Auditable AI-Assisted Research Writing: An Engineering Discipline with Pre-Registered Process Observation

The integration of AI into research production, particularly through language models, necessitates a robust framework for accountability and auditability. A proposed engineering discipline focuses on creating an auditable history of AI involvement at the point of production, rather than through retrospective detection. This discipline includes mechanisms like git sealing, hash-bound provenance, refusal logs for non-compliant artifacts, role separation among AI models, and programmatic assembly from registered sources. Adherence to these measures is assessed via 'metric cards' which pre-register potential blind spots and evidential standing.

Why it matters

As AI tools become more ubiquitous in knowledge creation, establishing clear accountability and auditability mechanisms is critical for maintaining integrity and trust in produced artifacts. This framework provides a methodology to ensure transparency regarding AI's contribution, which is essential for validating the credibility of research and other outputs in an increasingly AI-driven landscape.

Key insights

• Current AI-assisted research production lacks accountable historical records of machine involvement.
• A new engineering discipline proposes a proactive approach to auditability during the production phase.
• Key elements of this discipline include git sealing with anchor lineage, hash-bound provenance, and red-line gates.
• Non-compliant artifacts are refused, and these refusals are logged.
• Role separation for different AI models and programmatic assembly from registered sources are integral.
• Metric cards, pre-registered with blind spots and evidential standing, are used to instrument adherence.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.10858