In August 2025, the 1939 film The Wizard of Oz was projected onto the immersive screen at the Las Vegas Sphere, expanded and transformed using artificial intelligence. Filling a giant screen required three distinct actions, despite being marketed under the same label. The first was to increase the resolution of the original image: this is restoration. The second was to extend the sets beyond the frame captured by the camera in 1939, relying on archival materials: this is reconstruction. The third was to generate performances for characters present in the scene but never captured by the camera: this is invention. Three very different operations, yet a single word for all of them: “restoration”.

This confusion of terms is what I intend to dismantle, for it is not innocent. Artificial intelligence now permits intervention upon works of the past with a verisimilitude that previously did not exist, and the language used to present these interventions tends to equate them all under a reassuring term. Restoration implies rescue, care, and the restoration of life. Yet, between recovering what existed and fabricating what never existed lies an abyss, and the machine does not erase it: what erases it is the manner in which it is narrated.

Let us classify these operations, which range from minimal to maximal conjecture. Restoration involves stabilising, cleaning, or reintegrating a loss of something whose prior existence can be justified. Reconstruction involves formulating a hypothesis supported by documents, copies, or sketches. Completion involves continuing where the author left the work unfinished. Inventing in style involves generating content without evidence that such content existed or that the author would have intended it. AI may intervene in all four. The problem arises when the fourth is disguised under the name of the first.

An example of the honest extreme was provided in 2025 by an MIT researcher: they applied a microscopically thin, computer-calculated printed film to a damaged 15th-century oil painting, which visually reintegrates the lost areas. The decisive factor is not the technology, but two conditions that were met: the film is separable, meaning it can be removed without touching the original, and an exact digital record of what was added remains. Reversibility and documentation. A century from now, anyone returning to the work will know precisely what is the hand of the painter and what is that of the restorer. This is intervening with one's cards face up.

The opposite extreme is occupied by cases where the machine makes the dead speak. Since 2024, a museum dedicated to Dalí has allowed visitors to ask questions of a recreation of the painter, which responds with its synthesised voice to phrases Dalí never uttered. A company has licensed the voices of Judy Garland or Laurence Olivier so that, now deceased, they may read texts they never read. In the Sphere’s Oz, dead performers “act out” shots they never filmed. In all these cases, there is often authorisation from a museum or heirs. What cannot exist, and what is left unsaid, is the consent of the person who is no longer here. Heirs may cede a right; they cannot lend a will. And a voice that says what its owner never said is not an archive: it is high-quality ventriloquism.

There is one case that clarifies the difference better than any argument. In 2023, a social media user took Keith Haring’s Unfinished Painting and “completed” it with AI. Haring had deliberately left it half-finished in 1989, with a large empty area, as an image of the lives cut short by AIDS, his own among them. Filling that void is not finishing a work: it is erasing its meaning, which resided precisely in what was missing. There was no museum or foundation behind it; it was a private gesture, but for that very reason, it lays the operation bare: the machine can produce what is missing, and sometimes what is missing was the work itself.

None of this originates with artificial intelligence. When Mozart died leaving the Requiem unfinished, his widow commissioned Süssmayr to complete it, and that version became the standard, even though it coexists with others today: completing it did not recover a single original; it established a debatable version. When Puccini died without finishing Turandot, another composer wrote the ending. And at the premiere in 1926, Toscanini did something that remains the most honest response to this entire matter: he conducted up to the final bar written by Puccini, put down his baton, and stopped the performance there, so that the audience knew exactly where the author ended and another hand began. He did not hide the seam. He pointed to it.

That gesture of Toscanini is the criterion that AI renders urgent. It is not a matter of prohibiting the machine from restoring, reconstructing, or even imagining: it is a matter of not calling all four things by the same name. Each intervention should state from which hand each part originates, as with the MIT reversible film and Toscanini’s halted baton. The alternative—presenting invention with the voice and visage of the author and terming it restoration—restores nothing to the past: it fabricates a new past and requests that we confuse it with the one that existed. The problem is not that the machine completes what is missing. It is that we learn not to notice the difference.

On open conversation

This text is not opposed to the use of artificial intelligence in conservation, which may be valuable and reversible, but rather to the language that gathers under the word “restore” operations ranging from recovering what existed to inventing what never existed. The criterion I propose does not examine the medium, but rather the honesty regarding what is being done and from whose hand each part originates. I accept verified cases as certain and distinguish the documented from that which is not recorded, particularly regarding consent and rights. Should anyone wish to intervene from the fields of conservation and restoration, museology, film or music history, or artistic practice, this notebook remains open.

Sources

The Wizard of Oz at Sphere (Las Vegas, 28 August 2025), Sphere Studios with Google Cloud/DeepMind, Magnopus and Warner Bros. Discovery: super-resolution, outpainting and performance generation.

Now and Then, The Beatles (2 November 2023): AI source separation to isolate John Lennon’s voice from a demo; statements by Giles Martin.

Ask Dalí, The Dalí Museum (2024); synthetic voices licensed by ElevenLabs (2024): posthumous generative representations.

Alex Kachkine (MIT), reversible printed restoration mask, published in Nature (June 2025).

Non-institutional intervention on Keith Haring’s Unfinished Painting (social media, 2023), as a counter-example.

Conservation doctrine: Venice Charter (ICOMOS, 1964), art. 9 and 12; ICOM-CC, Terminology for Conservation (New Delhi, 2008); AIC, Code of Ethics. Precedents: Mozart’s Requiem (Süssmayr, 1792) and Puccini’s Turandot (Alfano; premiere conducted by Toscanini, La Scala, 25 April 1926).