Why Does AI Art Create That 'Uncanny Feeling'? - When the 'Author' Disappears from the Work

While browsing social media, I sometimes suddenly come across a beautiful picture.

The evening light grazes the contours of a cheek. The folds of the clothing, the reflections on the wet road surface, and the expression of a slightly tired person are all well-rendered. I stop scrolling to look at the details. However, the moment I notice the words 'AI-generated' at the end of the post, the image that was charming until then suddenly looks thin.

Of course, not everyone feels this way. Some people don't care if they know it was made by AI, and some are even impressed by the technology. Still, the experience of the same image looking different between 'when I thought a human drew it' and 'after I learned AI made it' is no longer rare.

This is a strange story. Not a single pixel of the image has changed. The colors, composition, and the person's face remain exactly the same. The only thing that has changed is the information about where that image came from.

If that is the case, we were not just looking at the appearance of the picture from the beginning.

The reason AI art is scary is not because AI is good at it. It is because the circuit that should have connected the picture to someone behind it is severed there.

The moment you know it is 'AI-made,' the picture changes

There are already several empirical clues regarding this sensation.

Harsha Gangadharbatla investigated how information about AI production affects the evaluation of works. In a 2023 study by Lucas Bellaiche et al., people also showed a tendency to rate works labeled as AI-produced lower than those labeled as human-produced. Differences were particularly likely to arise in evaluations that infer the mind or efforts behind the work, such as 'depth,' 'value,' and 'creativity.'

The important point is that this is not simply a story about AI images being visually inferior. Even when people cannot distinguish between human and AI works, or when different author labels are attached to the same work, information about the origin influences the judgment.

In 2025, a study was published that displayed the same painting as 'human-made' or 'AI-made' and examined reactions using EEG and near-infrared spectroscopy. In experiments involving a total of 125 people across two pre-registered studies, works labeled as human-made were rated higher, and differences were also seen in neural responses interpreted as being related to attention and semantic processing. However, one cannot conclude from the differences in brain activity that 'AI art has no meaning.' What can be said is that knowledge about the author seems to enter into the very way we appreciate art.

Art appreciation is not the task of grading stimuli that reach the retina. The viewer is looking for someone on the other side of the picture—someone who chose, someone who struggled, someone who failed, someone who tried to convey something.

That is why if the origin changes, the meaning of the same lines and colors changes as well.

'It has no soul' is not such an unscientific phrase

A phrase frequently used in criticism of AI art is 'it has no soul.'

It sounds a bit exaggerated and even mystical. If one starts saying that human works always have a soul and AI works absolutely do not, the discussion quickly becomes like a religious trial. There are human paintings where the artist cut corners. Conversely, there are people who use AI by adjusting it repeatedly, adding brushstrokes, and entrusting it with their own experiences.

Even so, it is a pity to discard this phrase as mere sentimentality. Because when you break down 'it has no soul,' quite concrete elements emerge.

Psychologists Heather Gray, Kurt Gray, Daniel Wegner, and others organized how people perceive the 'mind' of others mainly along two axes. Agency, which is the ability to think, plan, and control actions, and experience, which is the ability to feel pain, joy, fear, and so on.

Current generative AI is in a strange position for many people. In the sense that it responds to complex instructions and produces elaborate images, it appears to have high agency. However, it has not experienced the depicted sadness itself. It does not age, it does not suffer heartbreak, and it does not feel like throwing up before a deadline. In other words, while it appears to have high capability, almost no lived experience is attributed to it.

This combination creates an eeriness.

On the screen, there are 'traces of an attempt to draw a sad face.' However, the subject who knows that sadness cannot be found. Although emotions arise in the viewer, there is no one on the other side who intended to deliver those emotions.

AI art is somewhat like a letter without a sender. The writing is beautiful and touches our hearts. But when you finish reading it and turn the envelope over, there is no name. Strictly speaking, it is not that there is no name. There is the person who wrote the prompt, the person who designed the model, and the people who drew the works used for training. It is just that it is unclear whose voice we should receive it as.

What people have called a 'soul' may not have been a supernatural substance, but rather a continuity that allows one to trace back from the work to the author's body, experience, intentions, and responsibility.

Why is it seen as 'cheating' rather than 'taking it easy'?

The reaction to AI art is not just an aesthetic judgment, but also a moral one.

In psychology, there is a concept called the 'effort heuristic.' As Justin Kruger and others have shown, when people find it difficult to judge the value of an object directly, they use the effort they imagine was put into it as a clue. Even for the same work, it is sometimes rated higher if it is explained that it took a long time to create.

