The archival profession is facing a technological shift that differs from the digitization wave of the late twentieth century. While previous tools helped archivists scan, store, and search, generative AI introduces the ability to create. Generative AI refers to models, specifically Large Language Models (LLMs), capable of generating new text and summaries from existing records. This technology differs from rule-based automation, which follows logic, and predictive analytics, which forecasts trends.
The Society of American Archivists (SAA) Code of Ethics emphasizes professional judgment, transparency, and authenticity of records. When archivists introduce a machine that can summarize a complex 19th-century diary or synthesize a collection of emails, they move the role of archival description into the realm of interpretation. This epistemological shift changes how historians know what they know about the past.
Key Differences Between Description and Interpretation
When an LLM processes archival records, it uses patterns to generate human-like prose. For a researcher facing a hundred unindexed boxes, an AI-generated summary of a finding aid sounds like a miracle of accessibility. However, summarization is reductive. Unlike editorial practices or annotated editions where scholars make subjective choices, AI summaries present themselves as a neutral, authoritative voice.
The risk is the flattening of nuance. Archives are repositories of ambiguity, containing contested meanings and contradictory accounts. Generative AI seeks to resolve patterns and smooth over the productive friction that historians seek to uncover. It also distinguishes poorly between a paraphrase, which restates content, and an interpretation, which assigns meaning. Even more concerning is speculative reconstruction, in which the model fills in gaps.
The Specter of Hallucination
A hallucination occurs when a model generates information that is factually incorrect but highly convincing. In an archives, where provenance and original order are sacred, a fabricated date is a fundamental violation of the record’s integrity. If an AI attempts to fill lacunas, it invents history.
This process becomes problematic when examined through the concept of archival silences. Most collections are biased toward the powerful, leaving marginalized communities as shadows in the documentation. When AI extrapolates from these incomplete records, it risks amplifying those silences or, worse, hallucinating perspectives that further colonize the narrative. Applying AI to collections without consent or regard for sensitivity risks repeating historical harms.
The Challenges of Derivative Summaries
The legal landscape is equally complex. When an AI generates a summary or a finding aid, is the output a derivative work? If archivists train AI on copyrighted materials, the resulting summaries may expose the institutions to intellectual property risks. Who is accountable when the AI produces a summary that is libelous, culturally insensitive, or factually wrong?
Archives must implement transparency requirements, which include documenting the prompts used to generate text, the model versions, and the system configurations. Archivists must uphold disclosure standards, labeling AI-generated content just as museums display disclaimers on reconstructed artifacts.
Redefining Archival Literacy
As AI becomes an interface for discovery, archivists must rethink how they teach archival literacy. Researchers need to understand that the narrative they encounter might be an AI-generated contextualization that influences their interpretation before they consult the primary source. AI aids discovery, but it is not an agent of historical authority.
Sampling and manual description are labor-intensive, but they provide a trail of accountability. Archivists can use generative AI to manage record accumulation, but they must embed it in a framework of oversight. It should be used to point researchers toward the records, not to replace the act of reading them.
Generative AI is a powerful mirror, but it can distort, magnify, and invent what it reflects. By centering the SAA Code of Ethics and prioritizing the principles of provenance and transparency, archivists can ensure that AI illuminates collections. The “future of the past” depends on the ability to govern the machines that tell stories, ensuring they help support decisions rather than author memory.
0 Comments