Let me start by stating one of my conclusions here: our current text-to-image AI technology, while impressive, is still immature. It has the potential to go in two opposite directions—one of them will give artists ultimate control, the other removing them from the equation entirely. Both paths interesting and probably worth pursuing, but only one of them is not existentially alarming.
Two years ago, seemingly out of nowhere, we gained the ability to generate strikingly realistic images using text-to-image AIs like DALL·E, Midjourney, and Stable Diffusion. While other forms of text-to-media AI exist—including recent advances in text-to-video—this post will focus on text-to-image AI, though my conclusions likely apply more broadly.
We all experimented with it at first. We opened a browser, typed something ridiculous like "a robot painting Van Gogh's Starry Night," and moments later, we were confronted with an existential crisis: how can this "thing" create such appealing results in seconds, while trained professionals take hours or even days?
Yet, as remarkable as these results are—and putting aside the growing legal issues, which I won’t address here—artists face significant challenges when working with AI tools. At the end of the day, despite their disruptive potential, text-to-image AIs allow us to control only the concept. We cannot fully dictate colors, lighting, composition, or fine details.
Artistic Limitations of AI-Generated Images
Consider the above AI-generated image of a robot painting Van Gogh’s Starry Night. As an artist, here are some problems I’d like to solve:
The blue and yellow hues of Van Gogh’s painting blend too much with the surrounding colors, reducing the contrast needed to distinguish the painting from the robotic setting. The same issue applies to the robot’s colors.
The bright lights near the robot’s head are too intense, shifting the focal point away from where it should be: the contrast between the robot’s hand and the painting.
The harsh lab lighting is ill-suited for painting; it creates an unsettling inconsistency, making the robot’s act feel less believable. A mix of natural daylight from windows and scattered studio lights would be more fitting.
The conceptual contrast is off. Ideally, the robot would be painting in a futuristic room with a window overlooking nature, reinforcing the contrast between the robot’s world and Van Gogh’s world. Alternatively, placing the robot in an ancient studio could create a contrast between the robot and its surroundings.
The scene needs a wider perspective. Seeing more of the environment—perhaps other paintings the robot has made—would add depth. Are they all Van Gogh reproductions? Only Impressionists? Great masters? Such consistency would enhance the narrative.
The glass on the robot’s head is interesting, but the elements inside are unreadable. Ideally, we’d see mechanical eyes or some other visual representation of sight, raising philosophical questions about whether the robot possesses a form of perception—or whether we, in some way, are robots ourselves.
The brush makes little sense, and the painting’s surface lacks the characteristic texture of oil paint. Moreover, the painting appears to glow unnaturally, as if unaffected by the room’s lighting.
No one paints inside a frame—it’s added afterward. (Unless, of course, this is a robotic ability, in which case it should be more clearly intentional.)
The left leg of the easel is floating above the table, and the objects on the table lack coherence.
I could go on, but out of respect for your time, I won’t. In my own work, I make hundreds of such adjustments, and many artists do the same. Great art rarely happens by accident—it’s the result of time and refinement.
The Future of AI-Assisted Art
Could I have corrected some of these issues with a better prompt, more advanced AI techniques, or minor Photoshop edits? Perhaps. But as of today, achieving my desired result purely through AI is impossible.
And this raises a fundamental question: Can we get the results we want from AI? Or, perhaps more importantly, why do we need to get the results we want? The answers to these questions will potentially define the future of this technology—because both outcomes are possible.
One day, AI will function as a true artistic assistant. We will move beyond text prompts to more sophisticated interactions—speaking to AI, pointing at elements, using intuitive controls (whatever replaces the mouse and keyboard). Decades from now, we’ll be able to make all the adjustments I listed and more. Let’s call this type of AI human-oriented AI.
But there will also be independent AIs. Why should materializing my artistic vision be more valuable than experiencing an AI’s vision? What if independent AIs generate concepts as compelling as ours—and execute them with equal or greater mastery? What if they can even explain their artistic choices to us? Perhaps, in the not-so-distant future, an AI will conceive a better robot painting Van Gogh image than I ever could.
The Fork in the Road
We are at a crossroads: we can develop human-oriented AI to give artists ultimate control, or we can create independent AI that no longer needs us. In all likelihood, we will pursue both—because humans always push technology forward, regardless of the consequences. Not that I think the consequences will necessarily be bad, but history shows that we tend to worship technological progress.
One reason I’m drawn to technology is that it forces us to confront philosophical questions. When photography emerged, realist painters were dismayed—after all, even a modest camera could capture reality more accurately than the greatest painter. This existential crisis forced artists to reconsider the purpose of realism, which led to groundbreaking movements like Cubism. Instead of mimicking reality, artists explored new dimensions—showing multiple perspectives of a cube simultaneously, something a camera could never do.
It’s hard to predict what our current existential crisis will bring. As I said, I believe that in the future, artists will create more freely, quickly, and with greater control—removing one existential threat. But independent AIs will create their own art, without human input. Perhaps we will embrace both, appreciating human-made art for its connection to the artist’s story and AI-generated art for its alien or even divine qualities.
Since Barthes' The Death of the Author, many have argued that the artist’s intentions don’t matter—that the observer’s interpretation is more important. This remains the dominant view. Perhaps, in the future, we’ll take this theory even further—not only will the artist’s intent be irrelevant, but there may be no human artist at all.
This text was written by an AI.
Nahhhh, just kidding—it’s still just me. :)
Comments