However, the anger surrounding AI art cannot be fully explained by simply saying 'it has low value because the production time is short.' Tools have originally been created to reduce effort. Using a pen that is more convenient than a brush, a digital camera that eliminates the need for a darkroom, or Photoshop that makes corrections easier has not always been considered dishonest in itself.

AI art looks like 'cheating' because there is such a huge gap between the effort saved and the results obtained.

A person who has been sketching for years and a person who started generating images yesterday line up images of similar quality on a timeline. When competing for limited resources such as contests, commissions, praise, and revenue, that shortcut looks not like mere efficiency, but like a change in the rules of the game.

Here, it is not just the amount of effort that is being questioned, but 'who offered what.'

To put it a bit boldly, art has long existed within an 'economy of sacrifice.' The creator offers time, body, life, and the risk of failure. The viewer receives that irreplaceable expenditure as the weight of the work. This is a hypothesis that goes beyond the effort heuristic rather than an established psychological law. However, it explains well why no matter how much you increase the AI's computation time, it is difficult to feel it has the same value as human effort.

Even if a GPU runs for 10,000 hours, the GPU does not feel that it has lost 10,000 hours. A human's 10,000 hours are 10,000 hours that the person could have spent living another life.

Of course, there are cases where AI users themselves spend a lot of time. They change prompts, adjust models, discard hundreds of images, and add touch-ups. One should not conclude that there is no creativity in that act. However, that effort is not visible from the outside, and the relationship between each line and color that appears in the final image and the user's physical judgment is not as clear as in traditional drawing. Therefore, simply explaining 'I worked hard' does not dispel the doubt.

What people are feeling is likely a sense that there is unpaid effort. Something that should have been paid to obtain the result remains unpaid. If we call that 'effort debt,' the anger toward AI art can be understood as a retributive emotion toward those who receive only praise while carrying that debt.

However, this emotion is not always correct. More effort does not necessarily mean a better work, and a society that glorifies suffering excludes the weak. AI provides new means of expression to those who could not create art due to physical circumstances or lack of educational opportunities. That cannot be ignored.

What we should be asking is not whether they took it easy, but who is bearing the cost of the spared effort elsewhere.

'Plagiarism' is not a legal term, but a compressed moral vocabulary

When the word 'plagiarism' is used regarding AI art, the discussion often becomes solely about copyright law.

However, legal copyright infringement, similarity to specific works, imitation of style, unauthorized collection for training data, lack of credit, and lack of compensation are not the same thing. Whether a certain training act is legally permissible depends on the country, the purpose of use, and the specific generated results. It is crude to immediately conclude 'it is theft because it learned' or to dismiss it by saying 'there is no problem because it does not store copies.'

Even so, there is a reason why people feel that something has been 'stolen'.

Research on procedural justice has shown that people value not only the fairness of the outcome but also the legitimacy of the process that leads to that outcome. Was there an opportunity to have one's voice heard? Were the rules consistent? Was there an explanation? Could one lodge an objection? Was one respected?

This is precisely the part that many creators find problematic regarding the training of generative AI. Publishing a work is not the same as consenting to the training of a model that could potentially replace the author's work in the future. Moreover, it is difficult to see which works contributed to which model, and to what extent. While the original authors receive neither credit nor compensation, the companies that own the models and the users reap the benefits.

The word 'plagiarism' compresses these multiple grievances into a single term. It is not just image files that have been taken. It is the feeling that one's past work has been transformed, without one's knowledge, into a device that competes with one's future work.

This creates a unique kind of pain specific to generative AI that did not exist with traditional mechanization. It is not just that machines are replacing jobs. The results that one has published for years are also becoming resources that enhance the performance of the replacement. The past self is being incorporated into a machine that drives out the future self.

It is not that there is no author. It is that there are too many authors, and they are invisible.

Here, sociologist Howard S. Becker's 'Art Worlds' is useful.

Becker did not think of works of art as pure creations of a solitary genius. Works are established through the collaboration of many, including those who make paints and instruments, those who exhibit, criticize, transport, set up the institutions, and even the audience. In the first place, there is almost no such thing as a completely solitary author in human art.

If we adopt this perspective, the explanation that 'there is no author in AI art' is not accurate.

There are people who created the countless images that served as the training data. There are people who built the datasets. There are people who added tags, designed the models, maintained the computing environment, entered the prompts, and selected and corrected the generated results.

The problem is that those numerous contributions are not distributed in a visible way. Movies and anime are also collective productions, but at least they have end credits. They attempt to record, even if not perfectly, who directed, who filmed, and who was in charge of the art. In generative AI, the learned contributions dissolve into the weights of the model, and the path from the work back to the individual creators almost disappears.

