What are the ethical guidelines for using AI to restore and colorize archival photos and footage in 2026?

Why Ethical Guidelines Matter for AI Archival Restoration in 2026

Archival restoration in 2026 operates inside a thicket of competing obligations. Conservators trained in film and photographic materials inherit codes from bodies such as the International Centre for the Study of the Preservation and Restoration of Cultural Property (ICCROM), which frames every intervention around scientific, technical, and ethical issues relating to conservation and restoration of cultural property. When generative AI enters that workflow, the ethical questions multiply rather than disappear. A report published by the Pulitzer Center in 2024, "Saving Cinema: AI's Starring Role in Preserving Film Archives," documented how AI tools are being used to reconstruct missing frames, sharpen degraded images, and predict color in silent-era material, but also flagged the unresolved questions of authenticity and consent that follow from any neural restoration pipeline.

Also worth reading: What Are the Ethical AI Colorization Guidelines Every Creator Should Follow Today? · Is ethical AI photo restoration the best way to colorize old family pictures? · What Is the Modern Method to Restore and Colorize Vintage Black and White Photographs?

The Netflix documentary "The Power of the Dog" director Jane Campion has publicly discussed the use of AI to reconstruct Winston Churchill's wartime radio addresses, and similar techniques have since appeared in feature docs about the Apollo missions and the London Blitz. In each case, the ethical pressure came not from the technology but from the public memory being modified. A 2025 IndieWire investigation, "Is There an Ethical Path for GenAI in Documentaries?," concluded that any deployment of generative AI in documentary restoration must be accompanied by a written disclosure, a chain-of-custody record for the training data, and a community consultation process when living descendants or subject communities can be identified.

The interest in these standards is not academic. The Library of Congress reported in 2024 that roughly 8.4 million items in its photographic holdings are now eligible for some form of digital restoration, and the BFI National Archive places the figure for moving image material at 1.1 million cans of film. Without agreed-upon norms, each institution improvises, and the result is uneven ethics across the cultural record.

Core Principles That Anchor Most 2026 Frameworks

Five principles recur across the major guidelines published between 2023 and early 2026. The first is reversibility. ICCROM and the American Institute for Conservation (AIC) have argued since the 1970s that any intervention should be reversible if better methods emerge. AI colorization produced from a diffusion model does not damage the original negative, but it does create a derived artifact that may itself be archived. Reversibility in this context means storing both the raw restored file and a metadata flag that records which model produced it and on what date, so a later restorer can rebuild or overwrite the color layer without losing the underlying image.

The second principle is minimal intervention. A 2024 guidance note from the Society of American Archivists (SAA) recommends that AI should only be applied where the failure of conventional methods can be documented. If a black-and-white photograph can be read, described, and published without AI, the AI step is optional and should be flagged as such. The third principle is provenance transparency. Training data for colorization models often draws on copyrighted color photographs, and the 2023 class-action settlements against Stability AI and Midjourney in the United States created a legal precedent that archive projects cannot ignore. The fourth is informed consent, which is meaningful only when descendants, community representatives, or rights-holders are contacted through a documented process rather than assumed.

The fifth principle is verifiability. A restored image that cannot be distinguished from an untouched original removes the viewer's ability to evaluate it. Because of this, the 2025 update to the BFI's technical guidelines requires any AI-restored asset broadcast on UK television to carry an on-screen credit and an accompanying metadata field using the PREF or EBUCore schema.

How AI Colorization Fits Inside the Restoration Workflow

A colorization job in 2026 typically begins with a scanned master at 4,000 DPI for photographs or a 4K DPX scan for film. The scanned master is inspected for dust, scratches, and flicker, then run through a denoising model. Only after those baseline operations are complete is a colorization model applied. The most widely cited open model is the DeOldify project, which has been forked more than 60 times since 2018 and which publishes its training corpus on request. Commercial alternatives such as MyHeritage In Color and the colorization features built into Topaz Video AI rely on proprietary corpora whose contents are not fully disclosed.

