Faster research needs more provenance, not less

#Question

Why does AI-assisted acceleration make traceability more important rather than less?

#Answer in brief

Because speed increases the number of transformations, drafts and decisions that can occur before anyone reconstructs what happened.

A manual analysis might involve ten consequential choices over a week. An AI-assisted workflow can generate multiple data-cleaning strategies, models, sensitivity analyses, source lists and interpretations in minutes.

That is useful.

It also creates a provenance problem.

When work accelerates, memory becomes a weaker audit system.

#The old bottleneck was often execution

Researchers historically spent substantial time finding syntax, formatting tables, writing boilerplate, manually searching references, translating between software and drafting routine text.

AI can reduce those costs.

The new bottleneck increasingly becomes selection: which result should be trusted, which source is real and applicable, which code version generated the table, which model was accepted, which interpretation matches the estimand, and which AI suggestion was rejected and why.

Acceleration shifts effort from producing candidates to governing candidates.

#Speed multiplies branches

Consider a deliberately synthetic workflow example. An analyst asks an AI system to explore an association.

Within minutes it generates three cohort definitions, four regression specifications, two missing-data strategies, five plots and six possible explanations.

That is many plausible analytical paths before human review.

If the final report shows only one path, the reviewer needs to know which alternatives existed, which were actually executed, why one was selected, and whether selection depended on the observed result.

Without that history, flexibility can become hidden researcher degrees of freedom at machine speed.

#Provenance converts speed into reviewable work

W3C PROV models provenance around entities, activities and agents because understanding origin and transformation helps determine trust, compliance and reproducibility.^[1]

In an AI-assisted workflow, provenance can connect source data, extraction, AI-generated suggestion, human review, transformation, model, diagnostic, revision, result and interpretation.

The graph does not guarantee correctness.

It tells reviewers where to look.

#AI introduces a second provenance layer

Traditional analysis provenance asks which code and data produced a result.

AI-assisted work adds another question: which context, model or tool state produced a suggestion, and which parts were accepted by a human?

The answer does not always require storing raw conversations indefinitely.

A proportionate record can preserve task, input artifact identities, model or tool identity where relevant, generated candidate, human disposition and downstream artifact.

The point is to retain decision-relevant lineage.

#Citation work is a good example

AI can propose references rapidly.

Without provenance, a draft may contain a convincing citation whose origin no one remembers.

A controlled workflow records: AI proposed source, source retrieved, bibliographic identity verified, claim checked against source, citation accepted.

If retrieval fails, the claim remains unsupported.

The acceleration occurs in discovery, not in lowering the verification threshold.

#Code generation is another example

An AI assistant generates a data-cleaning function.

The function runs.

Later, a reviewer notices that it silently converts impossible dates to missing values.

If provenance is available, the team can identify which datasets used the function, which outputs depend on it, when the function changed, and whether affected analyses need regeneration.

Without lineage, the safest response may be to rerun everything.

Traceability reduces the blast radius of errors.

#Faster interpretation can be more dangerous than faster code

A model output can be produced deterministically.

Interpretation often involves compression and judgement.

AI can turn a dense result into persuasive prose instantly.

If the summary overstates causality, every downstream artifact can inherit the distortion.

Provenance should therefore extend beyond computation to claim formation.

#Accountability needs role separation

OECD's AI Principles call for accountability and traceability of datasets, processes and decisions across the AI system lifecycle.^[2]

Research workflows can adopt the same general discipline: AI system proposes, software executes, reviewer verifies, author interprets, accountable human releases.

These roles can overlap in one person, but they should remain conceptually distinct.

An AI tool is not the accountable author because it generated useful material.

#Speed can increase review quality if review capacity is preserved

Acceleration is not inherently anti-rigorous.

It can free time for sensitivity analyses, source verification, alternative models, clearer documentation and adversarial review.

But only if saved production time is reinvested in review.

If the same speed is used only to produce more outputs, the review-to-output ratio can collapse.

#More useful metrics than tokens per hour

For evidence work, meaningful questions include whether material claims are linked to verified sources, whether a reviewer can reconstruct why an analytical decision changed, whether affected outputs can be identified after an error, whether human approvals are attached to consequential decisions, and whether unsupported claims are held rather than polished.

Those properties measure controlled acceleration.

#Why this matters more at scale

Consider a hypothetical scaling example: one manually written report with ten material claims versus an AI-assisted process producing one hundred reports with ten claims each. The numbers are illustrative, not an empirical error-rate model. The point is simply that the number of opportunities for an undetected defect grows as output volume grows.

The solution is not necessarily slower generation.

It is stronger systematic verification and provenance.

Automation should scale the controls that make speed safe.

#Relation to human decision rights

The companion guide AI can assist the workflow; it should not own the epidemiological decision defines which choices need explicit human ownership.

This article addresses the infrastructure around those decisions: how to preserve the path so the ownership is meaningful.

#Connection to Weavidence

This principle is directly relevant to the design direction of Weavidence Lab: computation, interpretation and review are more useful when their dependencies and changes remain inspectable. That connection is a design rationale, not evidence that the product has achieved scientific effectiveness.

#Limitations and uncertainty

This is a methodological commentary, not an empirical estimate of how much provenance any AI-assisted workflow requires. WHO's 2025 LMM guidance addresses broad health, research and public-health governance rather than prescribing one provenance schema.^[3] OECD and W3C sources provide governance and provenance principles that require application-specific implementation.^[2][1] Excessive logging can create privacy, security and cognitive-burden problems; traceability should be proportionate to consequence.