The convenient phrase 'AI drew this' actually obscures a massive network of labor. What we are feeling may not be the absence of an author, but the invisibilization of the authors.

And responsibility is diluted in the same way. When a problematic expression appears, it is dispersed among the training data, the design of the development company, the user's prompt, or the judgment of the person who published the output. Although there are many involved, no one can be found to take responsibility for the whole.

Signing a work is not only about receiving credit, but also about taking on the place to answer for criticism. Authorship in the AI era needs to be redefined not by who did the most work, but by who took on the final choice and responsibility.

Whose power does the 'democratization of creation' strengthen?

AI image generation does indeed have an aspect of democratization.

Even people without the skill to draw can visualize the images in their minds. Small organizations and individuals can attempt expressions that were previously impossible due to budget constraints. For people with physical limitations, being able to create images from words can be a vital form of freedom.

If existing experts deny this possibility to protect their own status, it becomes mere gatekeeping.

To borrow Pierre Bourdieu's words, the art world is also a 'field,' and the skills, educational background, aesthetic sense, and career that are valid there function as cultural capital. Part of the backlash against AI is likely a defense against the devaluation of capital that has been accumulated over many years. In a world where anyone can create plausible-looking images, the superiority of not only those who can draw but also those who have considered themselves 'able to distinguish the real thing' is shaken.

However, it is also premature to conclude that all backlash against AI is merely about vested interests.

When old cultural capital weakens, it does not mean that power disappears. Access to high-performance models, computational resources, the cost of maintaining paid services, the knowledge to fine-tune models, and the aesthetic eye to select good results from massive outputs become the new cultural capital. And the position of greatest power is often occupied not by individual users, but by the platform companies that own the models and infrastructure.

From the perspective of the autonomy of the field of cultural production discussed by Bourdieu, the problem is not just that artists are excluding new entrants. It is that the value criteria of art and illustration themselves are becoming subordinate to the logic of economics, such as click-through rates, generation speed, usage fees, and the terms of service of model companies.

The democratization of creation and the centralization of cultural resources can happen simultaneously.

'Production without origin' that followed reproduction technology

When thinking about AI art, Walter Benjamin's concept of 'aura' is unavoidable.

Benjamin believed that the uniqueness of the 'here and now' that a work of art possessed was fading due to reproduction technologies like photography and film. However, he did not lament this as a mere degradation of culture. Reproduction also had the potential to liberate works from their privileged locations and open them up to many people.

Generative AI pushes that argument into even stranger territory.

In a photograph, there is something that was captured. Even if a photograph is manipulated, the form of photography has historically presupposed a causal contact between light and the subject. A reproduction also has an original work that was copied. But in an AI image, there is neither a single event corresponding to the depicted scene nor an original work that was reproduced.

What exists are patterns extracted from vast amounts of images, prompts, and a probabilistic generation process.

This is not so much a state where the aura has been lost, but rather a state where the origin that should house the aura is dispersed from the beginning. If we call this a 'reverse aura,' the problem is not that a single original was reproduced, but that it has become possible to create an appearance that looks like an original without a single origin.

As a result, what might hold value from now on is not the scarcity of the image, but the scarcity of its provenance.

Who, when, what tools were used, to what extent was AI involved, what was selected, and what was corrected? The reason technologies that record the provenance of digital content, such as C2PA, are attracting attention is not just for anti-fake measures. As images are generated infinitely, the causal history of 'this was created by this person through this process' itself becomes the value.

There is a slightly ironic future here.

While AI turns creation into cheap daily commodities, 'works that can prove they were made by humans' may become expensive crafts. Just as handmade bread, film photography, and artisan pottery became premium items in the age of mass production, human clumsiness and the traces of time will be re-commodified as luxury goods.

We could call this 'provenance capitalism.' Technology that promised the democratization of creation is turning proven humanity into a new scarce resource. Everyone can own an image, but not everyone can buy 'human time.'

Is AI criticism Luddism?

Some people call the backlash against AI 'modern-day Luddism.' It means people who are afraid of machines and try to stop progress.

However, the historical Luddite movement was not simply about hating machines. The textile workers who destroyed machines in early 19th-century England were resisting not the existence of technology itself, but the factory owners' use of machines to unilaterally lower wages, quality, skills, and community norms.

In that sense, calling the backlash against AI art Luddite is possible. However, that is not an insult.

What they are asking is not whether or not to use machines, but rather who benefits from the introduction of technology, who loses their bargaining power, and whose convenience dictates the rewriting of the value of certain jobs.

When photography appeared, and when Photoshop became widespread, there were voices saying it was 'not art' or 'cheating.' Therefore, some say that criticism of AI will eventually disappear. There is some truth to this comparison. Once a technology becomes established, new skills and forms of expression emerge, and social evaluation changes.

