cold-chain · deliverable

AI Production Disclosure & Responsible-Use Guidelines

Two parts: (A) transparent disclosure of AI's role building this kit — demanded implicitly by Exact's application question and Ricoh's tool-adoption duty; (B) a reusable responsible-AI-use guideline one-pager modeled on what UMC Utrecht asks candidates to produce ("kaders, richtlijnen en good practices").

A. How AI was used to build this kit (honest ledger)

Step	AI contribution	Human control point
Needs-analysis memo drafting	First-draft language; incident-pattern framing	All thresholds, numbers, failure patterns specified/human-edited; memo structure follows I Wayan Usada's analysis framework from 26 job analyses
Storyboard scenes & narration lines	Draft copies of scene beats	Every scene beat hand-tuned; EN/NL subtitle pairs human-reviewed
Quiz item bank (8 items)	Option generation for distractors	Correct-answer keys + rationale written by hand; every distractor checked against SOP logic (a domain-expert reviewer would sign off in production)
HTML/CSS/JS implementation	Code scaffolding and iteration speed	Design tokens chosen manually; interaction logic reviewed line-by-line; syntax-verified with Node; every interactive flow logically tested
Visual style system	Palette/constraint proposals within briefs	Final dark-navy palette hand-selected; contrast-checked against AA
Deliberately NOT AI:	—	Facts & thresholds (2–8 °C band, 30-min dock rule); assessment weighting; evaluation metrics; all final copy decisions

Why disclose: employers receive hundreds of portfolios claiming AI fluency. Showing where AI stopped and judgment started is the differentiator — it mirrors how professional teams actually govern AI-assisted production.

B. One-pager: Responsible AI Use in Learning Content (VanDijk-style policy)

Scope: any employee using generative-AI tools to draft, illustrate, translate, or assess learning material.

1. Purpose principle

AI accelerates production; humans own accuracy. No learner-facing fact may exist solely because a model produced it.

2. Verification ladder

Safety/compliance facts (thresholds, legal limits): require primary-source citation before use — never model memory.
Procedural steps: validated against current SOP version; when SOP changes, AI-assisted modules get a mandatory review ticket.
Illustrative scenarios: may be AI-drafted, but consequences shown must be plausible per operations lead.

3. Data protection (AVG/GDPR)

Never paste learner data, incident records with identifiers, or unreleased business figures into external tools.
Prefer enterprise/self-hosted models for anything touching operational data.

4. IP & provenance

AI-generated images/media used in courses carry a source note in course meta.
Generated code respects license hygiene (no unreviewed copyleft contamination in proprietary stacks).

5. Bias checks

Distractor options and example characters reviewed for demographic stereotyping — our receiver scenario deliberately features varied names/settings across language versions.

6. Transparency default

Learner-visible courses state at intro: "Drafted with AI assistance, verified by [role]. Feedback welcome." Trust survives honesty better than discovery.

7. Escalation

Uncertain case? Route to Learning Ops review rather than shipping quietly. When in doubt, leave it out.

Prefer the source? Download the raw markdown.