Turn Documents into Data with AI - A Short Course

Each session combines lecture and live demonstration: building codebooks, running extractions, diagnosing failures, and refining outputs. At times, you will watch the work unfold on the instructor’s screen; at others, you will make the same moves on your own, with supervision and feedback.

Along the way you will learn the transferable craft: how to spot where the AI is guessing rather than reading, how to refine a field to remove codebook ambiguity, and how to know when the answers are trustworthy enough to publish.

The course is built around four commitments that, together, provide an audit trail for trustworthy use of AI:

• • The codebook is the durable artifact. It is the document that holds your operationalization—every choice you made about what counts and what doesn’t, written down so that reviewers (and your future self) can see them. A good codebook outlasts the model that ran it.
  • AI is a reader, not a judge. Decomposing a research question into reading tasks an AI can do reliably is the core engineering move. In this course, you’ll learn decomposition as a craft, with concrete refinement techniques and a vocabulary for diagnosing failures.
  • Disagreement is the diagnostic. When multiple AI readers run the same codebook and disagree, those are the cells worth auditing first. Multi-reader agreement is treated as publication-defensible evidence of reliability, but not a stamp of approval.
  • Data are shared with codebook and reasoning. Your spreadsheet gets a URL that you can share. You can click any cell to see the codebook item, key quotes from your document, and the AI’s reasoning for the value in the cell.

The codebook can be reused on different documents by you or others and revised or refined as needed. The disagreements reveal where the AI readers interpret the codebook differently, or places where they make errors.

We will use Data Mint, a web-based platform built to do exactly what this course teaches. There is nothing to install, no command line, and no programming. All extraction work happens in a browser, and course accounts are provided so that no API keys or paid subscriptions are required.

To follow along with the hands-on exercises, you will need:

• • A laptop or desktop computer (not a mobile phone), running Mac, Windows, or Linux, with a modern web browser and a reliable internet connection.
  • A Data Mint account. Course accounts will be set up during the pre-course intake. Data Mint is open to anyone, so you can visit datamint.ing, request access with your academic email, and try it out before deciding whether to register.
  • A spreadsheet program (Excel, Google Sheets, or similar) for inspecting exported data.

Recommended (but not required): you should arrive with a research question and a corpus of documents.

You will leave with three things:

• A publishable codebook for AI-assisted extraction, which you can apply to the rest of your corpus.
• A spreadsheet with the data you came for.
• Reliability metrics—based on multi-reader agreement—at publication-defensible levels.

No experience with Python, R, or any programming language is required (the course will not instruct on how to analyze your data, only how to generate the data).

If you’re a researcher who needs to turn documents into data, this course is for you. That includes—but is not limited to—anyone doing data extraction, document coding, content analysis, chart abstraction, structured literature review, or large-scale qualitative coding.

It fits especially well for researchers working on systematic reviews (screening, extraction, and risk-of-bias assessment), but the craft applies to any project where a body of text must become structured, analyzable data.

Day 1

First extraction

• • Intro: How might we come to trust AI for research? What are key concerns? The idea of Minting Data.
  • Build a draft codebook for your own research question and produce your first AI-extracted outputs
  • Title-and-abstract screening as a low-stakes warm-up
  • Introduction to the vocabulary of scaffolding—the structural moves that make an instruction reliably followable
  • Hands-on exercise: a v0 codebook, a first round of extracted data, and at least one diagnosed failure

Day 2

The scaffolding deep dive

• • The signature work: learning to refine your codebook
  • Three legitimate fixes for a failing field:
    • Reword the question—clarify decision boundaries, add worked examples
    • Add scaffolding—upstream fields that break a multi-part judgment into steps the AI can handle
    • Rebuild the field entirely
  • Hands-on exercise: a cross-domain exercise that forces the meta-skill—diagnosing structural cracks—before you apply it to your own codebook

Day 3

Transfer and ship

• • Apply the craft to risk-of-bias assessment—a new domain that uses the same skill
  • Peer audits of each other’s codebooks
  • Leave with: a shipped codebook, an exported dataset, and reliability metrics from multi-reader agreement