Why AI Isn't Going to Make Art?

The Paradox: AI Art Is Booming While Its Artistic Status Collapses

Art is having its strangest moment. The generative AI art market is projected to grow 42% through 2029 to over $2.5 billion, according to Brookings analysis — even as audiences, courts, and working artists push back harder than ever against calling any of it art.

The tension crystallized in February 2025, when Christie's staged "Augmented Intelligence," the first AI art sale at a major auction house. The sale went ahead — one work, "Machine Hallucinations," fetched over $270,000 — but not before a public letter gathered more than 6,500 signatures demanding its cancellation. The same institution selling the art was being told the art shouldn't exist.

The backlash extends far beyond one auction. Consider the scale of resistance:

• 400+ Hollywood insiders signed a March 2025 open letter opposing Google and OpenAI's push for copyright exceptions for LLM training.
• At least 16 lawsuits have been filed against nearly every major AI company over copyright in the past two years, with Fair Use still unresolved.
• 200,000+ creative jobs have been lost across creative industries over four years, per figures cited by Current Affairs.

Meanwhile, the technology keeps improving at a pace that makes the cultural fight feel urgent. The old tells — malformed hands, garbled text — are gone. By 2025, Midjourney and DALL-E render anatomically correct hands and flawless typography, and one journalist fabricated an entire fake political scandal in 28 minutes for $8, per the GIJN's detection guide.

So we arrive at the question underneath all the noise: can AI actually make art, or just images? The distinction matters. Images are pixels arranged pleasingly; art, as critics like The Guardian's Eric Reinhart argue, requires an interior life reaching toward another person's experience. Machines "do not desire, do not live in a body, and have no interiority," he writes — producing what he calls spectral mimicry that "may deceive the eye but not the soul."

The courts, notably, are siding with the skeptics. The U.S. Copyright Office maintains that copyrightable works must "owe their origin to a human agent" — and that prompting a GenAI model doesn't qualify, because the model itself controls the output. Whatever AI produces, legally speaking, it isn't authorship.

That leaves businesses in a genuinely ambiguous spot, and it's the same ambiguity SkillzBench sees across the AI tooling landscape: capability is exploding while trust and verification lag behind. The market says yes; the culture, increasingly, says no. The rest of this article is about who's right.

The Evidence: Why Audiences and Law Both Demand a Human Hand

What if the exact same painting suddenly became worth less the moment you learned a machine made it? That is precisely what peer-reviewed research shows happens in the human mind — and the law agrees.

Two within-subjects studies (N=150 and N=151), published in Cognitive Research, presented participants with identical artworks and changed only the label. Works labeled "AI-created" were rated significantly lower on Liking, Beauty, Profundity, and Worth. The effect was moderated by perceived narrativity and effort: when participants watched robots physically producing art, the bias softened. Audiences, it turns out, are paying for the human hand and the story behind it — not just pixels.

The creativity itself may also be thinner than it looks. A Georgetown study of over 370,000 college admissions essays compared writing before and after ChatGPT's 2022 release. Post-ChatGPT essays showed greater word diversity but fewer original ideas — and creativity experts rated the AI-assisted essays as more creative anyway. Neuroscientist Adam Green calls this an "illusion of creativity": a single system produces homogeneous ideas, and "if we allow AI to do the creating, then we'll be stuck with the ideas that already exist."

The law draws the same line. The U.S. Copyright Office holds that copyrightable works must "owe their origin to a human agent," and that prompting alone doesn't qualify, because per Brookings analysis, "the GenAI model itself would control or determine the output rather than the human prompter." The Zarya of the Dawn case made this concrete:

• Midjourney-generated images in the comic were excluded from copyright registration.
• Only the author's human "selection, coordination, and arrangement" remained protected.
• At least 16 lawsuits over AI copyright have been filed against major AI companies in two years.

Philosophers close the argument. Marcus du Sautoy observes that "no machine is driven to express itself creatively. It is prompted by the intention of the human." And Eric Reinhart goes further: machines "do not desire, do not live in a body, and have no interiority," making AI art "spectral mimicry, a simulacrum that may deceive the eye but not the soul."

For teams building AI workflows, the lesson is to keep humans in the loop — choosing, arranging, verifying. That same human-in-the-loop philosophy drives SkillzBench, which is building a free directory and MCP connector where every agent skill is security-scanned and benchmarked so humans decide what to trust. Skills, proven.

