AI Product Incident Postmortems: Causal Chains for Model Behavior

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-ethics-safety-AI-and-product-incident-postmortem-adults

1. Which task can an AI tool appropriately assist with during an incident postmortem?

   1. Drafting a template that maps the causal chain across prompt, retrieval, model, and policy layers
   2. Assigning accountability to specific team members for the failure
   3. Determining which team member should be formally disciplined
   4. Deciding which remediation tradeoffs are acceptable to ship

2. Why is it problematic to identify a single root cause in an AI incident postmortem?

   1. It may hide failures that exist across multiple layers and leave future incidents unprotected
   2. It makes the postmortem too long to review
   3. Legal teams prefer simple, single-cause explanations
   4. Senior leadership will reject postmortems with multiple causes

3. What does a blameless postmortem review primarily seek to achieve?

   1. Determining which department should cover the costs of the incident
   2. Creating a record that absolves the engineering team of all responsibility
   3. Understanding the systemic factors that contributed to the failure to prevent recurrence
   4. Identifying which individual made the most errors during the incident

4. Which of the following layers should be included in a comprehensive AI incident causal chain analysis?

   1. The prompt, retrieval, model, policy filter, and UI layers
   2. Only the user interface where the error was visible
   3. The hardware layer and the network infrastructure
   4. Only the model layer where the prediction was generated

5. An AI incident postmortem reveals that a harmful response was triggered by a specific user prompt, exacerbated by a retrieval system that pulled irrelevant context, and not caught by the policy filter. What does this pattern suggest?

   1. The model is entirely responsible and should be retrained immediately
   2. The failure occurred across multiple layers and requires fixes at each layer
   3. The policy filter is working correctly but needs to be turned off
   4. The user prompt should be blocked for all future interactions

6. Who should make the final decision about which remediation tradeoffs are acceptable to implement after an AI incident?

   1. Human stakeholders with authority to balance business, safety, and technical considerations
   2. The AI system that generated the incident report
   3. The legal department without input from other teams
   4. The engineering team lead alone

7. What information can AI tools reliably extract from logs during an incident investigation?

   1. Event timelines spanning multiple system layers and components
   2. Whether the incident was caused by malicious intent
   3. The exact financial damage caused by the incident
   4. Which employee should be fired for the failure

8. A postmortem concludes that 'the model hallucinated, therefore the model is the problem.' Why is this conclusion inadequate?

   1. It ignores potential contributing factors from prompt design, retrieval, policy filters, and UI display
   2. The model cannot be the problem since it was deployed correctly
   3. Postmortems should never mention the model
   4. Models never hallucinate, so the conclusion is factually wrong

9. What is the primary purpose of adding regression tests after an AI incident?

   1. To punish the team that introduced the original bug
   2. To ensure the specific failure mode cannot reoccur in future deployments
   3. To replace the need for any future postmortems
   4. To satisfy regulatory requirements that mandate testing

10. Which statement best reflects the appropriate role of AI in incident postmortems?

    1. AI should lead the postmortem and make final determinations
    2. AI can draft templates and analyze logs but cannot assign blame or decide on actions
    3. AI should be excluded from postmortems entirely to protect data
    4. AI should write the final incident report for legal review

11. In a postmortem for an AI product that gave incorrect medical advice, which element would be most important to document in the causal chain?

    1. The employee's performance review history
    2. The lunch break schedule of the on-call team
    3. The marketing copy that described the product
    4. The exact user prompt that triggered the response, any retrieved context, the model version used, the policy filter configuration, and how the UI presented the answer

12. Why should a postmortem for an AI incident include documentation of user impact?

    1. To understand the severity and prioritize remediation efforts appropriately
    2. To compare the incident with competitors' incidents
    3. To determine who should be fired
    4. To justify the incident to shareholders

13. A team completes a postmortem that identifies only the retrieval system as the cause of an incident. Later, a similar incident occurs with a different retrieval setup. What is the most likely underlying issue?

    1. The postmortem process is fundamentally flawed
    2. The model was not properly retrained
    3. The second incident was unrelated to the first
    4. The original postmortem was incomplete and missed other contributing factors

14. What distinguishes a high-quality AI incident postmortem from a basic service outage postmortem?

    1. AI incident postmortems require legal review before publication
    2. There is no meaningful difference between the two types
    3. AI incident postmortems are shorter since AI is more predictable
    4. AI incident postmortems must trace prompt design, retrieval behavior, model versions, policy filters, and UI rendering, not just system uptime

15. During a postmortem, an AI tool suggests that the engineering manager was negligent. What should the team do with this suggestion?

    1. Accept the suggestion and begin disciplinary action
    2. Ignore it completely since AI cannot make suggestions about people
    3. Use it as the primary finding in the final report
    4. Treat it as one data point among many and verify with human analysis before making any accountability decisions