Young-ho Kim and Jeong-gi Kim’s study, “Shifting Perceptual Thresholds in the Acceptance of Generative AI-Based Content,” identifies an intriguing contradiction in recent debates surrounding generative AI content. Image- and video-generation technologies have advanced dramatically in only a few years. In the 2023 “Will Smith Eating Spaghetti” video, faces merged with food and bodies collapsed into impossible forms. Today, however, AI can produce results that even experts may struggle to identify immediately as machine-generated. One might expect resistance to diminish as technical defects become less visible. In practice, however, controversy has become more frequent as AI is integrated more seamlessly into films, advertisements, music, and games.

The paper argues that this occurs because the criteria by which people judge AI content shift as the technology develops. At first, the central question was whether the content looked real. Once the technology passed that threshold, the more important question became who audiences believed had created it—in other words, whether they trusted the attributed creator. Once that trust was stabilized, attention shifted again toward the rules and contexts under which the technology should be accepted. The paper explains this movement through three stages: the sensory threshold, the attributional threshold, and the playful threshold.

The First Threshold: The Body Rejects Before the Mind Interprets

The sensory threshold, the first stage, operates most rapidly. Before viewers consider the ethics of AI use or the intentions of a creator, they first determine whether the objects and bodies on screen obey the same physical rules. If a hand has the wrong number of fingers, a face suddenly changes shape, or a pair of glasses disappears and reappears from one shot to the next, immersion breaks immediately. The judgment at this stage is not “Is this good AI or bad AI?” but rather “Does this scene remain coherent as a single reality?”

The Will Smith video, a McDonald’s Netherlands advertisement, the opening sequence of Marvel’s Secret Invasion, and Toys “R” Us’s Sora advertisement—all analyzed in the paper—are rejected at this stage. The disappearance and reappearance of glasses in the Toys “R” Us advertisement is especially revealing. It suggests that continuity between shots may matter more than the aesthetic quality of individual frames. A single image may appear convincing, but if viewers can no longer believe that the preceding and following shots belong to the same world, the entire video collapses.

One of the paper’s important findings is that transparency cannot compensate for sensory failure. Valentino disclosed its use of AI collaboration, yet still faced rejection because of bodily distortions and unnatural compositing. The ethical act of openly acknowledging AI use and the sensory quality of the resulting work are judged on different levels. Moreover, the threshold becomes higher in fields such as luxury fashion and professional orchestras, where precision and mastery are central values. An error that might be dismissed as amusing in an individual’s experimental video can be interpreted as an affront to professional dignity when it depicts a musician’s hands as deformed. A technical defect does not remain confined to the screen; it spreads to the perceived identity of the people and institutions represented.

The Second Threshold: The Better the Result, the Greater the Sense of Betrayal

Once content passes the sensory threshold, the nature of the problem changes. When an image or performance appears sufficiently natural, audiences assume that there is a human actor, singer, director, or designer behind it. The paper calls this an “attribution structure.” We do not perceive only the formal qualities of a work; we also read it through the assumption that someone acted with a particular intention. We believe that an actor’s pronunciation reflects practice and performance, that a singer’s voice carries a body and a life history, and that documentary footage was recorded by a camera witnessing an actual event.

When the use of AI is revealed afterward, the pixels and sound quality remain unchanged, yet the evaluation of the work can be reversed. What audiences reconsider is not the quality of the finished product, but the interpretive relationship they had already formed with it. Examples include the AI recreation of the late actor Ian Holm’s appearance and voice in Alien: Romulus, and the use of AI to adjust an actor’s pronunciation in The Brutalist. Even when the production team did not actively intend to deceive, viewers must retrospectively revise an experience they had understood as a human performance.

The paper distinguishes intentional deception from unintentional nondisclosure in terms of production motives, but argues that the two can converge in the audience’s experience. A robocall that cloned Joe Biden’s voice with the explicit intention of deceiving listeners does not carry the same ethical weight as a film that simply failed to disclose one of its production tools. From the audience’s perspective, however, both involve the collapse of an attribution that had already been formed. A later explanation stating that “AI was used only in a limited capacity” or that “legal consent was obtained” cannot restore the initial experience.

Paradoxically, the more technically accomplished the content is, the greater the resulting sense of betrayal may become. Crude AI content appears machine-generated from the outset, so viewers never form a strong belief that it was created by a human. By contrast, highly polished content allows audiences to attribute human action and intention with greater confidence. When AI use is later revealed, more of those assumptions must be withdrawn. Technological success thus strengthens the conditions for attributional betrayal.

