Published: 08.04.2026
In my previous essay, I argued that the story behind a piece of art and the quality of its representation matter more than the labor required to produce it. Since writing it, I’ve come to think that claim needs a significant qualifier, and being upfront about this is more useful than saving it for the end: the principle holds only if the creator has actually done the work of having a story worth telling, knows which parts of the process were genuinely theirs, and is willing to say so. Without those conditions, “the story matters more than the labor” becomes a convenient way to dress up content without conviction.
What moved me toward this qualification was paying closer attention to how people actually respond when they encounter AI art in the wild. The patterns in the research are consistent enough that I can’t treat the discomfort I described in my first essay as a mere cultural lag, something that will fade once people get used to the technology. The reaction appears to be deeper and more structural than that.
What the reactions actually look like
When people are told a piece of art was made using AI, they rate it lower on beauty, profundity, and perceived worth, even when they themselves say the work is visually indistinguishable from something human-made. The label alone is enough to suppress the evaluation, which is a response to the the origin of art, rather than it’s quality. The reason this happens, as far as I can tell from the research, is that people experience art partly through a felt connection to whoever made it. They project empathy onto the creator, read intention into ambiguous choices, and build a kind of temporary identification with the person behind the piece. AI breaks that loop. Since there is no person to identify with, so the empathy has nowhere to go, and without it, the preference drops. And people seem to sense this even before they can articulate why. On the institutional side, the commercial art world reflects a similar caution. Gallery professionals report that collector interest in AI art remains limited, and the vast majority of artists represented by galleries are not using AI in their practice. There is not even a shared definition of what AI art is. The overall posture is one of wait-and-see, with skepticism about AI’s legitimacy as an artistic medium running well ahead of any enthusiasm for its potential. What I take from all of this is that the consumer bias I described in my first essay, the one rooted in perceived labor and creator context, operates at a deep psychological and even neurological levels. When a consumer learns that a work was AI-generated, they don’t dock a few points and move on. The knowledge works backward through every layer of evaluation. The gut-level response weakens because there’s no human act to anchor it to. The viewer stops reading meaning into ambiguity, because generosity toward the work depends on believing someone put it there deliberately. And the creator context that would normally give the piece biographical weight simply has nowhere to attach. The art may be identical pixel-for-pixel, but knowing its origin hollows out the encounter with it.
AI art needs to be its own category
I’ve come to think that much of this friction exists because we’re trying to evaluate AI art using standards that were built for something else. When someone looks at an AI-generated image and judges it against the implicit criteria they use for conventional art (how much skill was required, how long it took, what personal sacrifice it represents) the AI work will almost always come up short. Those criteria were never designed for it. This has happened before. Photography spent over a century fighting for recognition as fine art. Through much of the 1800s, critics dismissed it as a mechanical process that lacked the human spirit of painting. As late as 1960 in the UK, photography was still not widely accepted in that role. The objections will sound familiar: photography relied on a machine, it was too easy, it made image-making accessible to people who hadn’t earned the right to it. What eventually settled the question was photography developing its own language. Early photographers, lacking their own tradition, borrowed one from painters. The result was work that tried to look like paintings and failed on painting’s terms. Some historians described this as “photo-painting rather than photography,” and it led nowhere. Photography came into its own when practitioners stopped imitating painting and started exploring what a camera could do that a brush couldn’t: candid moments, documentary realism, the play of light on actual surfaces. (The analogy to AI art is useful, but it has a real limit worth stating plainly. A photographer still had to be physically present. They had to point the camera at something in the world and make compositional decisions in real time. The machine mediated the image, but the human body and eye were always part of the process. AI art removes that requirement entirely. You can generate a photorealistic street scene in a city you’ve never visited, depicting light you’ve never seen. The gap in embodied involvement is large enough that critics of AI art will reach for it immediately, and they won’t be wrong to. The reason the analogy still holds, despite that gap, is that the core dispute was never primarily about physical presence. Painters who rejected photography were not arguing that photographers failed to show up. They were arguing that the machine did the important work and the human merely triggered it. That objection is almost word-for-word what people say about AI art today. What photography’s history demonstrates is that this framing eventually collapsed, because practitioners showed that the creative decisions surrounding the machine (what to capture, when, how to frame and present it) constituted a legitimate artistic practice on their own. The open question for AI art is whether a similar case can be made when the machine’s contribution extends well beyond what a camera ever provided. I think it can, but it demands a clearer account of where human creative labor actually lives in the process, which is what the next section tries to address.) A more recent example is short-form video. TikTok and Reels content is not evaluated by the standards of filmmaking. Nobody watches a 45-second video essay and asks whether the cinematography holds up against a Nolan film. The format developed its own criteria: pacing, hook quality, relatability, the ability to convey something resonant within severe constraints. The people who make this content are called creators, not directors, and that distinction in language reflects a genuine difference in what the work is and how it should be judged. The format didn’t earn respect by competing with cinema on cinema’s terms. It earned respect by being good at something cinema wasn’t even trying to do.
