The Void of Data: When Information Ends

In a photograph from the late summer of 1976, my brother is eight years old. He is kneeling in our South Bend front yard, squinting against the glare of the sun. I can recall these details with clarity even after half a century.

Recently, I began experimenting with AI-driven colorization of old black-and-white scans to see where the technology's limits lie.

The results were an interesting mix of accuracy and guesswork. The green of the house was almost right. A neighbor's car in the background—a detail I remember only faintly—was rendered in a pale blue by ChatGPT, which felt plausible enough.

The Technical Struggle for Detail

My process involved different levels of hardware:

1. I started with a cheap \approx \100$ "digitizer" from Amazon.
2. I upgraded to an Epson V300 flatbed using SilverFast software.

Despite the better equipment, some things remained invisible. Specifically, my brother's eyebrows simply would not appear.

I then tried a different project: a small, three-inch print of a 1920s road crew constructing US 40 in east-central Illinois. I used a Canon CanoScan 9000F, which is a powerful tool, but it can only extract so much from a tiny print.

To help the AI, I provided reference photos I took in 2007 of the same bricks, now part of an abandoned road.

While the AI guessed the workers' clothes were shades of tan, gray, and blue, there is no historical record of what they actually wore that day. Furthermore, the text on a sign in the image remained a mystery; I could make out "Lockerbie," but the rest was gone.

Categorizing Information

This experience led me to realize that data falls into three distinct categories:

• Surviving Information: Data that still exists and can be verified (e.g., the physical bricks).
• Inferred Information: Data that can be reasonably guessed based on context (e.g., 1920s workwear).
• Lost Information: Data that was never captured or has vanished.

The AI's "Problem-Solving" Nature

I discovered that AI views a gap in information not as a void, but as a problem to be solved. Mathematically, it operates on probability: $P(\text{pixel color} \mid \text{surrounding context})$

When I told ChatGPT not to invent details, it still struggled. It added removed slats from a gable and sharpened edges that shouldn't have been sharp. I didn't realize I had to explicitly forbid the AI from inventing eyebrows until I saw it actually do so.

if information == None:
    return generate_statistically_likely_detail()
else:
    return existing_information

The Human Element

This process reminded me of the traditional darkroom, where a perfect print is never achieved on the first try. It also mirrored the hand-colored family photos I inherited from my mother; the original artist made guesses that were often wrong because they lacked the personal knowledge I possess.

My workflow evolved into a series of checks:

• Identify known colors.
• Accept plausible inferences.
• Prevent the AI from filling the "permanent gaps."

The human's role in AI restoration isn't just supervision. It is the act of communicating certainty where it exists, and—more importantly—reintroducing uncertainty where it must.

I know my brother's face. I know when the AI is lying to me.

---

Community Perspectives

Khurt Williams (AWS Certified) noted that AI isn't "thinking" but is simply calculating the most likely pixel. He agreed that the "darkroom" approach of iteration is the real work.

Jeff Sun offered a different take, suggesting that if a model can be built to fill gaps, a model could theoretically be built to refrain from doing so, though he noted that humans are inherently different in this regard.