# The Past, Remastered 2026-09-07 · Somerset County, New Jersey · Culture AI can make a street filmed in 1911 move like it was shot yesterday. It can color the coats, sharpen the faces, smooth the footsteps and manufacture the frames between them. The result may feel more real. That does not mean more of it is true. AI can make a street filmed in 1911 move like it was shot yesterday. It can color the coats, sharpen the faces, smooth the footsteps and manufacture the frames between them. The result may feel more real. That does no… --- There are specific kinds of video built to stop the scroll. The title usually promises some impossible recovery: New York in 1911 in 4K. Paris in 1896 at 60 frames per second. The streets of old London “brought back to life.” The footage moves smoothly. Faces appear startlingly clear. Black-and-white becomes color. Hooves strike pavement with newly added sound. A person who lived a century ago turns toward the camera and, for a few seconds, seems less like an archival figure than somebody standing on the other side of very clean glass. The emotional trick works because the distance collapses, and the historical problem begins for exactly the same reason. Artificial intelligence has become extraordinarily good at modernizing old images. Upscaling systems can infer texture where the source contains little detail. Colorization models can guess likely colors from objects they recognize. Frame-interpolation software can analyze two surviving frames and construct images to place between them, turning the uneven motion of early cinema into the fluid cadence contemporary viewers expect. Other tools can remove scratches, stabilize jitter, reconstruct damaged areas and even generate sound for images that were recorded silently. Some of those techniques are restoration. Some are enhancement. Some are invention but the most consequential change may be that the finished video rarely looks interested in telling us which is which. Restoration has a much stricter meaning Film archives have spent decades arguing over what it means to restore a moving image because the answer determines what future audiences will believe the original object was. The International Federation of Film Archives revised its Code of Ethics in 2025. Its standard is conservative by design: archives should preserve the integrity of their collections, document interventions, reduce damage caused by age or handling, respect the characteristics of the original material and resist deliberately modifying a work during restoration.1 FIAF's technical guidance gets more specific. Creating entirely new frames through interpolation or artificial intelligence produces what the organization calls ahistorical frames - images that never existed in the original. Artificial sharpening, smoothing and other forms of enhancement can likewise move beyond recovery because modern viewers have been trained to mistake sharpness for historical image quality.2 That distinction matters because “4K” itself tells us almost nothing about authenticity. An archive can scan a 35mm film element at 4K or higher resolution and capture information genuinely present in the surviving material. A consumer AI system can also take a low-resolution digital copy, enlarge it to a 4K file and generate plausible detail the source never contained. Both videos can be labeled 4K, one is a high-resolution record of an artifact, the other is partly a prediction. The train that arrived twice The modern genre of AI-remastered history announced itself spectacularly in 2020 when Denis Shiryaev published an enhanced version of the Lumière brothers' Arrival of a Train at La Ciotat, the landmark 1890s film of passengers meeting a train at a French station. Shiryaev used neural-network tools to upscale the image and interpolate enough new frames to present the film at 60 frames per second. He added sound as well. The result circulated widely because the old cinematic barrier seemed to disappear. People no longer moved with the familiar irregular speed associated with early film. They moved like people recorded by a modern camera.34 But that fluency was itself synthetic. The original was not filmed at 60 frames per second. The intermediate moments were calculated by software looking at what came before and after them. The algorithm was not uncovering hidden frames. It was drawing new ones. Shiryaev and others working in the form have generally been clearer about this distinction than the viral ecosystem surrounding their work. In reporting on the trend, WIRED noted that colorization systems such as DeOldify guess colors from learned associations, while upscalers and interpolation tools construct visual information that may be convincing without being historically exact.5 The enhancement can therefore make the footage easier to watch while making its provenance harder to see. The train really arrived. Not every image of its arrival did. We have been arguing about this longer than AI The ethical problem did not begin with generative AI. Peter Jackson's 2018 documentary They Shall Not Grow Old used First World War footage from the Imperial War Museums and audio from BBC archives, then cleaned, reframed, speed-corrected, colorized and transformed portions of the material with contemporary digital production techniques. Lip readers helped reconstruct speech and sound was designed for the once-silent footage.6 Jackson described the goal as reaching through the “fog of time” so the soldiers could regain their humanity rather than remain distant figures constrained by the technical limitations of early cinema.6 For many viewers, it worked magnificently. The soldiers suddenly seemed young in a way grainy black-and-white footage had taught generations not to see. BFI writers defended the film as an emotionally powerful renegotiation of archival material and argued that the familiar monochrome presentation was not itself some neutral window onto the Western Front.7 Archivists and historians also objected. Critics argued that color, reframing, smoothing and newly constructed sound could create the impression that viewers were finally seeing the war “as it really was,” when they were actually seeing