Evaluation standards for AI art
If AI art keeps being measured by how well it approximates traditional art, or by how much manual skill it required, it will keep losing. The interesting question is what evaluation criteria would actually make sense for the medium. Drawing on the creation framework from my first essay, I think the criteria would need to emphasize at least a few things:
- The quality and depth of the originating story. In the encoding system I described previously, art begins as a network of memories and lived experience in the creator’s mind. That origin point is identical whether you pick up a brush or open Midjourney. The richness of what you’re trying to say matters before any tool enters the picture.
- The skill involved in iterative refinement. I previously described how AI shifts the creator’s role from pure generation to generation-plus-curation: steering, selecting, and refining from a larger field of possibilities. That curation loop is where taste, judgment, and editorial instinct live. A creator who runs forty iterations, rejecting most and building on a few, is doing something meaningfully different from one who posts from the first few outputs.
- The intentionality of compression choices. In my earlier framework, compression refers to how much of the full story in the creator’s mind gets distilled into the final work. With AI, I argued that compression becomes a more purely creative choice because the cost of each iteration drops so low. But that freedom is only valuable if the creator actually exercises it. Did they make deliberate decisions about what to show and what to leave implied, or did they accept whatever the model gave them?
These are genuine creative skills. They look different from brushwork or darkroom technique, but they reward taste, judgment, and clarity of vision in ways that are not trivial to develop. Some of this is already being tested. Dedicated AI art competitions now evaluate work on criteria like creativity, originality, and technical execution, assessed by panels that include past winners, public voters, and even AI judges. These are early experiments, but they represent an attempt to build evaluation norms native to the medium rather than inherited from older ones. I suspect AI art will follow a trajectory similar to photography’s, although faster. The category won’t be established by institutional gatekeepers. Galleries and museums took decades to accept photography. The norms will more likely emerge from creator communities and platforms, and they may crystallise around the kinds of criteria I described above, ones that take the medium on its own terms.
Disclosure is necessary, and it’s harder than it sounds
For any of this to work, though, there’s a prerequisite: honesty about how AI was used. If AI art is going to develop its own category with its own evaluation standards, those standards are meaningless if creators can game them by obscuring how much of the work was theirs versus the machine’s. The regulatory world is starting to move on this:
- The EU AI Act now requires transparency obligations for AI-generated content, including disclosure requirements for synthetic media.
- The U.S. Copyright Office’s 2025 report established that AI-assisted works can qualify for copyright protection, but only if the human contribution is substantial and demonstrable. The mere use of AI doesn’t preclude eligibility, but basic prompts or trivial modifications are not enough.
- The C2PA coalition is developing tamper-evident metadata that embeds provenance information directly into image files.
These are useful steps. But they all share a limitation: they treat disclosure as a binary. You either used AI or you didn’t. In practice, AI can be involved at multiple stages of the creation process, and each stage carries different implications for how much of the creative work was genuinely human. Consider the stages I outlined in my first essay. AI could be involved in:
- Originating the story concept. Did the core idea come from the creator’s lived experience and imagination, or did AI generate the concept?
