Provenance first
Record what model or tool assisted, which source material it saw, what transformations occurred, and which human accepted or rejected the result.
Appendix H
AI-Assisted Research Guide
A practical guide to using AI in research while preserving provenance, validation, privacy, independence of evidence, uncertainty, and resistance to automation bias.
Resource map
Inside this resource
Use the sections below to work with this companion resource alongside the book.
- 01 Provenance First Available
- 02 Validation Proportional to Consequence Available
- 03 Search Is Not Evidence Available
- 04 Extraction and Meaning Available
- 05 Correlated Error Available
- 06 Automation Bias Available
- 07 Privacy Available
- 08 Visible Uncertainty Available
- 09 Date-Sensitive Tool Guidance Available
- 10 Human Authorization Available
AI-assisted research guide
Keep durable research controls visible as tools change.
These principles focus on provenance, validation, privacy, uncertainty, correlated failure, automation bias, and human authorization rather than on any particular model or vendor.
Validation proportional to consequence
A low-stakes draft summary and a rights-sensitive causal claim should not receive the same validation burden.
Search is not evidence
AI can widen discovery, but source quality, original context, independence, and claim fit must be checked at the source level.
Extraction can silently alter meaning
Tables, quotations, classifications, translations, and structured fields need spot checks or systematic validation against originals.
Generated synthesis can correlate errors
Multiple models can repeat the same source, benchmark, training-data misconception, or search result; model count is not evidence independence.
Automation bias is institutional
Design workflows so reviewers can disagree, inspect inputs, see uncertainty, and meaningfully depart from recommendations.
Privacy survives convenience
Do not expose sensitive personal, administrative, legal, health, or confidential material to tools whose data handling is inappropriate for the task.
Uncertainty should become more visible
Use AI to enumerate alternative explanations, missing data, contradictory evidence, and failure modes—not to cosmetically remove uncertainty.
Model-specific advice expires
Date capability claims, workflow recommendations, limitations, and evaluation evidence so readers can tell when guidance may have aged.
Human authorization remains a political boundary
Better research assistance does not create jurisdiction, democratic standing, legal authority, or the right to make a binding public choice.
Current-practice note
Model-specific recommendations can age quickly. When a course or research workflow depends on a particular capability, record the tool, date checked, validation evidence, and any material limitation.
Find related companion resources.
Browse tools, methods, references, and teaching material by what you are trying to do rather than by appendix letter alone.