Summary: A third and different meaning of “AI incident response”: a public official statement by the entity responsible for an AI incident, submitted to the AI Incident Database. What makes a good one, and what most of them are missing.
Sources: raw/AI-incident-response/Defining an "AI Incident Response".md
Last updated: 2026-07-28
The definition
“An AI Incident Response refers to a public official response to an incident in the AI Incident Database (AIID) from an entity (i.e. company, organization, individual) allegedly responsible for developing or deploying the AI or AI system involved in said incident” (source: Defining an “AI Incident Response”.md).
Note this is a disclosure artefact, not an operational process. It is the AI-sector analogue of a public post-mortem, not of an IR playbook. Do not confuse it with operational AI incident response — the two share a name and nothing else.
What a complete response contains
- What happened,
- Why the incident happened,
- What the entity has done or is going to do to
- remediate the incident, and
- prevent future incidents from recurring.
The pattern in real responses
The source works through two examples, and both fail on the same element.
Knightscope — a K5 autonomous security robot collided with a 16-month-old boy at Stanford Shopping Center, 7 July 2016. The response gave operational context (25,000 miles, 35,000 hours, ~1 mph, “nearly 30 sensors, including a multitude of laser ranging devices, sonar sensors”), an apology, and remediation with the affected party.
“What is missing from the response? The entity did not provide many details about prevention of future incidents.”
Facebook / The BL — a network of pages, groups and fake accounts using GAN-generated face photos bypassed moderation to push political narratives. The response detailed the takedown precisely: “39 Facebook accounts, 344 Pages, 13 Groups and 22 Instagram accounts” from Georgia; “610 accounts, 89 Facebook Pages, 156 Groups and 72 Instagram accounts” from Vietnam and the US. It also stated the principle applied: “We’re taking down these Pages, Groups and accounts based on their behavior, not the content they posted.”
“What is missing from the response? The entity did not include details on prevention of future incidents or recurring harms.”
Both examples fail element 3.ii. Entities describe what happened and what they did about it; they do not commit to what changes so it does not happen again. That is a consistent, structural gap in AI incident disclosure — and it is precisely the part that has external value.
Note also that Facebook was not the deploying party: “their involvement as a social media platform having a responsibility to protect their users from these attacks qualifies their response as an incident response.” Platform responsibility for adversarial AI use by third parties counts.
Why the AIID matters operationally
Not just as a transparency venue. The GPAI Code of Practice Measure 9.1 requires signatories to identify serious incidents by consulting external sources including “police and media reports, social media content, academic research, and incident databases” (source: raw/Laws/Overview of the Code of Practice...md).
So incident databases are a mandated detection input for GPAI systemic-risk providers. The AIID is a monitoring feed, not just an archive.
The disclosure argument
The source’s framing: publishing a response is an opportunity to “provide transparency to the public about the AI incident, the root incident causes, and the entity’s actions,” to “demonstrate proactivity and leadership,” and to “highlight the entity’s unique perspective and relevant analysis stemming from their technical knowledge of the underlying AI system components.”
“Full and complete incident responses are opportunities to simultaneously benefit society and corporate interests.”
Relationship to mandatory reporting
Voluntary public disclosure and mandatory regulatory reporting are different obligations with different audiences:
| AIID response | AI Act Art 73 / CoP C9 | |
|---|---|---|
| Audience | Public | Market surveillance authorities / AI Office |
| Mandatory | No | Yes |
| Timing | Any | 2–15 days |
| Root cause required | Expected, often missing | Required (CoP Measure 9.2) |
| Prevention commitments | Expected, usually missing | Corrective measures required |
The Code of Practice’s required root-cause schema — “model outputs that caused or contributed,” inputs used, “systemic mitigation failures or circumventions” — is far more demanding than anything in the two AIID examples above. See serious incident reporting.