Images from noise are noise
I’m not in glasses territory yet, but 20/20 my eyes are not. I’ve had too many birthdays not to be thinking about some form of corrective eyewear in the not-too-distant future. However, I do have a strangely tuned eye for spotting AI-generated images.
I recently stumbled, jet-legged, through an Osaka market strewn along a busy, vibrant, and overwhelming street. At least fifty meters ahead, a cluster of hoodies swung low from the front of a building. In a street overflowing with color and noise and things foreign to me, those hoodies stood out. I couldn’t make out what was on them at first, but I was immediately certain they were covered in AI-generated images. My certainty and discomfort grew with each step closer. These were some mad and wretched objects.
Can the barrier to entry for creativity be too low? Is there such a thing as bad art? In the context of AI, yeah. Now there is. There was a cruel but useful filter built into the old way images were made. People who, I do not count myself among, have developed the skills required to produce something have at least the ability to learn something. The time and skill investment encouraged the creator to consider whether the result would be worth the effort.
Lowering or removing that barrier completely opens a floodgate to people less interested in the craft of creativity and getting better at it, or in using it to say something, and more interested in commerce, with only the aim of creating things that could have remained uncreated. Or are not worth the energy externality cost to the environment even if they are worth the time cost of the prompt writer.
I’d like to know if those whose skills and interests are more aligned with pure commerce are delighted they no longer have to collaborate with creative people to fill their shelves with what appears to them to be costless equivalents. Are they concerned about the result if everyone else did the same?
The more art and designs we have, the better, right? More variety can be liberating. But what is it all to be used for? To sell? To deceive? To displace? When the primary goal is units shifted, I almost grieve the lost opportunity to have something made that might have been worth producing.
Supporting artists may or may not be of interest to you, but there are other reasons why knowing whether an image is AI-generated or not might be important or relevant to you. Ethical, legal, creative attribution, and ultimately, to provide a base level context for how to interpret the image. When buying a product, images are no longer enough proof of quality, proof that the item exists, or a good indicator of what will arrive at your doorstep.
AI tools are improving in quality (deceptiveness?) all the time. However, experts and tools exist for analyzing and identifying digital images, and these are also improving alongside AI tools. Still, out in the world, when you’re standing in a market staring at a hoodie, you’re on your own. So how do you tell?
Is it too good to be true? #
Tendency towards sensationalist, unlikely, garish, or exaggerated details. Where a human might draw or be likely to photograph a herd of 3-4 horses, a diffusion process will generate 10-100 horses. Real photography rarely has every detail in perfect focus, perfect framing of content, or every element of drama cranked to the visual maximum. Consider the source, do they have the skills, or the means to achieve this level of production?
A lack of real-world understanding #
Statistically averaging concepts from thousands of images can create convincing results at a glance, but one layer of knowledge down, these averages often lose the meanings, limits, and circumstances of the reference images and leave them behind in their munged-together rendition.
Would a garden of that size have a hose that would be wrapped that many times? Would someone put a full-sized window into an internal wall? Does that stage lighting set make sense? Why does the stitching of the seam end there? Would those three lanes merge into one that close together? Did that model of car even exist in that year? Deductive reasoning, deep observation, knowledge of weirdly specialised topics, use your inner Sherlock Holmes here.
Perspective problems #
Multiple vanishing points, misaligned geometry, objects that don’t quite meet or intersect naturally, and little spatial betrayals reveal a model guessing rather than observing (or making local compositional decisions that need a larger context).
Watch the shadows #
Lack of understanding and application of the physics of light. AI understands these concepts by example, but not to the level required to implement them at the calculated level of detail needed for convincing simulation. Look for shadows cast by illogical light sources or inconsistent color temperatures. Is that candle emitting a bright blue neon glow?
Boxy, geometric tendencies #
Tendency towards boxy geometric forms and non-intersecting overlapping edges, right-angled furniture, and avoidance of complex natural or vague edges. Definitive and distinct separation borders.
English everywhere #
AI has a habit of sprinkling English text into places where you’d normally expect a local language, it struggles to infer cultural context. Does the language match the location? Does the tone and standard of the language match the setting?
Uncanny composition #
Uncanny valley level of photo realism, unnaturally crisp subjects, unusual or ambiguous arrangement of objects in space. AI techniques see the world and human faces very differently from how we do. Human faces have subtle cues that always seem slightly off or absent in even the most impressive AI-generated face.
It’s hard to miss the high saturation, tendency towards vivid, unbalanced colors, and higher contrast AI seems to favor. Attention-grabbing, reality-distorting. Unrealistic outcomes, another “too good to be true” metric.
Absence of human intention #
This is the hardest one to articulate, but the easiest to feel. Hollow, soulless, lacking the intention or artistic conveyance of composition or expression. Ask yourself, “Why did/would someone produce this image?” Does this have a point of view? Does it have the tool marks of a culture?
To take a photo is to actively frame, selectively edit, and capture a slice of time and space, to record a version of reality. To paint or draw an image is to conjure forms and shapes to illustrate a feeling, capture a moment. Art communicates a relationship the artist has with society. AI images tend to lack depth of purpose or meaningful drama, or they’re sugary calories with no nutritional value, or they’re a ham-handed attempt at it. A lack of a point of view, a lack of tacit knowledge, because you never had to develop them through iteration and growth through participatory talent.
I didn’t buy that hoodie, if you could guess. Not because of the tool used to make it, but because I felt this particular instance wasn’t made by anyone for anyone in particular. I do believe the tool is not necessarily the problem and that creativity can be exercised, assisted by, and achieved using AI. Let’s hope those creative people will continue to develop the skills they need to use these tools wisely.