READING ROOM ↗
PSYBORG / LEARNING FROM HARM

Look at what happened. Then ask why.

Read an AI incident report critically: trace the evidence, separate harm from potential harm, and record what the case can and cannot establish.

Incident monitoring · Edited by PsyBorg · Updated 2026-11-16

Start with the evidence behind the headline

When an AI incident appears in your feed, begin with a question: what does the available source actually establish? Find the original disclosure, complaint, decision or investigation where possible. Record who produced it, when it was published and whether later information changes the account.

The OECD's 2025 common reporting framework includes supporting materials, affected stakeholders, the system's role and actions taken in response. Its Annex C describes different ways AI can be involved, including as a direct cause or contributing factor. More than one relationship may apply. [1]

PSYBORG's reading method below uses those distinctions to organize questions. It is an editorial aid, not the OECD's reporting form or a legal classification.

Separate an observed outcome from a possible one

The OECD distinguishes incidents involving actual harm from hazards involving plausible potential harm. Its definitions cover a wider set of harms than physical injury alone. [2]

When reading a case, write down the outcome the source reports. Was information disclosed? Was an action attempted or completed? Is harm documented, alleged or still being investigated? Keep those descriptions attached to the source that supports them.

An unresolved detail deserves an explicit note. “The report does not establish whether the message was delivered” gives a future reader something to investigate. Quietly replacing that gap with “no harm occurred” loses information.

A hypothetical example: three stories about one event

Imagine that a news article describes an AI-assisted service sending an incorrect account notice. Two other outlets repeat the first article. The service operator confirms that an incorrect notice was generated, but its disclosure leaves delivery and customer consequences unresolved.

In PSYBORG's suggested evidence note, record one reported event with three news stories, the operator's narrower statement, and two open questions: who received the notice, and what happened to them? Repetition of the first account should not be mistaken for three independent investigations.

This is a hypothetical reading exercise. It makes no claim about a real company, product or incident.

Build a short incident note

PSYBORG suggests recording the following before sharing a broad conclusion.

  1. Source and status: identify the original material and distinguish allegations, an institution's own account and established findings.
  2. System and setting: state the task, version and permissions where known. Mark missing details as unknown.
  3. People and outcomes: describe who may be affected and what the source supports about consequences.
  4. Mechanism: explain the system's reported role and any relevant human or organizational decisions, without assuming intention.
  5. Response and open questions: record what changed, who says it changed, and what evidence would help assess the response.

Keep an occurrence date separate from publication and review dates. A new article about an older event may add evidence without describing a new occurrence.

Be careful when comparing counts

The OECD's monitor methodology describes collecting news reports and recognizes that publicly reported cases represent only a subset of incidents and hazards. [2]

PSYBORG recommends asking what a count includes before treating it as a trend. Does it count articles, distinct events, attempts or affected people? Have collection rules or coverage changed? To estimate failures per use, you would also need a suitable measure of use over the same period. A list of reported cases alone does not supply that denominator.

The useful next question is often specific: what part of the workflow allowed this outcome, and what evidence would show that the response addresses it?

Sources and scope

[1] OECD, Towards a common reporting framework for AI incidents, 28 February 2025. See Table 2.2, printed pages 17–18, and Annex C, printed page 25. DOI: 10.1787/f326d4ac-en.
Read source (www.oecd.org)

[2] OECD.AI, Overview and methodology of the AI Incidents and Hazards Monitor. See Overview and Definitions.
Read source (oecd.ai)

Sources checked 16 November 2026. The reading method, hypothetical example and comparison questions are PSYBORG editorial analysis. They do not establish a legal finding, the prevalence of AI failures or the effectiveness of a particular intervention.

Continue reading

Explore the sourced incident records
Read source (psyborg369.com)

A confident answer is not a source.
Read source (psyborg369.com)

OECD topic overview — the original briefing's further-reading source
Read source (www.oecd.org)

SOURCE & FURTHER READINGOECD · Common reporting framework for AI incidents, 2025

Get the next briefing by email