a carefully authored modern interpretation of surviving evidence.5 That argument remains unresolved because both sides are right about something important. Old film technology creates distance that historical subjects themselves did not experience. Removing that distance requires us to add information they did not record. The scratch is not the past, but it belongs to the artifact There is an easy answer to some restoration questions. A scratch made during decades of handling was not part of a filmmaker's original intention. Mold is not authorship. A torn frame is not sacred because it is old. Archives repair physical damage precisely because preservation is not the same as freezing deterioration in place. But historical film contains more than depicted events. It also contains the history of the medium that depicted them. Film grain, tinting, exposure, frame rate, lens characteristics, instability, damage and projection practices all tell us something about how an image was made, copied, circulated and seen. BFI preservation work now includes multispectral scanning specifically because archivists are trying to understand original dyes and recreate the appearance of early color processes with evidence from the material itself.8 That is almost the opposite of automatic colorization. The archival question is: What evidence survives? The synthetic question is: What would look plausible here? Those can occasionally lead to the same answer. They are not the same method. Color is where plausibility masquerades as knowledge Colorization makes the epistemic problem unusually visible because black-and-white footage contains no hidden RGB values waiting to be extracted. A machine can recognize sky, grass, brick, military uniforms or human skin and assign colors statistically associated with those categories. A human colorist can research uniforms, paint samples, photographs and location records to make more informed decisions. Either way, the final color is an interpretation unless reliable evidence exists for that particular object. Scholars have increasingly treated digital colorization as a form of remediation rather than neutral restoration. Recent work on archival colorization argues that adding color changes how viewers understand the image and can obscure the historical conditions of the source.9 A 2025 study of AI-assisted restoration of early footage from China asks the same uncomfortable question: when scratches, flicker, grain and other material traces are erased while color and intermediate frames are added, is the viewer still encountering the same historical document or a new audiovisual object built from it?10 The danger is not merely getting a coat wrong. It is creating the feeling that no guessing occurred. The synthetic past is expanding beyond restoration By 2026, the problem has moved beyond people enhancing real archival footage. AI-generated “historical” videos now circulate on TikTok, YouTube, X and other platforms depicting scenes for which no camera footage exists at all: imagined views of ancient Pompeii, fabricated World War II scenes, invented surveillance-style footage and reconstructed disasters presented in the visual grammar of documentary evidence. Deutsche Welle reported this summer that such videos are popular partly because they satisfy audience expectations about how the past should look, often reinforcing familiar historical stereotypes in the process.11 This is a different category from restoration, but the aesthetics are beginning to converge. A genuinely filmed street from 1910 may be colorized, interpolated, sharpened and supplied with synthetic sound until it looks like modern video. A completely generated street from 1910 may be prompted to include film grain, lens flaws and period clothing until it looks archival. One begins with evidence and adds invention. The other begins with invention and adds the appearance of evidence. On a phone screen moving at algorithmic speed, the distinction can become a caption. Artificial nostalgia has a house style The most successful synthetic history often does not look historically messy. It looks beautifully legible. Faces are clear. Motion is smooth. Colors are rich but tasteful. Streets are busy enough to feel alive. People behave recognizably. Weather has cinematic atmosphere. The past arrives with the visual grammar of prestige television. That consistency is part of what might be called the artificial nostalgia industry: not one company or technology, but an ecosystem of creators, enhancement tools, social platforms and audience expectations that rewards historical imagery for feeling emotionally immediate. The incentive is obvious. A century-old film that still looks like a century-old film asks viewers to adjust themselves to another medium. A remastered version adjusts the medium to the viewer. The problem is that friction can contain information. The irregular motion reminds us that the camera was hand-cranked. The monochrome image reminds us what the recording process could preserve. The damaged edge reminds us that this object survived materially rather than descending intact from history. A flawless past can become a past with its manufacturing marks removed. The memory problem is not hypothetical Research into AI-edited personal imagery suggests that synthetic realism does more than change aesthetics. A 2025 MIT Media Lab study tested whether altered images and AI-generated videos could affect memory. Participants exposed to AI-edited visuals reported significantly more false recollections than controls; AI-generated videos made from edited images produced the strongest effect and higher confidence in memories that were not accurate.12 The study involved personal-memory scenarios rather than archival history, so it does not prove that watching colorized 1910 footage will implant false historical memories. It does establish something more basic: realistic synthetic imagery can change what people remember and how confident they feel about remembering it. Memory scholar Andrew Hoskins argued in a 2026 review that generative AI is transforming collective memory by altering how societies represent, access, sanitize, amplify and forget