickly became templated, formulaic, and virally disseminated—precisely in line with the logic of meme propagation in the internet era. “Human reverse imitation of 69 AI,” as a meme type requiring low cultural and technical literacy as well as minimal costs of imitation, has become a “traffic shortcut” for mass participation. By mimicking memes already validated through massive viral circulation, short-video creators can engage in repetitive production, thereby reducing creative risks while ensuring returns. In the digital age where attention has become a scarce resource, trending topics act as “traffic entry points” that rapidly aggregate attention. Consequently, they serve as low-cost and high-reward anchors of content production, enabling individuals to achieve exponential growth in visibility within a short time, and to convert attention into economic value. Creators strategically integrate such playful, absurd imitative performances into their works in hopes of achieving rapid popularity. As some scholars have noted, the phenomenon- level performance models that periodically appear on short- video platforms are continuously spliced, replicated, and virally spread; in this process, people cease to care about or remain attached to the original meaning of a given model. Instead, the pleasures, novelty, and shock effects generated by hyperreal imitations become the primary motivational force for “attention capture” [16]. This dual quality of being both topical and surprising enabled the imitation of AI behaviors to spread rapidly across the nation like wildfire. Yet, the viral success of such videos inevitably reflects deeper social concerns. Through participatory observation of comment sections, researchers discerned that in this era of rapid AI development, whether AI might replace humans has become a pressing question for many. Motivated by ambivalent feelings toward technology and a desire to disrupt the perceived threat of AI, creators highlight AI’s deficiencies through mimicry in order to reaffirm human originality. Although such videos are created by humans, platforms often label them as “content suspected to be AI-generated,” thereby provoking a reassertion of human creativity. This in turn reassures many viewers that AI can never truly replace humanity. 6. Conclusion Compared with one year ago, today’s artificial intelligence technology has undergone significant iterations, and the application of generative AI in creative fields continues to increase day by day. Many of the unreasonable phenomena once caused by algorithmic flaws have now been largely corrected, and the level of technology it can achieve is already capable of nearly perfect replication of reality. The coupling between humans and AI will become even closer, and questions concerning human–machine relations should always be placed at the center of scholarly consideration. Finally, this study also contains the following limitations: it merely provides a summary analysis of the features of the imitation phenomenon without probing into its underlying causes; the classification of categories is not strictly defined, but only a preliminary generalization derived from the observation of fifty selected cases; and the application of Norris’s multimodal analysis remains at a superficial level, without a comprehensive grasp of the framework. Future research will need to collect additional materials in order to further enrich the theory. Given that AI will remain a focal point of human attention for a long time to come, it is hoped that this article may serve as a point of departure and provide inspiration for subsequent researchers. References [1] Zheng Guocheng and Yuan Aiqing. Humor, Imitation, and Deconstruction of Generative AI in the Age of Smart Media— —Taking "Human Imitating AI" short videos as an example. Journal of Shanxi Youth Vocational College, 2025,38(01):7-12. [2] Du Botong. Imitation and Carnival: Analysis of the "Reverse AI Challenge" Video Phenomenon on TikTok Platform. Audiovisual, 2025, (15):92-96. [3] Norris, S. 2004. Analyzing Multimodal Interaction: A Methodological Framework. London: Routledge. [4] Zhong Jianqin. Human-machine collaborative content production mode. Youth Journalist, 2020(36):44-45. [5] Zhou Li and Chen Mu 'en. "Encountering the' Algorithmic Time Machine': The Technological Integration of AI-ized Memory and Subject Evasion". Journal of Journalism and Communication Studies, 2024,31(01):67-82+127. [6] Ding Fangzhou, Zhang Duo. "Is AI Just Your Tool?" Research on the Human-AI Relationship between Short Video Bloggers and AI in the--Intelligent Communication Ecosystem. Journal of Journalism University, 2025(01):1-14+119. [7] Jiang Qiaolei and Shi Mengsu. The Significance Clash in the Intelligent Era: Human-Machine Relationships and Meaning Perception in Collaborative Creativity. Journal of Journalism University, 2025(02):1-14+117. [8] Schmitt, B.Speciesism: An Obstacle to AI and Robot Adoption.Marketing Letters,2020, 31(1),3-6. [9] Ni Wan and Guo Jialiang. The Essence and Value of AI Artistic Creation from the Perspective of Mediaization. Modern Communication, Journal of Communication University of China, 2024,46(08):95-103. [10] Li Yabin and Liu Qingjun. A Preliminary Discussion on Imitation Theory from Ancient Greece to the Eighteenth Century. Art Observation, 2023(12):77-78. [11] Zhan Di and Fang Jieyun. Sensory synesthesia in digital audiovisual production under the perspective of "Technology- Culture Evolution Theory". Editorial Friends, 2025(02):56-63. [12] Sun Yuting. A Brief Analysis of Imitation Behavior in TikTok Short Video Spectacle. New Media Research, 2020,6(04):90- 91. [13] Zhang Tingyu. Initial Exploration on the Development of TikTok Platforms in the New Media Context. Journal of Journalism and Communication, 2023(05):46-48. [14] Suchman, Lucy. Human-Machine Reconfigurations. Cambridge: Cambridge University Press, 2007. [15] Haraway, Donna." A Cyborg Manifesto: Science, Technology, and Socialist - Feminism in the Late Twentieth Century." In Simians,Cyborgs and Women: The Reinvention of Nature.New York:Routledge,1991,149 - 181. [16] Zhan Di, Fang Jieyun. The Cyborg’s Mask: The Generative Mechanism of Collective Performativity in Short Video Production. Modern Communication (Journal of Communication University of China), 2025, 47 (01): 117-124.