Overview
A working demonstration of the principles my research argues for: a documented
workflow in which generative AI assists with production tasks — drafting,
organising, coding support, formatting — while a human remains identifiably
responsible for meaning, accuracy, ethics, safety and final approval of every
output. This website and my research materials are produced under this
workflow.
Problem
“Responsible AI use” is easy to endorse and hard to specify. Most guidance
stops at principles. The practical question is where exactly human judgement
must sit in a real production process, and what records make responsible use
demonstrable rather than merely claimed.
My role
Designed the workflow, applies it in daily research practice, and maintains
the documentation and checklists.
Research question
What does a defensible division of labour between human and AI look like in
research and content production — and how can it be documented so that
responsibility is traceable?
Method
Iterative design in practice: each stage of my research production (literature
work, drafting, coding assistance, formatting) was mapped, and for each stage
the workflow specifies what AI may do, what only a human may do, and what
record is kept.
The workflow
| Stage | AI may assist with | Human is responsible for |
|---|
| Planning | Structuring options, outlines | Goals, scope, framing |
| Research | Locating and summarising sources | Reading sources, judging relevance and quality |
| Drafting | First-pass prose, reformulation | Claims, argument, interpretation |
| Analysis / code | Boilerplate, refactoring suggestions | Analytic decisions, correctness |
| Verification | Flagging inconsistencies | Checking every factual claim against sources |
| Approval | — | Final sign-off on every published output |
Two rules are absolute: no AI-generated claim is published without human
verification, and the human approver is accountable for the final artefact
as if they had produced every word.
General-purpose LLM assistants, verification checklists, disclosure notes,
version history as the audit trail.
Process
Each output carries a short internal record of what AI assisted with and who
verified what. The overhead is deliberately small — a workflow that is too
heavy to follow produces false documentation, which is worse than none.
Findings
The clearest recurring observation from applying this workflow daily is that
the verification stage is where most corrections happen. AI-assisted
first-pass drafting is rarely wrong in an obviously detectable way — it is
usually wrong in ways that only surface when a human checks sourcing, nuance
or argument against the original material. This is an observation from my own
practice, not a formally evaluated finding.
Outputs
- The documented workflow (this page)
- Verification checklist — see the Technical Writing section for the derived guides
Impact
Its direct impact is this website and my research materials, all produced
under this workflow.
Limitations
- Self-imposed workflows depend on discipline; there is no external enforcement.
- The workflow is designed for individual research practice and would need
adaptation for team settings.
- It documents process, which constrains but cannot guarantee output quality.
Lessons learned
- The verification stage is where responsibility becomes real; every stage
before it is preparation.
- Recording AI assistance at the moment of use is cheap; reconstructing it
later is impossible.
Not written up as a standalone publication; it operates as documented
practice rather than a research output.
Downloadable materials
The verification checklist referenced above is being prepared for standalone
publication and is not yet posted here.