The Counterargument: AI as Collaborator, Not Replacement

The strongest case against "AI art" has a serious intellectual opposition — and pretending otherwise would be its own kind of dishonesty. Oxford mathematician Marcus du Sautoy, no techno-optimist lightweight, argues that AI "might help us to stop behaving like machines…and kick us into being creative again as humans," and that code exposed to data can genuinely "learn, mutate and evolve" into something deserving the label creativity (BBC Future).

Georgetown neuroscientist Adam Green takes a middle path with practical weight. He concedes "AI can be creative" but warns against treating it as a replacement — his recommendation is to start with your own ideas, then use AI to sharpen or edit them (Georgetown's creativity research). His larger worry is homogenization: AI's single system produces uniform ideas, and "if we allow AI to do the creating, then we'll be stuck with the ideas that already exist."

Then there is the philosophical grenade from Alice Helliwell of Northeastern University London, who invokes Marcel Duchamp: "It is hard to see why a urinal can be art, but art made by a generative algorithm could not be." Definitions of art have shifted before — photography once terrified painters — and machine learning systems are "black boxes" we don't fully understand (the same BBC piece notes). Even the peer-reviewed perception studies showing anti-AI bias acknowledge mixed evidence: some experiments found participants equally willing to call hypothetical AI products "art" (Cognitive Research).

But here is where the debate collides with an uncomfortable reality: you can no longer tell the difference. The old tells — malformed hands, garbled text, asymmetrical faces — stopped working by 2025, when Midjourney and DALL-E began rendering anatomically correct hands and flawless typography (GIJN's detection guide). The stakes of that shift are stark:

• One journalist fabricated a complete fake political scandal in 28 minutes for $8, using generative tools.
• Peer-reviewed research confirms GANs, diffusion models, and VAEs now synthesize high-quality multimedia while raising concerns about unethical usage and societal harm.
• The U.S. Copyright Office cannot reliably distinguish AI from human authorship without disclosure, leaving the burden of verification on humans.

This reframes the whole argument. If detection fails, the battleground moves to provenance — knowing where an image came from, who made it, and what tools touched it. That is the same principle SkillzBench is building on for agent skills: pinning every version to a commit SHA, security-scanning, and benchmarking independently, so trust comes from verification rather than eyeballing. In a world where a convincing fake costs $8, verification is the new authorship — and the human question shifts from "who made this?" to "who vouches for this?"

The Human-First Workflow: How to Use AI Images Without Losing Value (or Legal Protection)

So AI can't make art — but your business still needs images. The good news is you can use AI visuals responsibly, provided you keep humans at the center of the process.

Put humans first in the creative loop. Audiences reward perceived effort: peer-reviewed studies found identical artworks rated lower on liking, beauty, and worth when labeled AI-created. Human involvement also preserves legal protection — in the "Zarya of the Dawn" case, only the author's human "selection, coordination, and arrangement" remained copyrightable, per Brookings analysis. Georgetown neuroscientist Adam Green recommends starting with your own ideas, then using AI to sharpen or edit — never to originate.

Treat AI visuals as legally uncertain for commercial use. The U.S. Copyright Office holds that AI outputs don't "owe their origin to a human agent," and at least 16 lawsuits against major AI companies are pending. International rules diverge sharply: the UK grants copyright to whoever arranged a computer-generated work's creation, and China recognizes copyright for sufficiently detailed prompts, while the U.S. has taken no legislative action. Until Fair Use is resolved, treat any fully AI-generated hero asset as unprotected.

Benchmark output independently to dodge the creativity illusion. A study of 370,000+ college essays found post-ChatGPT writing showed greater word diversity but fewer original ideas — yet experts rated it as more creative. AI-assisted work can look better while actually being worse. The fix is measurement:

• Compare the same model's output with and without AI assistance, blind
• Measure originality, not surface polish
• Track whether human editors meaningfully change the result
• Re-test when you swap tools or models

Verify provenance of any AI tooling you adopt. Published research found roughly 26% of agent skills carry at least one vulnerability, so knowing exactly what code and instructions you're running matters. SkillzBench is building exactly this kind of infrastructure for agent skills: commit-SHA pinning for every version, security scanning, license checks, a 0–100 trust score, and Lift benchmarking that tests the same model with and without a skill — so you know what actually helps rather than trusting the illusion.

The pattern across all four steps is the same: AI contributes, humans decide, evidence verifies. That's the workflow that keeps your visuals valuable to audiences and defensible in court. To put it into practice, browse and test free skills on SkillzBench — every skill is pinned, scanned, and benchmarked so you adopt only what measurably lifts your agents.