Labelling AI-Altered Archives: A Disclosure Standard for Digital Restoration

Why restored archives need a disclosure standard

Digital preservation increasingly involves material that cannot be presented conveniently in its surviving form. Old screenshots may have unreadable interface text, compressed recordings may obscure small details, and chat exports may contain encoding errors. AI-assisted processing can improve access to these records, but an attractive derivative can also acquire an authority that its source never possessed. Readers need to know where the historical record ends and where processing begins.

This is particularly important for virtual worlds, where apparently minor details can carry historical meaning. A reconstructed username, a sharpened character emblem or a corrected line in an old conversation might influence how a community is remembered. Disclosure should therefore be treated as part of archival description, not as a disclaimer added after publication. Its purpose is to explain the relationship between an original record and every derivative made from it.

• Identify the original source and its archival reference.
• Describe each AI-assisted transformation.
• Separate visible evidence from inferred detail.
• Record uncertainty and subsequent corrections.

A four-part model for describing altered material

A useful disclosure record has four connected elements: source, operation, derivative and interpretation. The source identifies the material before processing. The operation describes what happened to it. The derivative identifies the resulting file or transcript. Interpretation records any editorial explanation that goes beyond the material itself. Keeping these elements separate prevents a restoration note from becoming an unsupported historical claim.

For example, a 2004 screenshot might survive as a compressed image with several unreadable labels. Its catalogue record should identify the original file, while a separate record describes an AI-assisted enlargement. If a curator subsequently supplies a likely interface label using another surviving screenshot, that addition belongs in an editorial annotation. It should not silently become part of the restored image or be presented as legible source content.

• Source: original filename, format, date if known and provenance.
• Operation: processing purpose, tool and relevant settings.
• Derivative: identifier, format and creation date.
• Interpretation: human additions, uncertainty and supporting evidence.

Disclosing changes to chat logs and oral histories

Transcribing old chat logs requires particular care because a transcript can look more definitive than the damaged export from which it was prepared. Record whether text was recovered from an existing file, converted from another encoding, corrected manually or reconstructed from incomplete fragments. If an AI tool proposes missing words, retain the original passage and mark the proposed reading separately. A grammatical sentence is not evidence that every word appeared in the historical conversation.

Oral history recordings from online communities present a different set of difficulties. Automated transcription can confuse usernames, technical vocabulary, accents and overlapping speakers. A disclosure record should distinguish the automatic transcript from a human-corrected version and identify passages where attribution remains uncertain. Participants' preferred names and any agreed restrictions on publication should also be recorded in the archive's access documentation.

• Retain the original chat export or audio recording.
• Mark uncertain words with a consistent convention.
• Distinguish speaker identification from speaker inference.
• Document human corrections and participant-approved amendments.
• Keep restricted personal information out of public derivatives.

A practical disclosure label for archive catalogues

A disclosure label should be short enough to appear alongside an individual file, but detailed enough to direct a reader towards its processing record. The label is an entry point, not a replacement for a complete technical description. A catalogue can use a small number of standard statuses, provided their meanings remain stable across collections and are explained in an accessible policy.

For example, an enlarged screenshot might carry the label 'AI-assisted enlargement; original preserved separately; fine interface details may be reconstructed'. Its full record would then identify the source image, processing date, tool, relevant settings and the person responsible for the editorial decision. The same approach works for transcripts, audio clean-up and image colourisation, with the wording adapted to the actual transformation.

• Original: unchanged source material.
• Processed: technical alteration without intentional reconstruction.
• AI-assisted: machine-generated enhancement or transcription is present.
• Reconstructed: one or more elements have been inferred or recreated.
• Composite: material from multiple sources has been combined.

Preserving provenance and correction histories

Disclosure becomes unreliable when processing records are stored separately from the files they describe. Every derivative should have a stable identifier and a direct relationship to its source. Where a collection contains several processing stages, record the sequence. For example, an original recording may produce a cleaned audio access copy, an automatic transcript and a corrected transcript. Each item should point to the relevant predecessor without implying that the later version supersedes the original.

Corrections also need a history. If a researcher identifies a username that was previously transcribed incorrectly, preserve the earlier transcript, record the amendment date and explain the evidence supporting the revision. This matters particularly when archive material has already been quoted in articles, research papers or community histories. A transparent correction record allows later readers to understand why two published versions differ.

Making disclosure useful to readers and future archivists

A disclosure policy should be published in the same place as the archive's catalogue guidance. Technical descriptions need not be identical for every audience: a general reader may need a concise label, while an archivist needs the complete processing history. Both should be able to identify the original source and understand whether a particular detail was observed, transcribed, corrected or inferred.

The standard should also cover material that is shared outside the archive. Downloaded images, embedded video and excerpts reproduced in articles can become detached from their catalogue records. Include the disclosure label in accompanying captions or metadata where practical. Before releasing a new derivative, confirm that its identifier, source relationship and processing description travel with the file. That is the next practical step towards an archive whose restorations remain understandable after its original custodians have moved on.

Questions readers ask