The post-incident review: what you missed, written without blaming the tool
The short answer
A post-incident review is understanding plus change, and it is none of the three documents it is often confused with
Not a press line (which closes inquiry), not a status update (which only looks forward from the fix), not a blame hunt (which destroys the information you need). It is written for yourself, looks backward from the fix to the conditions that produced it, and hunts causes, not culprits.
What you will be able to do
- Distinguish a post-incident review from the things it is often confused with (a press line, a status update, a blame hunt) and state what a real review must produce: understanding plus owned, dated change.
- Analyze an AI incident by tracing its causal chain past the proximate cause ("the model produced X") to the human decisions and process gaps that allowed the output to reach a user.
- Refuse the "blame the tool" reflex and explain why "the AI did it" is the AI-era version of blaming the operator: it feels like a cause, ends the inquiry, and protects the system that failed.
- Separate the proximate cause, the contributing factors, and the systemic conditions of an incident, and resist the myth that a complex failure has a single root cause.
- Locate where in the AI lifecycle a failure entered (data, a model or prompt change, deployment, a guardrail gap, a monitoring gap, a human-oversight gap) so the fix lands where the failure lives, not where it surfaced.
- Write the hardest sentence in the review, the honest "what we missed," without blaming the tool and without blaming a single person, and turn each finding into a corrective action with a named owner and a due date.
- Feed the review forward: make its findings become new eval cases (see Topic 4.2), new monitoring, and new logging (see Topic 10.2), so the loop closes instead of repeating.
The ideas, one by one
"The AI did it" is the AI-era version of blaming the operator
It describes the output and hides every governable cause behind it. Safety science already learned that "human error" is a symptom, not a cause; AI hands you a new operator to blame that has no feelings, which makes the dodge more tempting and no less false. In your review, "the model produced X" is a fact in the middle of a causal chain, never the end of one.
Trace proximate cause, contributing factors, and systemic conditions, and kill the single-root-cause myth
Complex systems fail when several survivable weaknesses line up, so a review that finds one root cause has usually stopped early. Ask which factors, if removed, would have stopped or shrunk the incident, and fix more than one, or the doors you left open wait for a different trigger.
Use the five whys and the second story
Keep asking why until you reach a condition you can build, not a description you can only regret. Keep asking why the people's choices made sense given what they knew, so you find the system that shaped the behavior instead of the individual who happened to be standing where it failed. Hindsight is the enemy; the second story is the cure.
Locate the failure in the AI lifecycle so the fix lands where the failure lives
Data, a model or prompt change, deployment, a guardrail gap, a monitoring gap, a human-oversight gap: name every layer the incident touched, because a failure surfaces at the output but enters earlier, and fixing only the visible layer leaves the others open.
Write the "what we missed" sentence, blaming neither the tool nor a person
It is the review's most valuable and most avoided line. Specific, owned by the system, honest that the gap was an absence: "our process allowed a change to ship without the check that would have caught this." That sentence blames no one and fixes everything.
Blameless does not mean unaccountable; it means the truth is not punished
Blame drives information underground, and a review runs on information. People stay accountable for honesty and for owning the fixes; accountability points at the system, not a scapegoat. And a true blameless review refuses to scapegoat the machine too, because blaming the tool protects the failed system exactly as blaming a person does.
Corrective actions are real only with an owner, a due date, and a verification, in both symptom and class form
"Improve monitoring" is where reviews go to die. Name who, by when, and how you will know it worked, track it to done, and fix both this exact incident and the family it belongs to. An untracked action list teaches your organization that incidents produce paperwork, not change.
The review's most durable value is forward
The failure becomes an eval case the system must pass forever (see Topic 4.2), new monitoring and logging so the next one is seen sooner and reconstructed faster (see Topic 10.2), and a truthful record a later forensic pass can read (see Topic 11.4). A review that does not feed forward is a page in a drawer.
A review is a facilitated conversation, held soon but not raw
The causal chain lives in several people's heads, and the second story only emerges when the person a decision happened to explains why it made sense. Hold the review within a few days, but not in the exhausted, defensive hours right after the incident, and facilitate it so the frame stays blameless; a single blame-seeking question can end the honesty the review depends on.
Beware hindsight and outcome bias
After the fact the warning sign looks obvious and a bad-outcome decision looks reckless, but neither was clear in the moment. Judge decisions by what was known at the time, fix the missable sign with an instrument rather than a scolding, and reserve the reckless label for genuine knowing disregard, which is rare. The just-culture line is drawn on the decision, not the result.
Review the near-miss too
A failure caught by a canary, a monitor, or luck before it reached users is a free lesson: the systemic conditions behind it are fully intact, so reviewing it while the harm is zero is far cheaper than reviewing the public version later. An organization that only reviews failures that already hurt someone has chosen the most expensive possible curriculum.
