UZH · Fall 2026 8 modules self-paced
AI as aResearch Tool
workflow design for the social sciences
§ overview
This is a self-paced course on using AI — specifically large language models — in social science research. It was developed from a three-day block seminar taught at the University of Zurich in September 2026 and expanded into eight standalone modules.
The course does not teach you to use specific AI products. Products change; the methodological questions do not. Instead, it teaches a framework for deciding where, if anywhere, AI should enter your research workflow, under what conditions, and on the basis of what evidence.
who this is for
- — MA and PhD students in the social sciences
- — Researchers who want to use AI thoughtfully rather than optimistically
- — No programming experience assumed
- — No commercial accounts, API keys, installation, or payment required
how to use this course
Each module is self-contained and can be taken in any order. The recommended path:
1 → 2 → 3 → 4 → 7 (validation) → 5 → 6 → 8 (synthesis)
You can also take Module 1 (foundations) and Module 7 (validation) first, then pick the application modules that matter to your research. Module 8 is the capstone.
Estimated total time: 24–30 hours (8 core modules × 3–4 hours each). Module 09 is a reference directory — use it any time.
§ modules
-
mod 01
How LLMs Work
A practical model of transformer-based language models for social scientists. Tokens, context, prediction, and why fluency is not truth. Introduces the four-question course lens and the capability map.
-
mod 02
Writing with AI
Drafting, editing, and style transfer with AI. Criteria-based evaluation: define "good" before generating. Derive a style guide from examples. Write a disclosure for AI-assisted writing.
-
Design a bounded literature search with explicit scope and stopping rules. Screen results with pre-registered criteria. Build an evidence table. Draft a source-bounded synthesis. Verifiy every cited source exists.
-
Turn a codebook into an AI prompt. Run a miniature annotation study. Measure reliability and report it honestly. Understand why high agreement ≠ construct validity.
-
mod 05
Code & Data Analysis
Use AI to explain, debug, and document code. Explore data with AI assistance. Indepedently verify statistics and interpretations. Understand common failure modes: halucinated functions, off-by-one errors.
-
mod 06
Agents & Workflow Automation
Distinguish chat, file-working assistants, and agents by autonomy level. Design a bounded agent workflow. Write persistent instructions. Stress-test an instruction artifact.
-
The evidence spine of the course. Seven validation checks. Reproducibility with LLMs. Data and institutional boundaries. The disclosure standard. Research integrity: fabricated references, confidentiality, accountability.
-
mod 08
Frontiers & Integration
Local and open-weight models. Multimodal AI. Institutional and ethical boundaries. Integrate two or more modules into a coherent research workflow. Build the portfolio that answers: where, if anywhere, should AI enter my research?
-
mod 09
Research Tools Directory
A curated directory of AI tools for academic research. Coding agents and terminal harnesses (Claude Code, Codex, OpenCode, Aider). Agentic systems (Hermes, OpenHands, CrewAI). Model access (HuggingFace, OpenRouter, Ollama). Literature and writing tools (OpenAlex, Zotero, Elicit). Research infrastructure.
§ the course lens
Every module uses four questions to evaluate AI use in research. Apply them before using AI, after using AI, and when reading research that uses AI.
question 1
What is being delegated?
Name the task and the decision separately. A task delegation ("summarise these three papers") is diferent from a decision delegation ("identify the most important papers in this field").
question 2
What would make the result believable?
Name evidence outside the model output. Did you check the sources? Did a human agree with the labels? Did you run it twice? The model's confidence is not evidence.
question 3
What could be lost or flattened?
Look for variation, minority cases, ambiguity, provenance, and tacit knowledge. Are edge cases being forced into the closest category? Is the source of a claim becoming invisible?
question 4
What must remain visible or human?
Which decision requires a researcher's judgment? What check must be performed by a person? What record must survive so a reviewer can inspect what happened?
§ shared resources
These reference documents are used across multiple modules:
- glossary— key terms with module references
- validation menu— seven checks for AI-assisted claims
- disclosure template— journal-ready AI use disclosure
§ portfolio
The course culminates in a portfolio that answers one question:
You may conclude: integrate, integrate with limits, continue experimenting, or do not integrate. All four conclusions can be equally strong. The grade depends on the quality of evidence and reasoning, not on whether you adopted AI.
constraints
- — No programming required. Code is demonstrated and provided as scafold.
- — No commercial accounts, API keys, installation, or payment required for any asessed component.
- — Every exercise has an equivalent no-account path. Constructed fallback outputs are provided.
- — Do not upload material you lack the authority to share. Unpublished co-authored drafts, interview transripts, administrative data, and identifiable records stay on your own device.
§ instructor
Dr. Colin Henry, Postdoctoral Researcher
Department of Political Science, University of Zurich
colin.henry@ipz.uzh.ch →
Source repository: github.com/cmhenry/uzh_ai_for_research_f2026 →