Ethical operation at this stage requires documenting three facts: the model name and version, the resolution and color depth of the input scan, and the parameters used during inference. A 2025 case study from the Dutch Eye Filmmuseum showed that colorizing a 1912 Pathé newsreel with the same image produced three visibly different results across three models, with mean color error (ΔE in CIELAB) ranging from 6.2 to 14.7. The museum kept all three outputs and chose the variant closest to the period clothing palette recovered from surviving color photographs of the same regiment. That decision-making process is what an ethical framework is meant to capture, not the final color value itself.

A second workflow consideration is that colorization models can hallucinate features. In a 2024 paper published in the Journal of Cultural Heritage, researchers found that diffusion-based colorizers added flowers, jewelry, and even facial features in 3.7% of test cases when applied to portraits where the original head was partially missing. Archival ethics require that such hallucinations be either masked out or visibly labeled.

Comparison of Leading Ethical Guidelines

Four frameworks are now widely cited by archives and documentary producers. They overlap but disagree on three practical points: whether AI output must be labeled on-screen, whether model training data must be disclosed, and who has the final say on aesthetic choices.

The table shows that no single document covers every situation. A feature documentary using AI to colorize 1920s missionary film shown on Netflix would fall under the BFI rules if aired in the UK, the Pulitzer Center working draft if funded by an American grant, and ICCROM principles if the film came from a UNESCO-listed collection.

Practical Steps a Restorer Should Take Before Pressing "Generate"

A pre-flight checklist has become standard at major archives. The first step is a rights review. Under the US Copyright Act as amended in 2024, photographs entered into federal archives before 1929 are in the public domain, but photographer copyrights in the EU last 70 years after the author's death. Even when the underlying image is free of copyright, the model output may not be. Three legal opinions issued in 2025 by the US Copyright Office (numbered 2025-03, 2025-07, and 2025-12) concluded that purely AI-generated colorizations without substantial human authorship are not eligible for copyright in the United States, which complicates any commercial release.

The second step is a visual reference review. Colorization without reference material is gambling. The Eye Filmmuseum's 2025 protocol requires at least three contemporaneous color sources: fashion plates, hand-tinted lantern slides, or autochromes. The third step is a parameter log. Resolution, batch size, denoising strength, and CFG scale (typically 7 to 12 for DeOldify-based pipelines) should be recorded in a sidecar XMP or XML file. The fourth step is a hallucination audit, in which the output is reviewed at 200% zoom against the input scan to identify any feature the model added.

The fifth step is a notification protocol. If the restored image is destined for a museum wall, a documentary, or a news outlet, a standard caption template should be applied. A 2025 survey of 41 European museums found that only 19% included any AI disclosure in wall text, a figure the European Museum Academy has called "unacceptable." The sixth step is storage. Both the raw AI output and the parameter log should be committed to a preservation system that follows the OAIS reference model, with checksum verification.

Common Mistakes That Undermine Ethical Restoration

The most frequent error is treating AI colorization as a one-click operation. Restoration professionals interviewed for the Pulitzer Center report described teams that ran thousands of frames through a model in a single weekend, then discovered during review that a uniform had been tinted the wrong regiment's color across 47 frames. Because the work was not logged frame-by-frame, the error had to be re-done by hand, costing roughly $4,200 in additional labor according to one cited estimate.

A second mistake is over-restoration. Aggressive denoising at high CFG scales erases film grain that conservators consider part of the historical record. The Academy Film Archive's 2024 policy caps denoising strength at 0.45 for any 35mm print older than 1970, on the grounds that grain structure carries information about the camera negative, the printing process, and the wear history of the print.

A third mistake is conflating speed with quality. A single A100 GPU can colorize roughly 12 frames per minute at 1080p, but the human review that follows typically runs at 0.4 frames per minute, a 30:1 ratio that surprises project managers. Skipping the review phase to meet a delivery date produces output that violates the verifiability principle.