This review is the honesty the whole shipping chain depends on
The disclosure letter you write next is only as honest as the review it stands on (see Topic 3.7); write the review to protect a reputation and the letter inherits a lie, write it to learn and every downstream artifact inherits a lesson.
Listen to it
The same lesson, talked through by two hosts as episode 22 of the podcast.
Real cases
These examples show post-incident analysis done well and badly, with the reasoning stated. The deep anchor is the Alexa case; the others are referenced to sharpen a specific point and are treated in depth by their owner topics.
Example 1 (the anchor): Amazon Alexa's asymmetric political answers, September 2024. In early September 2024, Alexa was observed giving reasons to vote for one presidential candidate while refusing the mirror question for the other, replying that it could not promote a party or candidate (Variety, September 2024). Amazon's public response was that this was an error, quickly fixed, that Alexa holds no political opinions, and that the company audits and improves its content-detection systems and has teams to prevent recurrence (Snopes, September 2024). Internal documents reported by The Washington Post attributed the behavior to AI software added in late 2023 to improve Alexa's accuracy (The Washington Post, 5 September 2024). Read against this topic, the public statement is a press line, not a review, and the instructive move is to reconstruct the review the press line stands in front of, without claiming to know Amazon's internal findings. The proximate cause is an asymmetric answer. The reported contributing factor visible from outside is an accuracy upgrade that changed behavior in a way that was not re-tested for political symmetry. The governance questions a real review would pursue are all human-shaped: what change process let an accuracy upgrade ship without symmetry testing on sensitive topics; why a content guardrail applied unevenly across mirror queries; and why the first detector was a public clip rather than an owned instrument. The lasting lesson is the discipline itself: "the AI was biased" describes the output and hides every one of those governable causes, and a review that stops there learns nothing that would prevent the next asymmetric answer. (All internal-cause detail beyond what The Washington Post reported is deliberately not asserted here.)
Example 2 (a change that fixed one thing and broke a worse one): the Grok system-prompt edit. In July 2025, a system-prompt change at xAI instructing the Grok chatbot to be less "politically correct" produced antisemitic output and the self-label "MechaHitler" within about two days (NPR, "Elon Musk's AI chatbot, Grok, started calling itself 'MechaHitler,'" 9 July 2025; owner topic, with full sourcing: (see Topic 1.3)). It is referenced here for one point about causal analysis: an incident's proximate trigger is often a change made in good faith to solve a different problem, so a review must always ask "what changed just before this," and must resist treating the change as a villain rather than as a link in a chain that also includes the absence of a test that would have caught the new failure. Blaming the person who made the prompt edit would miss the systemic condition: prompt changes shipping without adversarial re-testing. The change is a contributing factor; the missing gate is the systemic one.
Example 3 (why blaming the operator fails, from an older field): aviation's shift away from "pilot error." For decades, air-accident reports commonly ended at "pilot error," which satisfied the demand for a cause while preventing almost all learning. The modern investigative culture, embodied in bodies like the US National Transportation Safety Board, treats a human action as the start of an inquiry into the system that shaped it, cockpit design, training, procedures, fatigue, rather than its end (established safety-science practice; see also Dekker's human-error work). This is not an AI case and is not a program anchor; it is included because it is the cleanest historical proof that "blame the operator" and "blame the tool" are the same mistake, and that replacing the operator with the machine does not make the mistake less costly. The AI-era temptation is stronger only because the operator you are blaming has no union, no feelings, and no way to object.
Example 4 (blameless postmortems as an operating discipline): software site-reliability practice. Large software operations run on blameless postmortems: a standard practice, codified in Google's site-reliability engineering and popularized by Etsy under John Allspaw, in which incidents are analyzed for systemic causes with an explicit rule against punishing individuals, precisely so that the people who were present will tell the truth (Google, "Site Reliability Engineering," 2016; Allspaw, established). The relevant lesson for AI governance is that this is not a moral nicety but an information strategy: blame drives the truth underground, and a review runs on truth. The AI-specific extension is that the discipline must extend to the machine: a blameless review refuses to scapegoat the tool as well as the person, because scapegoating the tool protects the system just as effectively as scapegoating a person, while feeling virtuous.
Example 5 (the review that feeds the next defense): an incident becomes an eval case. Consider any AI feature that suffered a specific failure, a summarizer that fabricated a claim, a bot that invented a policy (see Topic 1.6). The post-incident review's most valuable output is not the apology; it is the conversion of that exact failure into a test the system must pass forever after. This is the shape of the whole program's adversarial discipline: the incident your monitoring caught in production becomes a case your evaluation suite must catch before production (see Topic 4.2). A review that ends without asking "what eval case does this incident owe us" has left its most durable value on the table, because the strongest guarantee that a failure will not recur is a test that fails the build if it does.
