Deep Insight // Appendix worksheetAll worksheets  /  Appendix C

Appendix C | Companion to Chapters 1 and 8

SOP: AI Bias Mitigation

A standard operating procedure for catching bias before it reaches a report.


Autosaves on this device PDF version

Purpose. A repeatable procedure for identifying, documenting, and mitigating bias when AI tools support intelligence analysis.

Scope. Any AI-assisted task where model output informs a judgment, flags a person or entity, or ends up in a report.

Part 1

Pre-use screening

Complete before the tool touches a live task.

Part 2

During use

Follow at the point of analysis.

Part 3

Post-output validation

Complete before any AI-assisted finding moves forward.

Part 4

Where did the bias get in?

Bias can enter at any stage of the AI lifecycle. When an output looks wrong, use this to trace it, then check the stage(s) you suspect.

StageRisk
Data collectionData that doesn't represent the real world produces a model that inherits that skewed view.
Model designHow inputs and features are structured can hard-code assumptions about what is normal.
Optimization goalsA system built for speed or accuracy without fairness can learn shortcuts that harm certain groups.
DeploymentReal-world conditions differ from training conditions. Without monitoring, errors compound.
Part 5

Escalation

If a bias concern could affect a live case, decision, or person:

  1. Halt forward distribution of the affected output.
  2. Document the concern and the lifecycle stage it likely came from (Part 4).
  3. Escalate to a supervisor or designated review authority before the output is used in any operational decision.
  4. Record the resolution and any corrective action.
Escalation log
DateConcernStageEscalated toResolution
Part 6

Quick bias mitigation checklist

The short version from Chapter 1, for when you're moving fast. Initial each step.

#TaskComplete?Initials
1Review tool limitations and disclosures
2Use well-structured and neutral prompts
3Ensure input sources are diverse and balanced
4Validate AI outputs against alternative sources and tools
5Conduct a human review and critical analysis of outputs
6Document tool use and approvals
7Ensure proper classification and handling of outputs

We can't eliminate bias altogether. We can shrink it a lot with design, oversight, and a documented procedure like this one. Human oversight is non-negotiable.

Deep Insight: AI for the Intelligence Analyst by Rae Bakerdeepinsightbook.com