Transparency Is Not a Disclaimer but the Arrangement of Viewing Conditions

In this study, transparency is not reduced to the administrative act of attaching an “AI-generated” label to a piece of content. The timing and manner of disclosure function as mechanisms that alter the order in which audiences interpret a work. People do not first encounter content and then simply add information afterward. They organize everything that follows according to the information they receive first. Disclosing AI use after a work has been experienced and disclosing it beforehand are therefore not equivalent acts, even when the factual information is identical.

Post-viewing disclosure dismantles an already established assumption of human creation. Advance disclosure, by contrast, prevents that false assumption from forming in the first place. Audiences enter the work knowing that AI has been involved and begin to explore which elements resulted from human decisions and which emerged through machine generation. The central question shifts from “Why did they pretend this was human-made?” to “What did the human creator use this technology to accomplish?”

The practical function of transparency, then, is to establish in advance the rules under which creators and audiences encounter one another. When disclosure comes late, it becomes an explanation. When it comes first, it becomes the frame of the work. This temporal difference can separate deception from experimentation, even when the same technology is used.

The Third Threshold: Enjoying It Because We Know It Is AI

The playful threshold, the third stage, is where the paper’s perspective becomes most distinct. Here, “play” does not refer only to lighthearted or humorous content. It describes a mode of engagement in which audiences voluntarily participate in a new set of rules while remaining aware that AI is involved. Serious documentaries and horror games can also be included within playful acceptance in this sense.

In the game MIMESIS, an AI player imitates humans and deceives other players. The same behavior would be considered deceptive in everyday life or politics, but in the game, identifying the AI is introduced in advance as a central rule. The uncertainty of attribution is not removed; instead, it is preserved as a source of tension and enjoyment. Deception has not disappeared. Its boundaries and purpose have been disclosed.

The documentary Another Body uses deepfake technology to protect victims of deepfake abuse and reconstruct their experiences. At the technical level, the same tool is used both to produce harm and to protect those who have been harmed. Because the filmmakers disclose their intentions and methods in advance, however, the AI is attributed not as a device for falsifying reality but as a means of enabling testimony. The meaning of a technology is not fixed within its function. It changes according to who uses it and under what conditions.

The Beatles’ “Now and Then” offers another clear example. AI was not presented as an entity that generated a new John Lennon voice, but as a restoration tool that separated his existing voice from the noise of an old recording. By revealing both the production process and the manner in which the technology was used, the role of creative agency was not transferred to the AI. Grimes’s Elf.Tech takes the opposite approach, making an AI version of her voice available to fans and encouraging collaborative creation. One preserves attribution, while the other opens it up. Their methods differ, but both establish the rules before participation begins.

The tendency of Neuro-sama’s fans to attribute personality and emotion to the character despite knowing that she is an AI follows the same structure. People do not participate because they mistake the AI for a human. They temporarily grant it personhood while continuing to recognize it as AI, and they enjoy the unpredictability of its behavior as content. The experience does not arise from erasing the boundary between humans and machines, but from knowingly moving back and forth across it.

Platforms Prepare Attitudes of Acceptance Before the Technology Appears

Alongside transparency, the paper identifies the context of reception as an important variable. The same AI-generated synthesis may be understood as a violation of factual integrity when used in news, but accepted as a rule when used in a game. In advertising, audiences judge it according to the quality standards and labor ethics historically promised by the brand. In a museum, visitors may arrive with the prior expectation that they will encounter experimental technologies.

Artechouse’s AI exhibitions organize the exhibition space, titles, and explanatory system around the premise of collaboration between humans and AI. Rather than questioning only afterward whether AI was used, visitors examine what kinds of sensory experience the collaboration produces. A platform is not a neutral container for content. It is a mechanism that preconditions what audiences should question and what they may be willing to permit.

This perspective moves beyond the familiar division of audiences into supporters and opponents of AI. It is not contradictory for one person to reject AI-generated advertising while enjoying an AI-based game, or to condemn criminal deepfakes while supporting a documentary that uses deepfake technology. Each form of content demands a different type of trust and establishes different rules of participation. Audiences do not necessarily maintain a consistent attitude toward AI itself. Rather, they judge the kind of relationship that the technology is attempting to establish with them.

What Must Now Be Designed Is Not Only the Result, but the Order of Disclosure

The central question reframed by this paper is no longer “Can AI create content equivalent to that made by humans?” but “What kind of relationship was assumed when this content was presented?” As generative technology advances, the sensory threshold recedes, but judgment does not disappear. It moves from the surface of the image toward relationships among creators, institutions, and platforms.

The task of producing AI content, therefore, is not a matter of choosing between quality and transparency. Sensory coherence is the condition for passing the first threshold. Advance disclosure stabilizes the structure of attribution. The purpose of use and the platform context provide audiences with a reason to experience the technology. None of these can substitute for the others.