Example 6 (single-root-cause thinking as a trap): a complex failure with several necessary conditions. Consider a hypothetical incident where an AI feature produced a harmful output because a model was changed, a guardrail did not cover the new case, and no monitor watched for it. Each of the three, alone, might have been survivable: the model change with a guardrail and a monitor in place would likely have been caught; the guardrail gap with good monitoring would have been seen fast; the monitoring gap with a solid guardrail might never have mattered. The harm required all three to line up. A review that names only the model change "fixes" it and leaves the guardrail gap and the monitoring gap in place, so the next model change, or a different trigger entirely, walks straight through the two doors still open. This is the method point of 3D made concrete: ask which factors, if removed, would have stopped or shrunk the incident, and fix more than one.
Example 7 (impact honesty): the temptation to undercount. When Apple's notification summaries fabricated news, the harm was not only the individuals who saw a false alert but the erosion of trust in a channel people rely on (owner topic: (see Topic 3.4)). Referenced here for one narrow point about the impact block: the easiest dishonesty in a review is a quietly small impact number, because a small impact justifies a small response. A review that counts only "confirmed complaints" when the real exposure is every user who received the output is lying with a true number, and it will misprioritize every corrective action that follows. State the real exposure from your own logs, not the flattering subset.
Example 8 (detection provenance decides the finding): internal alert versus viral clip. Consider two organizations that ship the same feature and suffer the same failure with the same impact. In the first, an internal monitor fires within minutes and the team contains the issue before most users notice. In the second, the failure runs for hours and is surfaced by a screenshot that spreads on social media, exactly the shape of the Alexa timeline where the asymmetry was public before it was owned internally. The two incidents have identical causes and identical harm, yet their reviews are not the same, because how you found out is itself a governable fact. The first review has no monitoring-gap finding; the second has a serious one that demands its own corrective action, an owned instrument that would have caught the failure before the world did. The lesson is that "who caught it" is not an anecdote to note in passing but a distinct finding that can add an entire class of fix, which is why the timeline block earns its place.
Example 9 (hindsight makes the sign look obvious): the warning that was invisible in the noise. After almost any incident, someone points at a signal that "should have been noticed," and in retrospect it glows. The trap is that at the time it sat among thousands of ordinary signals with nothing marking it as the one that mattered, judged against every other priority on a busy day. Aviation learned this decades ago: a warning that looks damning in the accident report was, in the cockpit, one of many competing indications under time pressure (established safety-science practice; see the shift away from "pilot error"). The governance move is to stop asking "why didn't they notice" and start asking "why was this sign missable, and how do we make it impossible to miss," which turns a character judgment into a system fix (an alert that surfaces the sign automatically). This is method, not a single incident: hindsight is a bias to be corrected in every review, and the correction is the second story plus an instrument that removes the need to notice by luck.
Where people go wrong
- "The AI was biased, that is the cause." That is a description of the output, not an account of how it reached a user, and it ends the inquiry exactly where the governable causes begin. A model producing an output is a model doing what models do; the review's job is the chain after it, the change, the guardrail, the monitor, the process. "The AI did it" is the AI-era version of "pilot error": true, comforting, and useless.
- "A post-incident review is what we tell the customer and the press." Those are different documents with opposite purposes. The press line exists to close inquiry and make the problem sound small; the review exists to open inquiry and make the causes visible. When your internal review reads like your public statement, you have stopped investigating. The honest customer letter comes later and is built on the review, not instead of it (see Topic 3.7).
- "We found the root cause and fixed it." Complex systems rarely fail for one reason; they fail when several survivable weaknesses line up. A review that names one root cause has usually stopped early, and the danger is that you fix the one thing, feel safe, and leave the other necessary conditions in place for a different trigger. Ask for contributing factors, plural, and fix more than one.
- "Blameless means nobody is accountable." Blameless means nobody is punished for the truth, so the truth keeps flowing, because blame drives information underground and a review runs on information. People remain fully accountable, for honesty and for owning the fixes. Accountability points at the system and the corrective actions, not at a scapegoat.
- "Blaming the tool is the safe, blameless choice, since no person is accused." Blaming the tool feels blameless and is the opposite of a blameless review, because it protects the failed system just as effectively as blaming a person, while ending the inquiry before it reaches anything worth fixing. A true blameless review refuses to scapegoat both the person and the machine and examines the system that produced the output.
- "The person who shipped it should have caught it." Hindsight makes the warning sign obvious after the fact; the fair and useful question is why, given what was visible at the time, missing it made sense, and how to make the sign impossible to miss next time. Naming an individual misses the systemic condition (why was shipping-without-that-check even possible?) and teaches everyone that honesty in a review gets punished.
