...
Your brain auto-corrects white balance so a white shirt looks white under yellow light and blue light. AI colorizers don't have that ability — and the failures expose fascinating limits in computer vision.
Look at a white piece of paper under an incandescent bulb. It looks white. Now look at the same paper under fluorescent office lighting. Still white. Under the noon sun. Still white. Your brain performs an automatic white balance adjustment so seamless you never notice it. The paper is objectively reflecting different wavelengths of light in each scenario — yellowish under incandescent, greenish under fluorescent, bluish under midday sun — but your visual system subtracts the color of the light source and reports "white."
This ability is called color constancy. AI colorizers do not have it. And their failures to achieve it reveal something profound about the gap between computer vision and human perception.
Your visual system does not measure absolute color. It measures relative color — how each surface compares to the average color of the entire scene. If the scene average is yellowish (incandescent lighting), your brain subtracts yellow from every surface, making the white paper appear white. If the scene average is bluish (shade), it subtracts blue.
This is why the famous dress illusion (#TheDress, 2015) split the internet. Some people's visual systems assumed warm lighting and subtracted yellow, seeing the dress as blue and black. Others assumed cool lighting and subtracted blue, seeing it as white and gold. Neither group was wrong — their brains made different assumptions about the illuminant. The dress photo was accidentally perfectly ambiguous.
Color constancy is not perfect. Optical illusions like the checker shadow illusion (Edward Adelson, 1995) exploit its limits — squares A and B are objectively the same shade of gray, but B looks lighter because your brain compensates for the perceived shadow. Knowing it is an illusion does not make it stop working. The compensation is involuntary and pre-conscious.
An AI colorizer is trained on millions of color photos. It learns that grass is usually green, sky is usually blue, skin is usually in a range of warm tones. When it receives a black-and-white photo, it predicts colors based on these statistical priors — what color things "should" be.
But black-and-white photos erase the illuminant information. A B&W photo taken under warm incandescent light looks identical to one taken under cool fluorescent light — the film or sensor captured only luminance, not chrominance. The AI has no way to know what the original lighting was. So it guesses, and its guess is always "neutral daylight" because that is the statistical average of its training data.
This produces photos that look plausible but wrong. A B&W photo of a room lit by warm candlelight gets colorized with neutral colors, losing the warm amber glow that defined the original scene's mood. A photo taken during golden hour gets colorized as if it were noon. The AI produces technically competent colorization that misses the emotional content carried by the original lighting.
Even more interesting: AI colorizers sometimes reproduce the same illusions humans fall for. If a B&W photo contains a checker shadow pattern — identical gray values that a human perceives as different due to context — the AI may color them differently too, because it learned from human-labeled data that encoded the illusion. The AI did not develop color constancy. It learned to mimic human color constancy failures.
This is a recurring theme in AI: models learn to reproduce human biases and perceptual quirks, not because they share our biology, but because they were trained on data produced by humans who have those quirks. The AI colorizer's "mistakes" are a mirror of human perception.
When you colorize a black-and-white photo, understand that the AI is making a best statistical guess about colors it cannot possibly know. For family photos, the result will look natural because skin tones and grass and sky are statistically predictable. For photos with unusual lighting — candlelight, neon, colored stage lights, underwater — the AI will neutralize the lighting and produce colors that are plausible but historically wrong.
If lighting mood matters for your photo, manually adjust the white balance and color temperature after AI colorization. The AI gives you a starting point, not the final answer. The gap between "looks natural" and "matches the original scene's lighting" is where human judgment still beats the algorithm.
Try it yourself at AI colorizer — upload a black-and-white photo and see how the AI interprets lighting it cannot measure.