However, we cannot assume that the same history will repeat itself.

The camera did not read a massive amount of a portrait painter's past work and output countless portraits in that painter's style. Photoshop helped the human drawing the lines make decisions, but it did not learn from a body of work to propose composition, lines, and colors all at once. Generative AI blurs the boundary between tool and author more deeply than ever before, and incorporates past cultural achievements into the means of production.

What the resistance to AI art is questioning is not whether we progress or regress. It is whether, when rewriting the cultural contract through technology, we have negotiated with the people who were there.

What we sought in art

Thinking this far, it becomes impossible to praise all discomfort toward AI art as justice, or to mock it all as ignorance.

Within it, there is a sense of privilege that only humans should be creative. There is also an exclusivity that only those who have trained for a long time have the right to create works. The judgment that human works are always deep and AI works are always empty can also become a form of label discrimination that decides the conclusion before even looking at the work.

On the other hand, the optimism that everyone will be free if they use AI only sees half of the reality. The distrust of a system that sucks up someone's work as a resource, makes its origin invisible, and separates responsibility from profit is not technophobia.

For AI art to be accepted, it is not enough for it to simply get better.

What kind of data was it made with? Could the creators refuse its use? Where did humans make choices? Who gets the credit and the profit? When a problem occurs, who explains it and takes responsibility? We need to make those chains of causality visible once again.

And on the user's side, there are things that should be accepted if one calls oneself an author. Not just the fact that you wrote a prompt, but why you chose this image, what you are expressing, how much you depend on the resources of others, and how you respond to criticism after publication. When you accept that responsibility, AI can become not just an automatic generation device, but a tool incorporated into the human expression process.

That there is someone on the other side of the work who lived a different time than oneself. That there is a person who chose lines I did not choose, saw a world I could not see, and offered something even though they might fail. We looked at art to touch those traces.

What is scary about AI art is not that there are no human traces.

The beauty is there. But one cannot see whose time that beauty came from. What generative AI has broken is not art itself. It is the chain of causality that connected work and author, effort and evaluation, citation and consent, expression and responsibility.

Whether we can reconnect that chain in a new form. What determines art in the AI era is probably that, rather than the performance of the model.

Main References

Becker, Howard S. Art Worlds. University of California Press, 1982.

Benjamin, Walter. "The Work of Art in the Age of Mechanical Reproduction," 1935-1936.

Bourdieu, Pierre. The Field of Cultural Production. Columbia University Press, 1993.

Bellaiche, Lucas, et al. “Humans versus AI: Whether and Why We Prefer Human-Created Compared to AI-Created Artwork.” Cognitive Research: Principles and Implications, 8, 42, 2023. https://doi.org/10.1186/s41235-023-00499-6

Gangadharbatla, Harsha. “The Role of AI Attribution Knowledge in the Evaluation of Artwork.” Empirical Studies of the Arts, 40(2), 125-142, 2022. https://doi.org/10.1177/0276237421994697

Gray, Heather M., Kurt Gray, and Daniel M. Wegner. “Dimensions of Mind Perception.” Science, 315(5812), 619, 2007. https://doi.org/10.1126/science.1134475

Kruger, Justin, et al. “The Effort Heuristic.” Journal of Experimental Social Psychology, 40(1), 91-98, 2004. https://doi.org/10.1016/S0022-1031(03)00065-9

Kirk, Colleen P., and Julian Givi. “The AI-Authorship Effect: Understanding Authenticity, Moral Disgust, and Consumer Responses to AI-Generated Marketing Communications.” Journal of Business Research, 186, 114984, 2025. https://doi.org/10.1016/j.jbusres.2024.114984

Zhang, Wenyu, et al. “Neural Correlates of Evaluative Bias against Artificial Intelligence-Labeled versus Human-Labeled Artworks.” Social Cognitive and Affective Neuroscience, 20(1), nsaf071, 2025. https://doi.org/10.1093/scan/nsaf071

van Hees, Jules, et al. “Human Perception of Art in the Age of Artificial Intelligence.” Frontiers in Psychology, 15, 1497469, 2025. https://doi.org/10.3389/fpsyg.2024.1497469

Epstein, Ziv, et al. “Who Gets Credit for AI-Generated Art?” iScience, 23(9), 101515, 2020. https://doi.org/10.1016/j.isci.2020.101515

Tyler, Tom R. Why People Obey the Law. Princeton University Press, 1990.

C2PA. “C2PA Technical Specifications.” Coalition for Content Provenance and Authenticity. https://spec.c2pa.org/

Agency for Cultural Affairs, "On the Approach to AI and Copyright," 2024.