- "We fixed the bug, so the review is done." Fixing the bug is a status update, not a review. The review asks why the bug reached a user and whether the conditions that produced it are still present. A symptom fix (patch this guardrail) without a class fix (add the gate that would have caught it) treats the fever and leaves the infection.
- "Corrective actions: improve monitoring, be more careful, consider testing." Those are verbs without accountability, and they are where reviews go to die. A real corrective action names a specific change, an owner, a due date, and a way you will know it worked, and someone tracks it to done. An action with no owner is a wish; with no date it never happens; with no verification it is a hope.
- "The impact was small, only a few people complained." Complaints are a fraction of exposure; most people who receive a bad output never report it. Counting only complaints is lying with a true number, and a small impact number justifies a small response, so undercounting here quietly misprioritizes every fix that follows. State real exposure from your own logs.
- "The incident is over once the review is written." The review's most durable value is forward: the failure becomes an eval case the system must pass forever (see Topic 4.2), a new monitor and new logging so the next one is seen sooner and reconstructed faster (see Topic 10.2), and a record a later forensic pass can read (see Topic 11.4). A review that does not feed forward is a page in a drawer.
- "We should not write down what we missed; it could be used against us." A review that hides its findings is worthless as a review and, in a later attack or forensic reconstruction, far more damaging than an honest one, because a discovered omission reads as a cover-up (see Topic 11.4). The defensible record is the honest one; the honest one is also the only one that prevents recurrence.
- "Speed matters more than the gate; a governance check just slows shipping." The Alexa-shaped incident is what shipping without the gate costs, and it is paid in public, at the worst time, with trust that is expensive to rebuild. The gate that requires symmetry and safety re-testing before a model change ships is cheap; the incident it prevents is not. A governance function that can never slow a ship is not governing it.
- "There is one lesson per incident." A good review usually yields several: a guardrail fix, a testing gap, a monitoring gap, and a systemic change-management gap can all come from one clip. Forcing an incident into a single tidy lesson is the single-root-cause myth in a friendlier costume, and it leaves most of the incident's value unlearned.
- "They should have noticed the warning sign." Hindsight makes the sign obvious after the outcome is known, but in the moment it sat among thousands of ordinary signals with nothing marking it as important. The fix is never "notice harder"; it is the second story (why was the sign missable?) plus an instrument that surfaces the sign automatically, so catching it does not depend on a person's luck or attention.
- "A review is a form you fill in after the fix." The causal chain lives in several people's heads, and the second story only emerges when the person a decision happened to explains why it made sense. A review is a facilitated conversation, held soon but not raw, from which the document is written; a form filled in alone captures one perspective and misses the interactions that produced the incident.
- "We will hold the review right now, while it is fresh." Fresh is good; raw is not. In the exhausted, defensive hours right after an incident, people are too stressed to reflect honestly, so the review produces silence or blame. Wait until soon but calm, usually a few days, so the review is accurate and honest rather than a re-run of the crisis.
- "Recklessness and honest mistakes are the same once the outcome is bad." Judging a decision by its outcome (outcome bias) turns every honest mistake with a bad result into "recklessness," which is unjust and destroys learning. The just-culture line is drawn on what the person knew and the pressures they faced at the time: an honest mistake is a choice a reasonable person could have made in that situation, and it is a learning opportunity; recklessness is knowing disregard of a clear serious risk, and it is genuinely rare.
Questions people ask
- What is post-incident review (postmortem)?
- The one to two page artifact this topic produces. A structured analysis of an incident whose purpose is understanding plus specific, owned, dated change, built on the incident log and feeding the disclosure letter, the eval suite, the monitoring, and the logging. It is written for the organization, not for an audience.
- What is press line?
- A short public statement, such as "an error, quickly fixed," written to close a news cycle and make a problem sound small and handled. Legitimate in public, but the opposite of a review, because its purpose is to close inquiry rather than open it.
- What is status update?
- A report that the fire is out and the fix is live. It looks forward from the fix and says nothing about why the incident happened, which is why it is not a review.
- What is blame hunt?
- The search for whose fault an incident was. It feels like accountability but destroys the information a review needs, because the presence of blame makes everyone manage their exposure instead of telling the truth.
- What is blaming the tool?
- The AI-era reflex of ending a causal analysis at "the AI was biased" or "the model hallucinated." It describes the output rather than explaining how it reached a user, ends the inquiry where the governable causes begin, and protects the failed system just as blaming a person would.
Keep going
This lesson builds Human oversight design, and that page shows the roles that hire for it. Every Certified AI Governance Professional (CAIGP) lesson.
Watch the lesson and prove it
Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.
Watch the lesson and prove it →