the past.13 That is a much larger issue than whether an old film looks nicer after denoising. Visual evidence has always shaped collective memory. We are entering a period when the evidence can be aesthetically improved faster than most viewers can learn what was added. History has always been mediated There is a temptation here to romanticize the original archive. That would be its own mistake. Historical images have never been transparent reality. A camera records where somebody pointed it. Early film stocks captured some colors and skin tones differently from others. Colonial archives were built by institutions with power. Newsreels were edited. Photographs were staged. Captions misidentified people. Governments censored material. Surviving collections reflect what someone considered worth preserving. The original object is not automatically the truth. It is the evidence. That is exactly why provenance matters. The American Historical Association warned in 2026 that generative AI can produce simulations of the past but cannot create historical images in the meaningful archival sense. A generated image can illustrate an argument, provoke a question or make an artistic point. It cannot retroactively become evidence that a camera captured an event.14 The distinction sounds obvious when stated plainly. Modern image systems are very good at making it stop feeling obvious. The answer may be two versions, not one There is nothing inherently unethical about an artist making an enhanced historical film. Colorization can make viewers curious. Correcting projection speed can prevent early film from turning ordinary people into accidental slapstick. Stabilization can reveal gestures previously difficult to see. Careful reconstruction can help audiences understand damaged or incomplete material. Synthetic sound can create a compelling interpretive experience when everyone knows it is synthetic. The problem begins when interpretation inherits the authority of the source. A useful standard already exists in archival ethics: document the intervention, preserve the original and make the nature of the restoration visible. FIAF requires restoration decisions to be documented. Libraries, archives and museums are now pushing further, developing provenance practices capable of identifying when AI has affected a digital object during its lifecycle. A 2026 Library of Congress initiative urged cultural institutions to adopt stronger content-authenticity systems so AI-affected collection material remains transparent and verifiable.115 The practical answer may therefore be less dramatic than banning enhancement. Keep the evidence. Label the interpretation. Do not make the second one overwrite the first. A past optimized for the present The artificial nostalgia industry is powerful because it solves a real problem. The past is difficult. Its images are damaged, incomplete and technologically strange. Its streets do not move at our frame rate. Its photographs lack the colors we expect. Its recordings contain noise. Its documents require context. Dead people do not explain themselves in contemporary visual language. AI can reduce that distance in seconds. And sometimes reducing the distance is generous. A teenager who would never sit through flickering footage from 1906 may stop for the remastered version and discover that the anonymous people crossing the frame had ordinary faces, impatient gestures and places to be. But history should resist us a little. The past was not a lower-resolution version of the present waiting for software to finish rendering it. People did not live in black and white, but neither did they live in AI's best guess at Kodachrome. A missing frame is missing. A color not recorded is unknown. Silence in a silent film is not an invitation for a model to remember the sound. There is a profound difference between helping us see the evidence and helping the evidence resemble us. Artificial nostalgia erases that difference beautifully, that is why it works and why the label matters. Reporting & references International Federation of Film Archives (FIAF), Code of Ethics, revised April 2025. SourceFIAF Technical Commission, Digital Statement: recommendations on digital restoration and manipulation.Ars Technica, “Someone used neural networks to upscale a famous 1896 video to 4K quality,” Feb. 4, 2020.Engadget, “How AI helped upscale an antique 1896 film to 4K,” Feb. 4, 2020.WIRED, “YouTubers are upscaling the past to 4K. Historians want them to stop,” 2020; and “AI Magic Makes Century-Old Films Look New,” 2020.British Film Institute, announcement and background on Peter Jackson's They Shall Not Grow Old, 2018.BFI Sight and Sound, “They Shall Not Grow Old review: Peter Jackson brings controversial colour to WWI footage,” 2018.BFI National Archive, Scan2Screen and multispectral work on historical film color and projection, July 2025.Tom Livingstone, “The Colourised Archive,” in Hybrid Images and the Vanishing Point of Digital Visual Effects, Edinburgh University Press, 2024.Kaiqi Zhu, “Reanimating the Archive: AI and the Afterlives of China’s Early Images on Digital Platforms,” Baltic Screen Media Review, 2025.Deutsche Welle Fact Check, “Why are ‘historical’ AI videos so popular?”, July 19, 2026.MIT Media Lab, “Synthetic Human Memories: AI-Edited Images and Videos Can Implant False Memories and Distort Recollection,” CHI 2025.Andrew Hoskins, “AI & collective memory,” Current Opinion in Psychology, Vol. 67, February 2026.American Historical Association, “What Altering Historical Images Can Teach Us About History,” March 17, 2026.Library of Congress, “Content Authenticity and Provenance in the Age of Artificial Intelligence: A Call to Action for the Libraries, Archives and Museums Community,” April 14, 2026.Peter Geimer and Luca Beisel, “What Is the Color of the Past? The Truth of the Archive and the Truth of Simulation,” International Journal for Digital Art History, 2021. --- ProbleMattic is written and maintained by Matthew Kulcsar, a software engineer, project manager, technologist, platform builder, emergency-services-trained helper, grandfather, and lifelong collector of broken systems, odd behaviors, and useful nonsense.