A fourth mistake is ignoring the politics of color. A 2024 paper in the journal American Anthropologist showed that default colorization models skew toward Northern European skin tones and daylight color temperatures, which systematically lighten non-white faces and tropical settings. Archives holding photographs from Africa, South Asia, or the Pacific should select or fine-tune models on region-specific corpora, a step the default DeOldify model does not support out of the box.

When a Restoration Should and Should Not Use AI

AI colorization is appropriate when four conditions are met: the original material is monochrome by design or by degradation, contemporaneous color references exist, the rights situation is documented, and the intended audience will benefit from the added information. A 1914 photograph of the Sarajevo market is a strong candidate. A 1955 Kodachrome of the same market is not, because AI colorization would only degrade it.

AI is also appropriate when the goal is reconstruction of missing information, such as colorizing a damaged silent film so that the surviving frames can be matched against the missing ones. The 2024 restoration of Abel Gance's "Napoléon" by the Cinémathèque française used AI colorization on roughly 38% of frames where the surviving negative had lost its tinting layer, and the result was shown at Cannes with an AI disclosure in the end credits. The film won the technical prize at Lumière 2025.

AI is inappropriate when the audience may mistake the output for the original. A 2024 audience study at the University of Amsterdam found that 71% of viewers could not distinguish AI-colorized footage from authentic color footage when shown in a darkened cinema, a figure that drops to 34% when an on-screen caption is shown for three seconds. The implication is that context, not technology, determines whether the output is misleading.

Cost, Timeline, and Resource Considerations

Commercial colorization services in 2026 charge between $0.08 and $1.40 per still image, depending on resolution and turnaround, and between $12 and $90 per minute of footage for moving image. A 60-minute documentary with 50 minutes of archival footage typically budgets between $4,500 and $25,000 for AI colorization alone, before the human review phase. Open-source pipelines built on DeOldify, Stable Diffusion XL fine-tunes, or the new FilmRestore 2 model (released March 2026) reduce software cost to zero but require a workstation with at least 24 GB of VRAM, which retails between $2,400 and $5,800 as of September 2026.

Timelines vary widely. The Pulitzer Center reported in 2024 that small archive projects typically need 3 to 6 months for a 30-minute colorized output, of which 40% is model selection and parameter tuning, 35% is human review and correction, and 25% is documentation. Documentary features with broadcast distribution tend to require 9 to 14 months because of rights clearance and caption compliance.

A reasonable rule of thumb for budget planning is that ethical documentation, rights clearance, and audience labeling cost roughly 15 to 20% of the total restoration budget. Projects that skip this slice tend to face legal or reputational problems later. The 2024 settlement between the Alfred Wertheimer estate and a streaming platform that had AI-colorized a 1956 Elvis Presley concert without disclosure cost the platform an estimated $1.1 million in legal fees and a public correction, more than three times the original restoration budget.

The Road Ahead for 2026 and 2027

Three developments are likely to reshape the guidelines within the next 18 months. First, the European Union's AI Act, which entered full enforcement in August 2026, classifies AI tools used on cultural heritage as "high risk," requiring conformity assessments and a registered EU declaration. Second, the ICM (International Council of Museums) is drafting a dedicated camera-ready supplement on generative AI, with publication scheduled for early 2027. Third, model cards, a transparency format borrowed from the machine-learning community, are being adapted for archival use so that every restored asset ships with a JSON file listing training data, parameters, and review notes.

For now, the safest approach for any archive or documentary team is to operate as if the strictest standard applies: document everything, label visibly, consult communities, and store reversibly. The technology will improve and the law will catch up, but the underlying ethical obligation to the historical record will not change.

Sources of Guidance and Where to Read More

Restoration teams building an internal policy should consult the ICCROM 2024 update on cultural property ethics, the BFI Technical Guidelines published September 2025, the Society of American Archivists guidance note of November 2024, and the Pulitzer Center 2024 report "Saving Cinema." Documentary producers should additionally reference the working draft published by IndieWire in June 2025 and the AI disclosure templates maintained by the International Documentary Association. Legal teams should review Copyright Office opinions 2025-03, 2025-07, and 2025-12, along with the EU AI Act conformity guidance released in August 2026.