- Transforming fragments into a coherent narrative. Did the human shape scattered ideas into a structure, or did AI do the organising? This is where the line between creative director and button-pusher tends to live.
- Determining the compression level. Did the creator make deliberate choices about how much abstraction, how much detail, how much to show and what to leave implied? Or did they accept the AI’s default level of representation?
- Executing the final rendering. Did AI generate the visual or textual output? This is the stage most disclosure focuses on, and arguably the least important for artistic value.
- Refining through iteration. How much did the human select, reject, steer, and rework across multiple cycles?
A person who has a deeply personal story, carefully structures its arc, makes deliberate compression choices, and only then hands the execution to Midjourney has done something fundamentally different from someone who typed a four-word prompt and posted the first output. Yet under current disclosure frameworks, both might check the same “AI-assisted” box. The right direction, I think, is toward graduated disclosure: something that communicates not just whether AI was involved but where in the creative process it entered and how much steering the human did. Harder to implement than a checkbox, but far more honest, and it would give consumers the information they actually need to evaluate the work on appropriate terms. But even graduated disclosure runs into a verification problem. From the outside, it is extremely difficult, perhaps impossible, to determine at which stage AI was used and to what extent. Did the creator really conceive the story independently, or did a conversation with an LLM seed it? Did they genuinely iterate across dozens of outputs with critical judgment, or accept the third one that looked good enough? I doubt any metadata standard or watermarking scheme fully closes this gap, and it’s likely that some of it never gets completely resolved. If that’s the case, the burden of proof shifts onto the creator. Consumers who can’t verify the process will default to judging the output even more harshly, raising the quality stakes for AI art further. The category will need to earn trust the hard way, through work that is so clearly vision-driven and intentional that the question of who or what rendered it becomes secondary.
The self-deception layer
There’s a deeper complication worth flagging. Even if creators are acting in good faith, they may genuinely misjudge how much of the creative contribution was theirs. A widely discussed thread on LessWrong captured this well. One commenter pointed out that it’s easy to convince yourself the AI is just expressing your own thoughts, but this often turns out to be an illusion. A follow-up post offered a useful analogy: editing AI output does involve some generative effort, but it’s magnitudes less than writing from scratch, in the same way that watching a professional tennis player teaches you something about the sport but is nothing like the internal experience of playing at that level. Reviewing and approving AI output gives you the surface pattern of creative work without requiring the underlying cognitive process. And there’s growing anecdotal evidence that this gap compounds over time. Several writers have described how regular use of AI tools gradually eroded their ability to produce work independently, even in domains where they were previously strong. The implication for disclosure is uncomfortable: it’s possible for a creator to sincerely believe they directed the creative process when the AI shaped it more than they realise. This doesn’t make them dishonest. It makes the problem harder. The disclosure conversation needs to go beyond “tell the truth about what you did” and into the more difficult territory of “develop enough self-awareness to know what you actually did.”
The honest assessment of where things stand is that AI art’s future is bottlenecked less by the quality of the tools and more by trust. The category can’t develop its own evaluation standards if creators aren’t transparent about their process. Transparency is hard to verify even with good intentions. And good intentions themselves are suspect when the tools make it so easy to mistake curation for creation. I don’t think this resolves through any single framework. It will more likely happen the way most hard cultural problems do: slowly, unevenly, through norms that accumulate before anyone can fully articulate them. But I also think the burden does not fall on creators alone, because if AI art is going to develop into something worth taking seriously, consumers need to come toward it with some openness. That means being willing to evaluate work on terms native to the medium, rather than reflexively downgrading it because an AI model was involved. The research I described earlier shows that consumers often penalise AI art even when they admit the work itself is indistinguishable from human-made art. That’s a bias worth examining, not just a reasonable preference. The ask to creators is: be honest about what you did and what the machine did, and develop the self-awareness to know the difference. The ask to audiences is: if that honesty is present, engage with the work on its own terms rather than dismissing it for how it was made. Neither side of that exchange works without the other. And for now, neither side has fully committed to their part.