Editing Openai/6944034d-9850-8003-8632-66d16158b32f (section)

==== Key Takeaways — Organized by Topic ====

===== 1. Core Principle: Decompose Before You Prompt =====

Summary:
Effective AI-assisted genealogy requires breaking complex research questions into smaller, manageable components before involving AI.

Fact Extractions:
* The article adapts the carpenter’s saying “measure twice, cut once” into a genealogical corollary: “decompose first, prompt second.”
* Large genealogical questions given to AI are often projects, not tasks.
* AI performs poorly when asked to resolve entire research problems in a single step.

===== 2. Limits of AI in Genealogical Judgment =====

Summary:
AI lacks the human genealogist’s judgment, interpretation skills, and domain expertise.

Fact Extractions:
* AI can “stumble” when asked to make genealogical judgments.
* Errors occur not because of lack of capability, but because of misaligned task assignment.
* AI may produce authoritative-sounding answers that are unreliable when asked to perform interpretive reasoning.

===== 3. Steve Little’s Three Best Practices (Referenced) =====

Summary:
The framework builds on Steve Little’s long-taught principles.

Fact Extractions:
* The three practices are explicitly named: 1. Know Your Data 2. Know Your Model 3. Know Your Limits
* This article focuses primarily on the second and third practices.

===== 4. The Three Task Types Framework =====

Summary:
Not all genealogical tasks are suitable for direct AI assignment. Tasks fall into three distinct types.

====== Type 1: Narrow Tasks ======

Summary:
Constrained, factual, and easily verifiable tasks.

Fact Extractions:
* Have one correct answer.
* Require minimal context.
* Errors are easy to detect.
* Example questions include: - Definitions of terms - Record availability by place and time - Typical informants for records

====== Type 2: Contextual Tasks ======

Summary:
Specific analytical tasks that require user-supplied documents and direction.

Fact Extractions:
* Require context such as documents, records, or datasets.
* The genealogist provides the interpretive framework.
* AI performs systematic analysis.
* Examples include: - Household analysis - Classification of information as primary or secondary - Comparison tables of census data

====== Type 3: Open-Ended Tasks ======

Summary:
Complex research questions that should never be assigned directly to AI.

Fact Extractions:
* These are multi-step, interpretive research questions.
* Examples include: - Identifying parents - Resolving conflicting records - Determining identity across records
* AI cannot reliably perform genealogical judgment required for these tasks.
* Directly assigning these tasks to AI poses a risk of misleading results.

===== 5. The Decision Heuristic =====

Summary:
A practical test for determining task suitability.

Fact Extractions:
* The test question is: “Would you ask a new colleague to do this on their first day?”
* If the answer is no, the task is a Type 3 task.
* Type 3 tasks should be decomposed into Type 1 and Type 2 subtasks.

===== 6. Decomposition in Practice =====

Summary:
Breaking a large research question into actionable subtasks makes AI assistance effective.

Fact Extractions:
* Example transformation of a Type 3 question into: - One Type 1 task (record types) - Multiple Type 2 tasks (household analysis, candidate evaluation)
* Decomposition results in: - Verifiable questions - Actionable next steps - Retained human control over reasoning

===== 7. Why Decomposition Matters =====

Summary:
Decomposition preserves genealogical ownership and research integrity.

Fact Extractions:
* Decomposition ensures the genealogist: - Chooses sources - Evaluates evidence - Resolves conflicts
* AI handles: - Systematic - Checkable - Tedious work
* Researchers who succeed with AI understand that AI assists but does not replace judgment.

===== 8. Ethical and Intellectual Positioning of AI =====

Summary:
AI is framed as a tool, not a decision-maker.

Fact Extractions:
* AI is described as a “very capable assistant.”
* The genealogist remains responsible for conclusions.
* The article closes with the reaffirmation: - “Know your model. Know your limits.” - Research conclusions should remain the genealogist’s own.

===== 9. Attribution and Licensing =====

Summary:
The framework has explicit origins and licensing.

Fact Extractions:
* The framework is adapted from: - “Stop Asking AI to Build the Whole Feature” by BekahHW.
* Genealogical adaptation by: - Steve Little - AI-Jane
* The accompanying infographic is licensed under Creative Commons 4 BY-NC.