Better Answers, Better Institutions

How to study

Work through the lessons in order, or use the correction-plan capstone to identify the sections you need. Allow about six hours for the teaching, exercises and initial capstone. A workplace application will take longer because it requires evidence from the actual process.

Use a fictional organisation or public example. Do not put confidential, personal or client information into working notes. The self-checks run in your browser; the course does not require an account or transmit your answers. Keep your written work in your own files.

Each lesson identifies related chapters of Better Answers, Worse Decisions. The course contains new explanations, hypothetical cases and exercises, not the book or chapter downloads. The book is intended for separate purchase; a verified purchase link will be added when publication details are available.

Attempt each problem before opening its worked feedback. Agreement with my position is not a marking criterion. A strong answer may reject a proposed control if it gives evidence, identifies the resulting exposure and defends a better alternative.

What you will be able to do

• Compare AI-assisted work with the realistic alternative, using both task and institutional outcomes.
• Trace the evidential dependence behind apparently separate recommendations.
• Specify the competence, authority and resources needed to detect, challenge, reverse and remedy an error.
• Evaluate a claim without treating illustrative calculations, selected cases or coder agreement as causal proof.
• Produce a costed, testable correction plan, including reasons to reject unnecessary safeguards.

Eight decisions to work through

1. A better answer is not the same outcome

   40 minutes · Companion reading: Introduction; Chapters 1, 4 and 9

   • Separate task-level gains from effects on a wider decision process.
   • Choose a realistic comparator rather than an idealised human or machine.

2. When three confirmations are one path

   45 minutes · Companion reading: Chapters 2, 3 and 10

   • Distinguish different outputs from independent evidence.
   • Explain why a shared failure can defeat apparent corroboration.

3. Decide what to delegate

   45 minutes · Companion reading: Chapter 4

   • Specify the operating boundary of an autonomous task.
   • Distinguish a meaningful intervention from ceremonial approval.

4. Keep the ability to recover

   45 minutes · Companion reading: Chapters 5, 6, 7 and 11

   • Identify which capabilities must survive routine automation.
   • Design an observable recovery exercise rather than relying on confidence.

5. Give a challenge somewhere to go

   45 minutes · Companion reading: Chapters 8 and 12

   • Distinguish permission to object from authority to change a decision.
   • Map who controls each consequential part of an automated workflow.

6. Build the connected correction path

   45 minutes · Companion reading: Chapter 13

   • Connect detection, challenge, diagnosis, authority, remedy and learning.
   • Test a handoff rather than counting safeguards.

7. Who pays for the capacity to correct?

   45 minutes · Companion reading: Chapter 14

   • Identify a gap between private adoption incentives and wider correction costs.
   • Cost a proportionate safeguard without presenting hypothetical values as evidence.

8. Make the argument survive its strongest objection

   50 minutes · Companion reading: Note on evidence; Chapter 14; Conclusion

   • Distinguish illustrative cases, coding agreement and causal identification.
   • Revise a claim when an alternative reading changes its apparent support.

Apply and defend

Design and red-team a correction plan

Choose a fictional institution deploying an automated decision system. Produce a 1,500–2,000-word decision paper and a one-page correction-path diagram. Use the book’s questions as a framework, not as conclusions you must agree with.

What to produce

1. Define the task, affected population, operating boundary and realistic alternative.
2. Give task-level and institutional outcomes, with observation periods and denominators.
3. Map shared evidential dependencies and retained diagnostic competence.
4. Trace one material error through detection, challenge, diagnosis, authority, remedy and learning, including an interrupted handoff.
5. Cost one proposed control, show a break-even or sensitivity calculation, and identify who pays and benefits.
6. State the strongest objection, the evidence that would change your decision and the controls you would reject.

Self-assessment rubric: 100 points

Decision and comparator 20 points

A bounded decision, realistic alternative and explicit trade-offs are supported rather than asserted.

Connected correction architecture 25 points

Information, competence, authority and remedy are connected through testable handoffs, including an adverse scenario.

Evidence and claim discipline 25 points

Assumptions, observations and causal claims are separated; limitations and sensitivity affect the conclusion.

Resources and incentives 15 points

Costs, bearers, time horizon and uncertainty are explicit, including the cost of the safeguards.

Counterargument and revision 15 points

A serious objection leads to a defensible modification, rejection or falsifiable defence.

Use the rubric for self-assessment or discussion. The site does not award an automatic mark, certification or academic credit. A completed capstone is not proof that a real deployment is safe or legally compliant.