AI Art: Why the Label Changes How We Judge an Image

Imagine this conversation. Someone is looking at an image. They like the light, pause over a figure and ask what it means. Then they hear: “I made it with AI.” Suddenly, the enthusiasm fades.

The image has not changed. What they know about how it was made has.

That moment interests me. How much are we judging the work itself, and how much the story we tell ourselves about its maker, their effort and their tools? AI has made the question louder. It also touches a broader argument about digital creativity and the dismissive “Oh, so it was done on a computer.”

An outstretched hand beneath a floating, dissolving form against grey — Illustration by the author. A hand beneath a suspended form. This image is not from the studies discussed.

What does the “AI” label do to an image?

Lucas Bellaiche and colleagues showed participants AI-generated images attributed either to a human or to AI. Images labelled as human-made received more favourable evaluations. All the images had the same actual origin; what differed was the information given to viewers. ^[1]

Another study, comprising six experiments with 2,965 participants, also found lower ratings across several dimensions when artworks were labelled as AI-made. This happened even when participants reported that they would not have been able to identify the images' origin without labels. ^[2]

This does not prove that everyone who dislikes AI is a hypocrite. It shows that authorship information helps shape an evaluation. You can like an image while objecting to the circumstances of its creation. It is worth knowing which of those things you are judging.

“The computer did it” and imagined effort

One possible mechanism is the effort heuristic. In studies by Justin Kruger and colleagues, participants evaluated works, including a painting, more favourably when they believed that producing them had required more time and effort. Perceived effort served as a cue for quality. ^[3]

This is not a law that always holds. Replication attempts published by Ziano and colleagues in 2023 produced mixed results. Some relationships were reproduced and others were not. The effect depends on context and on what is being measured. ^[4]

In a conversation about generated images, it is easy to imagine the shortcut: if the result appeared quickly, the person cannot have done much. Yet rendering time tells us little about someone's understanding of light, composition and colour, or their ability to recognise a work worth developing.

I do not pretend that entering a prompt is the same as painting an image by hand. During generation, I am not making brushstrokes or modelling every part. My contribution centres on mental work: imagination, visual knowledge, choices and revisions. This does not describe all digital art. Digital painting, sculpting and animation can demand extensive manual skill.

What am I doing when I create with AI?

A short prompt does not necessarily mean little thought. In my recent work, a dark mouse with negative space or a dog, barking, close up, negative space around was enough. Those words do not reveal why I left so much empty background or why I chose that particular result.

Further decisions followed. Should the text occupy the top of the image or remain a small caption? Should its colour compete with the background? After adding a word, am I still interested in the same work?

Not every generation involves that reflection. You can accept the first result because it looks attractive. You can also spend a long time on something that remains banal. Neither hours spent nor iteration counts settle a work's value.

Technology is a tool, but it is not transparent. It suggests solutions and can steer attention towards what is easiest to get from it. Knowledge also helps you notice when a result looks good while moving away from your own idea.

A pale network of fine fibres against a dark green background — Illustration by the author. Fibres as a visual association with connections and thought.

More images do not automatically mean more ideas

Eric Zhou and Dokyun Lee analysed over four million artworks by more than 50,000 users of an art-sharing platform. In their analysis, adopting text-to-image tools was associated with a 25% increase in productivity and a 50% increase in the favourites-per-view measure. The latter measures platform users' responses, not the objective value of art. ^[6]

Meanwhile, average content and visual novelty declined, although peak content novelty increased. The authors highlight the importance of ideation and filtering outputs. This is an analysis of data from a particular platform, not a promise that every artist will become more inventive.

In an experiment by Wadinambiarachchi and colleagues involving 60 participants, using an image generator during a design task was associated with greater fixation on an initial example and fewer, less varied and less original ideas than in the baseline group. It was a bounded task, but it demonstrates a risk: a tool's suggestion can narrow the search. ^[7]

I can therefore produce more while increasingly repeating the same idea. Access to a generator alone does not solve that problem.

A title also suggests what we should see

I later added “RAGE” to the dog image. The word suggests anger as an interpretation. If I called the same work “Play” or “Fear”, I would invite the viewer to look for different meanings. This is a proposed thought experiment, not a study I have conducted.

Leder, Carbon and Ripsas studied how titles affect the understanding and appreciation of paintings. In one experiment, elaborative titles improved understanding of abstract paintings without increasing aesthetic appreciation. Effects also varied with viewing time. ^[5]

Text placed inside an image also changes its composition. A title beneath it can change the context without changing the image itself. And the word “AI” adds another story: who made the work, and how.

I shared both image pairs and the exact prompts in a practical article about negative space. Here, I am interested in what viewers add after the maker has finished.

Another marketplace, another new era

I find it slightly funny how quickly we come full circle. A new tool, big promises, then a marketplace, a drop, a limited edition and a certificate. With NFTs, we also spent a lot of time discussing what exactly we were buying. Now we add AI to the shop's name.

This is not an imaginary scenario. Lux et Ars describes an offering combining a digital artwork, a fine art print and a physical certificate, with auctions, editions and provenance tracking. That describes the platform's operation; it is not evidence of speculation or a judgment of its artists. ^[8]

Jesus, people, I would still like to talk about the image itself.

Selling art does not bother me. What bothers me is expecting the technology to do all the explaining: why this is interesting, why it deserves attention, why I should care.

I am comparing the way the market is talked about, not the technologies themselves. NFTs and image generators serve different functions. Neither a certificate nor a model name answers my question about what I see in a particular work.

A finger reaching towards a luminous blue and orange surface — Illustration by the author. A gesture of touching an image.

“It did not really happen” can be a different objection

If someone admired a photograph because its maker captured a real event, learning that the scene was generated changes the basis of that admiration. That is an understandable distinction. In documentary work, the truth of the event is part of what the image promises.

I do not want to reduce every objection to bias. Concerns may also involve training data, working conditions or how an image's origin was presented to its audience. Research on AI labels does not invalidate those questions.

I want us to distinguish between discussions about production, facts and the experience of looking. With the next work you encounter, first write down what caught your attention. Then read about the process and check what changed in your evaluation. If the admiration suddenly disappears, try to name the reason.

Sources and scope

These studies examine particular samples, tasks and measures. They do not settle what counts as art. The passages about my practice and the comparison with NFTs are the author's commentary. The marketplace description was checked on 25 October 2026.

You liked it until you heard it was AI