What Restoration Destroys

The China Film Foundation has announced a project to restore 100 classic martial arts films using AI. The titles include Jackie Chan's Police Story, Bruce Lee's Fist of Fury, Jet Li's Once Upon a Time in China. The stated goal is enhancement of image quality, sound, and production values while — explicitly, in the Foundation's own language — "preserving original story and aesthetics."

That phrase is carrying more weight than it appears to know.

In the United States, a startup called Fable (formerly Showrunner) is pursuing something adjacent and considerably more contested: it plans to recreate missing footage from Orson Welles' The Magnificent Ambersons, the director's follow-up to Citizen Kane, cut and then destroyed by RKO in 1942 against Welles' explicit objections. Fable's method involves filming live-action scenes on physically recreated sets, then using AI to overlay digital recreations of the original cast's faces and voices onto new actors. They are not restoring a film. They are producing one that was never finished and can no longer be made.

These two projects are presented as different things. They are both, at root, the same question: what is a film, and who is authorized to decide?

Film preservation has always been a philosophical problem dressed as a technical one. The question seems practical: how do you stop nitrate from decomposing, vinegar syndrome from corroding acetate, digital formats from becoming unreadable? But underneath the technical question is a prior one: what are you trying to preserve? The answer is not obvious.

The physical artifact is one answer. The film as it exists — degraded, scratched, missing frames, faded color — is a historical object. Its damage is part of its history. The scratches on a print of Metropolis are evidence of a century of screenings, of hands and projectors and storage conditions. To remove them is to remove evidence. The restored print is not the same object.

The director's intention is another answer. What Welles intended The Magnificent Ambersons to be is, in principle, reconstructable — we have the script, production documents, the footage that survives, accounts from people who saw the original cut. Fable is arguing that it can reconstruct the intention. This is a different preservation project from the one that saves celluloid. It is the preservation of an idea about what a film was supposed to be.

The audience experience is a third answer. A contemporary viewer watching a degraded print of Once Upon a Time in China is not watching what audiences saw in 1991. The color has shifted, the resolution has dropped, the soundtrack may have deteriorated. The AI-restored version offers something closer to the original experience — not the original object, not the original intention necessarily, but the original effect, the way it was supposed to look and sound.

These three answers produce three different restoration projects with three different claims to legitimacy. The current AI restoration boom collapses them into a single category called "restoration" and proceeds as if the question of which preservation goal is being served has already been settled.

The phrase "digital hallucination" has entered the criticism of AI restoration, and it is precise in ways that matter. When an AI model fills in a damaged frame, it is not recovering what was there. It is generating a plausible estimate of what might have been there, based on its training data and the surrounding context. The estimate may be visually seamless. It may even be aesthetically appropriate. But it is a fabrication — the AI's inference about what the film would have shown, not a record of what it did show.

In a film where the photography is itself the argument — where Gregg Toland's deep focus in Citizen Kane or Gordon Willis' darkness in The Godfather is doing critical work — an AI hallucination in a damaged frame is not a neutral repair. It is a reinterpretation. The AI has made a cinematographic decision that the original cinematographer did not make. The restored frame may look like what Willis might have shot. It was not shot by Willis.

This is the problem with "preserving original aesthetics." Aesthetics are not separable from the specific decisions made by specific people under specific conditions. The aesthetic of a Bruce Lee film is not just the look — it is the look produced by those cameras, those lenses, those lights, in those locations, by those crew members. An AI that upscales the image, sharpens the edges, corrects the color, and fills in damaged frames has produced a different set of aesthetic decisions. The result may be more watchable. It is not the same film.

The Magnificent Ambersons project makes this problem acute. Fable is not repairing damage to existing footage. It is generating footage that does not exist, of performances that were never given, in scenes that were filmed and destroyed. What they will produce is not The Magnificent Ambersons. It is a film about what The Magnificent Ambersons might have been, made with AI and actors and sets in 2026. Whether that is a worthwhile project is a separate question from whether it is restoration.

The reason this matters beyond archival policy is that film is the primary art form through which a culture records and transmits its visual history. What we preserve is what we remember. What restoration decisions we make determines what version of cinema history survives.

The AI restoration of 100 Chinese martial arts films will likely be the version most people see from now on. The degraded originals will exist in archives and be accessible to scholars. The enhanced versions will circulate — on streaming platforms, in theaters, in home collections. The aesthetic decisions made by AI inference in 2026 will become, for most practical purposes, the aesthetic of those films.

This is not unprecedented. Every restoration is an interpretive act. The color reconstruction of Metropolis, the multiple versions of Blade Runner, the ongoing debates about which cut of Brazil is authoritative — film history is already full of contested restorations and the aesthetic decisions they embed. AI makes the scale of the problem different: the speed, the cost, and the seamlessness of the intervention mean that AI-restored versions will proliferate faster and more completely than any previous restoration technology.

What is new is that the decision-maker is a model rather than a human with documented aesthetic commitments. When Martin Scorsese's Film Foundation oversees a restoration, there is a person accountable for the interpretive choices made. When an AI fills in damaged frames, the choices are made by a system trained on a corpus of other films, optimizing for visual coherence by criteria that are not fully transparent. The hallucination is plausible. Its logic is opaque.

The China Film Foundation says it intends to preserve the original aesthetics. It is worth asking: which aesthetics, exactly? The aesthetics of the films as they were shot, or the aesthetics of the films as an AI model, trained on decades of cinema, believes they should look? These are not the same thing. The gap between them is what restoration destroys.