February 5, 2026
This talk:
What is AI?
Here: Focus on LLMs, preferably open models such as Mistral 24B via chat.kiconnect.nrw, etc. that generate code based on prompts, used via a terminal, IDE, software integration, or browser. Most models are closed source (not transparent), not EU-based, dependent on company support, etc, see also KI-Whitelist der Universität zu Köln.
Example use-cases:
Many guidelines, laws, regulations:
AI Act (as based on FOI compliance checker):
“universities need to take their role seriously to safeguard higher education, critical thinking, expertise, academic freedom, and scientific integrity” (Guest et al., 2025)
Critical AI literacy criticism includes (e.g. Guest et al., 2025):
“While automation may give the false impression of rigor and efficiency, it leads to conceptual and scientific deskilling, deteriorates reflexive theorizing…” Van Rooij & Guest (2025)
Critical review as a way to reduce risks, and address KI-Richtlinie §2.4 (human control) and §9.1 (output verification)?
Idea: start from classic good practice rules of software engineering
| Practice | Idea |
|---|---|
| KISS | Keep it simple |
| DRY | Don’t repeat yourself |
| YAGNI | You ain’t gonna need it |
| SoC | Separation of concerns |
Phase 1 → 2: From Script to Modular Design
Phase 2 → 3: Adding more Features
Phase 3 → Adding Quality Checks
Problems with classic good practices:
benchmark.py, unused dependencies (spacy, scikit-learn, …)Some additional critical developments:
“Fast” Iteration
Generate (AI) → Review (Human) →
Refine → Integrate (one-way)
Advantages:
Problems:
Clausemate Example:
Three working parsers, but with much overlap. Never asked “why 3?”
“Slow” Questioning**
AI Pro-active 'Review'
─────────── ─────────────────────
Suggests x ←→ Question necessity - YAGNI
Verbose code ←→ Demand simplicity - KISS
Duplication ←→ Enforce reuse - DRY
Mixed concerns ←→ Separate clearly - SoC
Advantages:
Problem:
Clausemate Example:
Better understand text structure first before developing three different parsers.”
Fast vs slow
Some principles for slow review
Warning signs
Conclusions
Take-Home Message:
AI builds faster than we can understand easily. Train your critical thinking skills and safeguard research integrity.
“Für jede KI-gestützte Anwendung ist eine namentlich benannte Rolle mit menschlicher Letztverantwortung festzulegen.” (UzK KI Richtlinie §2.4b)
Some References: