Caught Between Gates
Digging Deeper into Explainable AI and the Archival Mind
I’m still in motion — or rather, not.
Somewhere between weather delays and the latest air-traffic reshuffle, my flight’s been pushed back again. So, like any archivist stranded with a charging port and a restless mind, I’ve reopened the laptop.
This post picks up where the last one left off — that first sketch on how Explainable AI (XAI) might look when viewed through the lenses of records management, archival theory, and information science. The last essay was written in a quieter moment, when I began connecting archival transparency with algorithmic explainability. Now, grounded (temporarily) and sipping a hazy IPA, I’ve been digging deeper into the research itself — looking for the scaffolding behind those ideas.
Revisiting the recordkeeping view
The starting point, still, is Jenny Bunn’s “Working in Contexts for Which Transparency Is Important” (2020) — Bunn reframes XAI not as a technical add-on but as a recordkeeping condition: explainability becomes part of what makes a record accountable. She asks what new record forms emerge when the act of decision-making itself is mediated by machine learning. That’s the right provocation.
Next comes Giovanni Colavizza’s “Archives and AI: An Overview of Current Debates and Future Perspectives” (2021) — a panoramic survey of how automation is bending traditional archival principles: provenance, original order, appraisal. His observation that algorithmic mediation subtly shifts what archivists consider context has been echoing in my mind.
And there’s Patricia Franks, whose paper “Positioning Paradata as AI Processual Documentation” (2022) introduces paradata — the metadata of metadata creation — as a form of accountability artefact. In the XAI world, that translates neatly: process traces, decision logs, model explanations. They all belong in the record.
L. Zhang’s 2024 conference paper on explainability and enterprise information management extends this to the governance layer. It’s one thing to document an AI decision; it’s another to integrate that documentation into compliance and lifecycle frameworks.
Together, these pieces sketch a field quietly redefining itself: archives not as passive repositories of evidence, but as systems that must themselves explain.
The information-theoretic undercurrent
This time around, I also wanted to see what the information scientists have been saying. And that led me into the Information Bottleneck (IB) literature — the elegant, mathematical cousin of archival provenance.
The 2024 survey by Hu et al. “A Survey on Information Bottleneck” lays out the logic: every explanation sits somewhere on a curve between compression and fidelity. The more concise the explanation, the less information it carries — and vice versa. That’s not far from our own archival tensions between appraisal and preservation.
Then, in 2025, the arXiv review of Self-Interpretable Neural Models took the idea further. These are systems designed to generate their own explanations by construction, optimizing the same balance we struggle with in documentation — enough information to be intelligible, but not so much that it collapses under its own weight.
And Sanneman’s MIT study on “Workload Trade-off in Human-Centered Explainable AI” (2024) reframes the entire problem as a human workload trade-off. The more detailed the explanation, the more cognitive load it imposes. That feels painfully familiar to anyone who’s ever tried to write or read an overly dense metadata standard.
So, while the archival world debates which AI logs to keep, the information theorists are quietly quantifying how much explanation is enough. It’s an unexpected dialogue: entropy meets accountability.
Why it matters — still
If the first post asked whether archives and XAI belong in the same conversation, this one answers why.
For records and information governance, the connection is direct. Provenance, accountability, transparency — these aren’t optional virtues; they’re the preconditions for trust. XAI is simply the next frontier where those conditions must be asserted.
And for information theory, the relevance flows back the other way. Concepts like mutual information and compression offer a vocabulary — even a calculus — for understanding what it means to preserve the essence of an explanation without drowning in data.
I keep coming back to two questions that, to me, define the intersection of these worlds:
What must we record when an AI system makes a decision?
And what explanation fidelity is necessary — or even possible — given the limits of information theory?
Between those questions lies the emerging practice of AI recordkeeping — one that bridges the qualitative ethics of archives with the quantitative logic of information science.
As my gate number flickers to yet another update, I’m convinced this is the conversation we need most: how to make the machines explain themselves — and how to keep those explanations as records.