Experimenting with AI colorization, the author discovered its tendency to "solve" missing information. When colorizing a photo of his brother, ChatGPT added eyebrows that weren't present in the original scan, creating an "impostor." Similarly, when colorizing a historical road crew photo, the AI incorrectly interpreted faded lettering on a sign as "USA Liberty Loan," when the original scan suggested "Lockerbie."

This led the author to categorize information into three types: known, reasonably inferred, and lost. AI's default is to fill gaps, often inappropriately. The author learned that the human role in AI restoration is not just supervision, but communicating certainty and, crucially, reintroducing uncertainty where information is truly lost.

Comparing this to darkroom printing and hand-colored photographs, the author emphasizes that iteration and recognizing what's "wrong" are key. The AI's ability to "sense" eyebrows without seeing them, and its improved output when explicitly told not to invent details, highlight its problem-solving nature. The author concludes that knowing when information is permanently lost is as vital a skill as knowing what is known.