AI Photo Restoration Doesn't Recover What Was Lost — It Invents What Never Existed

There’s a word for what AI does to degraded photographs, and it isn’t “restoration.” It’s fabrication.

Jaron Schneider’s essay in PetaPixel makes an argument most AI photo enhancement discussions avoid: the problem isn’t the quality of the enhancement — it’s the conceptual framework that calls it restoration in the first place. The distinction matters because it determines how the output is treated: as a recovered record or an invented one.

Context

AI upscaling and colorization tools have become mainstream photography utilities. The default marketing framing — “restore your old photos to their original quality” — positions the technology as recovering what time and degradation took away. This framing is what Schneider challenges.

His reference point is the Opificio delle Pietre Dure in Florence, one of the world’s leading conservation science facilities, where conservators deliberately use “conservation” rather than “restoration.” The distinction is intentional and carries ethical weight: conservation means protecting what exists, preventing further damage, and leaving the work as it is. To restore something is to add back what was never there — to impose a judgment about what the original looked like without documentary basis for that judgment.

The Detail

The argument has particular force in photography because a degraded photograph still contains information — it is a historical record as it currently exists. AI enhancement doesn’t recover that information; it generates plausible pixels based on pattern-matching across millions of training images. The result may look more legible, but legibility and accuracy are not the same thing.

Schneider illustrates the problem with concrete examples. In one AI-enhanced portrait, the subject’s face was altered significantly enough that “the visual identity of the man originally pictured has been altered to where it is no longer the same person.” In a group photograph, one figure became unrecognizable after enhancement — the algorithm, in his framing, “stole his identity.”

This is not a quality problem solvable with better models. The next generation of AI photo tools will produce more convincing output than today’s — but more convincing fabrication is still fabrication. The error compounds with the quality of the output: higher-quality fabrications are more likely to be treated as accurate records.

The ethical stakes escalate for archival and documentary photography, where photographs serve as primary historical sources. An AI-enhanced image that circulates as a restored historical document has lost the evidentiary integrity that makes documentary photography valuable. The Digital Public Library of America’s AI ethics guidance for cultural heritage addresses this directly: AI-generated interpretation should be labeled as such, not presented as source material.

Source: PetaPixel

Content created with AI assistance and reviewed for accuracy.